Title of Invention	ENCODER, SYSTEM AND METHOD FOR COMPRESSING AUDIO SIGNALS
Abstract	The invention relates to an encoder (200) comprising an input (201) for inputting frames of an audio signal in a frequency band, at least a first excitation block (206) for performing a first excitation for a speech like audio signal, and a second excitation block (207) for performing a second excitation for a non-speech like audio signal. The encoder (200) further comprises a filter (300) for dividing the frequency band into a plurality of sub bands each having a narrower bandwidth than said frequency band. The encoder (200) also comprises an excitation selection block (203) for selecting one excitation block among said at least first excitation block (206) and said second excitation block (207) for performing the excitation for a frame of the audio signal on the basis of the properties of the audio signal at least at one of said sub bands. The invention also relates to a device, a system, a method and a storage medium for a computer program.

Title of Invention

ENCODER, SYSTEM AND METHOD FOR COMPRESSING AUDIO SIGNALS

Abstract

The invention relates to an encoder (200) comprising an input (201) for inputting frames of an audio signal in a frequency band, at least a first excitation block (206) for performing a first excitation for a speech like audio signal, and a second excitation block (207) for performing a second excitation for a non-speech like audio signal. The encoder (200) further comprises a filter (300) for dividing the frequency band into a plurality of sub bands each having a narrower bandwidth than said frequency band. The encoder (200) also comprises an excitation selection block (203) for selecting one excitation block among said at least first excitation block (206) and said second excitation block (207) for performing the excitation for a frame of the audio signal on the basis of the properties of the audio signal at least at one of said sub bands. The invention also relates to a device, a system, a method and a storage medium for a computer program.

Full Text	Classification of audio signals Field of the Invention The invention relates to speech and audio coding in which the encoding mode is changed depending whether input signal is speech like or music like signal. The present invention relates to an encoder comprising an input for inputting frames of an audio signal in a frequency band, at least a first excitation block for performing a first excitation for a speech like audio signal, and a second excitation block for performing a second excitation for a non-speech like audio signal. The invention also relates to a device comprising an encoder comprising an input for inputting frames of an audio signal in a frequency band, at least a first excitation block for performing a first excitation for a speech like audio signal, and a second excitation block for performing a second excitation for a non-speech like audio signal. The invention also relates to a system comprising an encoder comprising an input for inputting frames of an audio signal in a frequency band, at least a first excitation block for performing a first excitation for a speech like audio signal, and a second excitation block for performing a second excitation for a non-speech like audio signal. The invention further relates to a method for compressing audio signals in a frequency band, in which a first excitation is used for a speech like audio signal, and second excitation is used for a non-speech like audio signal . The invention relates to a module for classifying frames of an audio signal in a frequency band for selection of an excitation among at least a first excitation for a speech like audio signal, and a second excitation fora non-speech like audio signal. The invention relates to a computer program product comprising machine executable steps for compressing audio signals in a frequency band, in which a first excitation is used for a speech like audio signal, and second excitation is used for a non-speech like audio signal. Background of the Invention In many audio signal processing applications audio signals are compressed to reduce the processing power requirements when processing the audio signal. For example, in digital communication systems audio signal is typically captured as an analogue signal, digitised in an analogue to digital (A/D) converter and then encoded before transmission over a wireless air interface between a user equipment, such as a mobile station, and a base station. The purpose of the encoding is to compress the digitised signal and transmit it over the air interface with the minimum amount of data whilst maintaining an acceptable signal quality level. This is particularly important as radio channel capacity over the wireless air interface is limited in a cellular communication network. There are also applications in which digitised audio signal is stored to a storage medium for later reproduction of the audio signal. The compression can be lossy or lossless. In lossy compression some information is lost during the compression wherein it is not possible to fully reconstruct the original signal from the compressed signal. In lossless compression no information is normally lost. Hence, the original signal can usually be completely reconstructed from the compressed signal. The term audio signal is normally understood as a signal containing speech, music (non-speech) or both. The different nature of speech and music makes it rather difficult to design one compression algorithm which works enough well for both speech and music. Therefore, the problem is often solved by designing different algorithms for both audio and speech and use some kind of recognition method to recognise whether the audio signal is speech like or music like and select the appropriate algorithm according to the recognition. In overall, classifying purely between speech and music or non-speech signals is a difficult task. The required accuracy depends heavily on the application. In some applications the accuracy is more critical like in speech recognition or in accurate archiving for storage and retrieval purposes. However, the situation is a bit different if the classification is used for selecting optimal compression method for the input signal. In this case, it may happen that there does not exist one compression method that is always optimal for speech and another method that is always optimal for music or non-speech signals. In practise, it may be that a compression method for speech transients is also very efficient for music transients. It is also possible that a music compression for strong tonal components may be good for voiced speech segments. So, in these instances, methods for classifying just purely for speech and music do not create the most optimal algorithm to select the best compression method. Often speech can be considered as bandlimited to between approximately 200Hz and 3400 Hz. The typical sampling rate used by an A/D converter to convert an analogue speech signal into a digital signal is either 8kHz or 16kHz. Music or non-speech signals may contain frequency components well above the normal speech bandwidth. In some applications the audio system should be able to handle a frequency band between about 20 Hz to 20 000 kHz. The sample rate for that kind of signals should be at least 40 000 kHz to avoid aliasing. It should be noted here that the above mentioned values are just non-limiting examples. For example, in some systems the higher limit for music signals may be about 10 000 kHz or even less than that. The sampled digital signal is then encoded, usually on a frame by frame basis, resulting in a digital data stream with a bit rate that is determined by a codec used for encoding. The higher the bit rate, the more data is encoded, which results in a more accurate representation of the input frame. The encoded audio signal can then be decoded and passed through a digital to analogue (D/A) converter to reconstruct a signal which is as near the original signal as possible. An ideal codec will encode the audio signal with as few bits as possible thereby optimising channel capacity, while producing decoded audio signal that sounds as close to the original audio signal as possible. In practice there is usually a trade-off between the bit rate of the codec and the quality of the decoded audio. At present there are numerous different codecs, such as the adaptive multi-rate (AMR) codec and the adaptive multi-rate wideband (AMRWB) codec, which are developed for compressing and encoding audio signals. AMR was developed by the 3rd Generation Partnership Project (3GPP) for GSM/EDGE and WCDMA communication networks. In addition, it has also been envisaged that AMR will be used in packet switched networks. AMR is based on Algebraic Code Excited Linear Prediction (ACELP} coding. The AMR and AMR WB codecs consist of 8 and 9 active bit rates respectively and aiso include voice activity detection (VAD) and discontinuous transmission (DTX) functionality. At the moment, the sampling rate in the AMR codec is 8 kHz and in the AMR WB codec the sampling rate is 16kHz. It is obvious that the codecs and sampling rates mentioned above are just non-limiting examples. ACELP coding operates using a model of how the signal source is generated, and extracts from the signal the parameters of the model More specifically, ACELP coding is based on a mode! of the human vocal system, where the throat and mouth are modelled as a linear filter and speech is generated by a periodic vibration of air exciting the filter. The speech is analysed on a frame by frame basis by the encoder and for each frame a set of parameters representing the modelled speech is generated and output by the encoder. The set of parameters may include excitation parameters and the coefficients for the filter as well as other parameters. The output from a speech encoder is often referred to as a parametric representation of the input speech signal. The set of parameters is then used by a suitably configured decoder to regenerate the input speech signal. For some input signals, the pulse-like ACELP-excitation produces higher quality and for some input signals transform coded excitation (TCX) is more optimal. !t is assumed here that ACELP-excitation is mostly used for typical speech content as an input signal and TCXexcitation is mostly used for typical music as an input signal. However, this is not always the case, i.e., sometimes speech signal has parts, which are music like and music signal has parts, which are speech like. The definition of speech like signal in this application is that most of the speech belongs to this category and some of the music may also belong to this category. For music like signals the definition is other way around. Additionally, there are some speech signal parts and music signal parts that are neutral in a sense that they can belong to the both classes. The selection of excitation can be done in several ways: the most complex and quite good method is to encode both ACELP and TCXexcitation and then select the best excitation based on the synthesised speech signal. This analysis-by-synthesis type of method will provide good results but it is in some applications not practical because of its high complexity. In this method for example SNR-type of algorithm can be used to measure the quality produced by both excitations. This method can be called as a "brute-force" method because it tries all the combinations of different excitations and selects afterwards the best one. The less complex method would perform the synthesis only once by analysing the signal properties beforehand and then selecting the best excitation. The method can also be a combination of pre-selection and "brute-force" to make compromised between quality and complexity. Figure 1 presents a simplified encoder 100 with prior-art high complexity classification. An audio signal is input to the input signal block 101 in which the signal is digitised and filtered. The input signal block 101 also forms frames from the digitised and filtered signal. The frames are input to a linear prediction coding (LPC) analysis block 102. It performs a LPC analysis on the digitised input signal on a frame by frame basis to find such a parameter set which matches best with the input signal. The determined parameters (LPC parameters) are quantized and output 109 from the encoder 100. The encoder 100 also generates two output signals with LPC synthesis blocks 103, 104. The first LPC synthesis block 103 uses a signal generated by the TCX excitation block 105 to synthesise the audio signal for finding the code vector producing the best result for the TCX excitation. The second LPC synthesis block 104 uses a signal generated by the ACELP excitation block 106 to synthesise the audio signal for finding the code vector producing the best result for the ACELP excitation. In the excitation selection block 107 the signals generated by the LPC synthesis blocks 103, 104 are compared to determine which one of the excitation methods gives the best (optimal) excitation. Information about the selected excitation method and parameters of the selected excitation signal are, for example, quantized and channel coded 108 before outputting 109 the signals from the encoder 100 for transmission. Summary of the Invention One aim of the present invention is to provide an improved method for classifying speech like and music like signals utilising frequency information of the signal. There are music like speech signal segments and vice versa and there are signal segments in speech and in music that can belong to either class. In other words, the invention does not purely classify between speech and music. However, it defines means for categorize input signal into music like and speech like components according to some criteria. The classification information can be used e.g. in a multimode encoder for selecting an encoding mode. The invention is based on the idea that input signal is divided into several frequency bands and the relations between lower and higher frequency bands are analysed together with the energy level variations in those bands and the signal is classified into music like or speech like based on both of the calculated measurements or several different combinations of those measurements using different analysis windows and decision threshold values. This information can then be utilised for example in the selection of the compression method for the analysed signal. The encoder according to the present invention is primarily characterised in that the encoder further comprises a filter for dividing the frequency band into a plurality of sub bands each having a narrower bandwidth than said frequency band, and an excitation selection block for selecting one excitation block among said at least first excitation block and said second excitation block for performing the excitation for a frame of the audio signal on the basis of the properties of the audio signal at least at one of said sub bands. The device according to the present invention is primarily characterised in that said encoder comprises a filter for dividing the frequency band into a plurality of sub bands each having a narrower bandwidth than said frequency band, that the device also comprises an excitation selection block for selecting one excitation block among said at least first excitation block and said second excitation block for performing the excitation for a frame of the audio signal on the basis of the properties of the audio signal at least at one of said sub bands. The system according to the present invention is primarily characterised in that said encoder further comprises a filter for dividing the frequency band into a plurality of sub bands each having a narrower bandwidth than said frequency band, that the system also comprises an excitation selection block for selecting one excitation block among said at least first excitation block and said second excitation block for performing the excitation for a frame of the audio signal on the basis of the properties of the audio signal at least at one of said sub bands. The method according to the present invention is primarily characterised in that the frequency band is divided into a plurality of sub bands each having a narrower bandwidth than said frequency band, that one excitation among said at least first excitation and said second excitation is selected for performing the excitation for a frame of the audio signal on the basis of the properties of the audio signal at least at one of said sub bands. The module according to the present invention is primarily characterised in that the module further comprises input for inputting information indicative of the frequency band divided into a plurality of sub bands each having a narrower bandwidth than said frequency band, and an excitation selection block for selecting one excitation block among said at least first excitation block and said second excitation block for performing the excitation for a frame of the audio signal on the basis of the properties of the audio signal at least at one of said sub bands. The computer program product according to the present invention is primarily characterised in that the computer program product further comprises machine executable steps for dividing the frequency band into a plurality of sub bands each having a narrower bandwidth than said frequency band, machine executable steps for selecting one excitation among said at least first excitation and said second excitation on the basis of the properties of the audio signal at least at one of said sub bands for performing the excitation for a frame of the audio signal. In this application, terms "speech like" and "music like" are defined to separate the invention from the typical speech and music classifications. Even if around 90% of the speech were categorized as speech like in a system according to the present invention, the rest of the speech signal may be defined as a music like signal, which may improve audio quality if the selection of the compression algorithm is based on this classification. Also typical music signals may fall in 80-90 % of the cases into music like signals but classifying part of the music signal into speech like category will improve the quality of the sound signal for the compression system. Therefore, the present invention provides advantages when compared with prior art methods and systems. By using the classification method according to the present invention it is possible to improve reproduced sound quality without greatly affecting the compression efficiency. Compared to the brute-force approach presented above, the invention provides a much less complex pre-selection type approach to make selection between two excitation types. The invention divides input signal into frequency bands and analyses the relations between lower and higher frequency bands together and can also use, for example, the energy level variations in those bands and classifies the signal into music like or speech like. Description of the Drawings Fig. 1 presents a simplified encoder with prior-art high complexity classification, Fig. 2 presents an example embodiment of an encoder with classification according to the invention, Fig. 3 illustrates an exampleof a VAD filter bank structure in AMRWB VAD algorithm, Fig. 4 shows an example of a plotting of standard deviation of energy levels in VAD filter banks as a function of the relation between low and high-energy components in a music signal, Fig. 5 shows an example of a plotting of standard deviation of energy levels in VAD filter banks as a function of the relation between low- and high-energy components in a speech signal, Fig. 6 shows an example of a combined plotting for both music and speech signals, and Fig. 7 shows an example of a system according to the present invention. Detailed Description of the Invention In the following an encoder 200 according to an example embodiment of the present invention will be described in more detail with reference to Fig. 2. The encoder 200 comprises an input block 201 for digitizing, filtering and framing the input signal when necessary. It should be noted here that the input signal may already be in a form suitable for the encoding process. For example, the input signal may have been digitised at an earlier stage and stored to a memory medium (not shown). The input signal frames are input to a voice activity detection block 202. The voice activity detection block 202 outputs a multiplicity of narrower band signals which are input to an excitation selection block 203. The excitation selection block 203 analyses the signals to determine which excitation method is the most appropriate one for encoding the input signal. The excitation selection block 203 produces a control signal 204 for controlling a selection means 205 according to the determination of the excitation method. If it was determined that the best excitation method for encoding the current frame of the input signal is a first excitation method, the selection means 205 are controlled to select the signal of a first excitation block 206. If it was determined that the best excitation method for encoding the current frame of the input signal is a second excitation method, the selection means 205 are controlled to select the signal of a second excitation block 207. Although the encoder of Fig. 2 has only the first 206 and the second excitation block 207 for the encoding process, it is obvious that there can also be more than two different excitation blocks for different excitation methods available in the encoder 200 to be used in the encoding of the input signal. The first excitation block 206 produces, for example, a TCX excitation signal and the second excitation block 207 produces, for example, a ACELP excitation signal. The LPC analysis block 208 performs a LPC analysis on the digitised input signal on a frame by frame basis to find such a parameter set which matches best with the input signal. LPC parameters 210 and excitation parameters 211 are, for example, quantised and encoded in a quantisation and encoding block 212 before transmission e.g. to a communication network 704 (Fig. 7). However, it is not necessary to transmit the parameters but they can, for example, be stored on a storage medium and at a later stage retrieved for transmission and/or decoding. Fig. 3 depicts one example of a filter 300 which can be used in the encoder 200 for the signal analysis. The filter 300 is, for example, a filter bank of the voice activity detection block of the AMR-WB codec, wherein a separate filter is not needed but it is also possible to use other filters for this purpose. The filter 300 comprises two or more filter blocks 301 to divide the input signal into two or more subband signals on different frequencies. In other words, each output signal of the filter 300 represents a certain frequency band of the input signal. The output signals of the filter 300 can be used in the excitation selection block 203 to determine the frequency content of the input signal. The excitation selection block 203 evaluates energy levels of each output of the filter bank 300 and analyses the relations between lower and higher frequency subbands together with the energy level variations in those subbands and classifies the signal into music like or speech like. The invention is based on examining the frequency content of the input signal to select the excitation method for frames of the input signal. In the following, AMR-WB extension (AMR-WB+) is used as a practical example used to classify input signal into speech like or music like signals and to select either ACELP- or TCX-excitation for those signal respectively. However, the invention is not limited to AMR-WB codecs or ACELP- and TCX- excitation methods. In the extended AMR-WB (AMR-WB+) codec, there are two types of excitation for LP-synthesis: ACELP pulse-like excitation and transform coded excitation (TCX). ACELP excitation is the same than used already in the original 3GPP AMR-WB standard (3GPP TS 26.190) and TCX is an improvement implemented in the extended AMR-WB. AMR-WB extension example is based on the AMR-WB VAD filter banks, which for each 20 ms input frame, produces signal energy E(n) in the 12 subbands over the frequency range from 0 to 6400 Hz as shown in Fig. 3. The bandwidths of the filter banks are normally not equal but may vary on different bands as can be seen on Fig. 3. Also the number of subbands may vary and the subbands may be partly overlapping. Then energy levels of each subband are normalised by dividing the energy level E(n) from each subband by the width of that subband (in Hz) producing normalised EN(n) energy levels of each band where n is the band number from 0 to 11. Index 0 refers to the lowest subband shown in Fig. 3. In the excitation selection block 203 the standard deviation of the energy levels is calculated for each of the 12 subbands using e.g. two windows: a short window stdshort(n) and a long window stdlong(n). For AMR-WB+ case, the length of the short window is 4 frames and the long window is 16 frames. In these calculations, the 12 energy levels from the current frame together with past 3 or 15 frames are used to derive these two standard deviation values. The special feature of this calculation is that it is only performed when voice activity detection block 202 indicates 213 active speech. This will make the algorithm react faster especially after long speech pauses. Then, for each frame, the average standard deviation over all the 12 filter banks are taken for both long and short window and average standard deviation values stdashort and stdalong are created. For frames of the audio signal, also a relation between lower frequency bands and higher frequency bands are calculated. In AMR-WB+ energy of lower frequency subbands LevL from 1 to 7 are taken and normalised by dividing it by the length (bandwidth) of these subbands (in Hz). For higher frequency bands from 8 to 11 energy of them are taken and normalised respectively to create LevH. Note that in this example embodiment the lowest subband 0 is not used in these calculations because it usually contains so much energy that it will distort the calculations and make the contributions from other subbands too small. From these measurements the relation LPH = LevL / LevH is defined. In addition, for each frame a moving average LPHa is calculated using the current and 3 past LPH values. After these calculations a measurement of the low and high frequency relation LPHaF for the current frame is calculated by using weighted sum of the current and 7 past moving average LPHa values by setting slightly more weighting for the latest values. It is also possible to implement the present invention so that only one or few of the available subbands are analysed. Also average level AVL of the filter blocks 301 for the current frame is calculated by subtracting the estimated level of background noise from each filter block output, and summing these levels multiplied by the highest frequency of the corresponding filter block 301, to balance the high frequency subbands containing relatively less energy than the lower frequency subbands. Also the total energy of the current frame TotEO from all the filter blocks 301 subtracted by background noise estimate of the each filter bank 301 is calculated. After calculating these measurements, a choice between ACELP and TCX excitation is made by using, for example, the following method. In the following it is assumed that when a flag is set, other flags are cleared to prevent conflicts. First, the average standard deviation value for the long window stdalong is compared with a first threshold value TH1, for example 0.4. If the standard deviation value stdalong is smaller than the first threshold value TH1, a TCX MODE flag is set. Otherwise, the calculated measurement of the low and high frequency relation LPHaF is compared with a second threshold value TH2, for example 280. If the calculated measurement of the low and high frequency relation LPHaF is greater than the second threshold value TH2, the TCX MODE flag is set. Otherwise, an inverse of the standard deviation value stdalong subtracted by the first threshold value TH1 is calculated and a first constant C1, for example 5, is summed to the calculated inverse value. The sum is compared with the calculated measurement of the low and high frequency relation LPHaF: C1+(1/( stdalong -TH1))> LPHaF (1) If the result of the comparison is true, the TCX MODE flag is set. If the result of the comparison is not true, the standard deviation value stdalong is multiplied by a first multiplicand M1 (e.g. -90) and a second constant C2 (e.g. 120) is added to the result of the multiplication. The sum is compared with the calculated measurement of the low and high frequency relation LPHaF: M1* stdalong +C2 If the sum is smaller than the calculated measurement of the low and high frequency relation LPHaF, an ACELP MODE flag is set. Otherwise an UNCERTAIN MODE flag is set indicating that the excitation method could not yet be selected for the current frame. A further examination is performed after the above described steps before the excitation method for the current frame is selected. First, it is examined whether either the ACELP MODE flag or the UNCERTAIN MODE flag is set and if the calculated average level AVL of the filter banks 301 for the current frame is greater than a third threshold value TH3 (e.g. 2000), therein the TCX MODE flag is set and the ACELP MODE flag and the UNCERTAIN MODE flag are cleared. Next, if the UNCERTAIN MODE flag is set, the similar evaluations are performed for the average standard deviation value stdashort for the short window than what was performed above for the average standard deviation value stdalong for the long window but using slightly different values for the constants and thresholds in the comparisons. If the average standard deviation value stdashort for the short window is smaller than a fourth threshold value TH4 (e.g. 0.2), the TCX MODE flag is set. Otherwise, an inverse of the standard deviation value stdashort for the short window subtracted by the fourth threshold value TH4 is calculated and a third constant C3 (e.g. 2.5) is summed to the calculated inverse value. The sum is compared with the calculated measurement of the low and high frequency relation LPHaF: C3+(1 /( stdashort-TH4)) > LPHaF (3) If the result of the comparison is true, the TCX MODE flag is set. If the result of the comparison is not true, the standard deviation value stdashort is multiplied by a second multiplicand M2 (e.g. -90) and a fourth constant C4 (e.g. 140) is added to the result of the multiplication. The sum is compared with the calculated measurement of the low and high frequency relation LPHaF: M2* stdashort+C4 If the sum is smaller than the calculated measurement of the low and high frequency relation LPHaF, the ACELP MODE flag is set. Otherwise the UNCERTAIN MODE flag is set indicating that the excitation method could not yet be selected for the current frame. At the next stage the energy levels of the current frame and the previous frame are examined. If the rate between the total energy of the current frame TotEO and the total energy of the previous frame TotE-1 is greater than a fifth threshold value TH5 (e.g. 25) the ACELP MODE flag is set and the TCX MODE flag and the UNCERTAIN MODE flag are cleared. Finally, if the TCX MODE flag or the UNCERTAIN MODE flag is set and if the calculated average level AVL of the filter banks 301 for the current frame is greater than the third threshold value TH3 and the total energy of the current frame TotEO is less than a sixth threshold value TH6 (e.g. 60) the ACELP MODE flag is set. When the above described evaluation method is performed the first excitation method and the first excitation block 206 is selected if the TCX MODE flag is set or the second excitation method and the second excitation block 207 is selected if the ACELP MODE flag is set. If, however, the UNCERTAIN MODE flag is set, the evaluation method could not perform the selection. In that case e either ACELP or TCX is selected or some further analysis have to be performed to make the differentiation. The method can also be illustrated as the following pseudo-code: if (stdalong SET TCX_MODE else if (LPHaF > TH2) SET TCX_MODE else if ((C1+(1/( stdalong -TH1))) > LPHaF) SETTCX MODE else if ((M1* stdalong +C2) SET ACELP_MODE else SET UNCERTAIN_MODE if (ACELP_MODE or UNCERTAIN_MODE) and (AVL > TH3) SET TCX_MODE if(UNCERTAIN_MODE) if (stdashort SET TCX_MODE else if ((C3+(1/( stdashort -TH4))) > LPHaF) SET TCX_MODE else if ((M2* stdashort+C4) SETACELP_MODE else SET UNCERTAIN_MODE if(UNCERTAIN_MODE) if ((TotEO/TotE-1)>TH5) SET ACELP_MODE if (TCX_MODE \|\| UNCERTAIN_MODE)) if (AVL > TH3 and TotEO SET ACELP_MODE The basic idea behind the classification is illustrated in Figures 4, 5 and 6. Fig. 4 shows an example of a plotting of standard deviation of energy levels in VAD filter banks as a function of the relation between low and high-energy components in a music signal. Each dot corresponds to a 20 ms frame taken from the long music signal containing different variations of music. The line A is fitted to approximately correspond to the upper border of the music signal area, i.e., dots to the right side of the line are not considered as music like signals in the method according to the present invention. Respectively, Fig. 5 shows an example of a plotting of standard deviation of energy levels in VAD filter banks as a function of the relation between low and high-energy components in a speech signal. Each dot corresponds to a 20 ms frame taken from the long speech signal containing different variations of speech and different talkers. The curve B is fitted to indicate approximately the lower border of the speech signal area, i.e., dots to the left side of the curve B are not considered as speech like in the method according to the present invention. As can be seen in figure 4, most of the music signal has quite small standard deviation and relatively even frequency distribution over the analysed frequencies. For the speech signal plotted in figure 5, the tendency is other way around, higher standard deviations and more low frequency components. Putting both signals into the same plot in figure 6 and fitting curves A, B to match the borders of the regions for both music and speech signals, it is quite easy to divide the most of the music signals and the most of the speech signals into different categories. The fitted curves A, B in the figures are the same than presented also in the attached pseudo-code above. The pictures demonstrate only a single standard deviation and low per high frequency values calculated by long windowing. The pseudo code contains an algorithm, which uses two different windowings, thus utilising two different versions of the mapping algorithm presented in Figures 4, 5 and 6. The area C limited by the curves A, B in Figure 6 indicates the overlapping area where further means for classifying music like and speech like signals may normally be needed. The area C can be made smaller by using different length of the analysis windows for the signal variation and combining these different measurements as it is done in our pseudo-code example. Some overlap can be allowed because some of the music signals can be efficiently coded with the compression optimised for speech and some speech signals can be efficiently coded with the compression optimised for music. In the example presented above the most optimal ACELP excitation is selected by using analysis-by-synthesis and the selection between the best ACELP-excitation and TCX-excitation is done by pre-selection. Although the invention was presented above by using two different excitation methods it is possible to use more than two different excitation methods and make the selection among them for compressing audio signals. It is also obvious that the filter 300 may divide the input signal into different frequency bands than presented above and also the number of frequency bands may be different than 12. Figure 7 depicts an example of a system in which the present invention can be applied. The system comprises one or more audio sources 701 producing speech and/or non-speech audio signals. The audio signals are converted into digital signals by an A/D-converter 702 when necessary. The digitised signals are input to an encoder 200 of a transmitting device 700 in which the compression is performed according to the present invention. The compressed signals are also quantised and encoded for transmission in the encoder 200 when necessary. A transmitter 703, for example a transmitter of a mobile communications device 700, transmits the compressed and encoded signals to a communication network 704. The signals are received from the communication network 704 by a receiver 705 of a receiving device 706. The received signals are transferred from the receiver 705 to a decoder 707 for decoding, dequantisation and decompression. The decoder 707 comprises detection means 708 to determine the compression method used in the encoder 200 for a current frame. The decoder 707 selects on the basis of the determination a first decompression means 709 or a second decompression means 710 for decompressing the current frame. The decompressed signals are connected from the decompression means 709, 710 to a filter 711 and a D/A converter 712 for converting the digital signal into analog signal. The analog signal can then be transformed to audio, for example, in a loudspeaker 713. The present invention can be implemented in different kind of systems, especially in low-rate transmission for achieving more efficient compression than in prior art systems. The encoder 200 according to the present invention can be implemented in different parts of communication systems. For example, the encoder 200 can be implemented in a mobile communication device having limited processing capabilities. It is obvious that the present invention is not solely limited to the above described embodiments but it can be modified within the scope of the appended claims. We Claim: 1. An encoder (200) for compressing audio signals in a transmitting device (700), said transmitting device (700) being configured to transmit said compressed audio signals to a receiving device (706) via a communication network (704), said encoder (200) comprising an input (201) for inputting frames of an audio signal in a frequency band, at least a first excitation block (206) for performing a first excitation for a speech like audio signal, and a second excitation block (207) for performing a second excitation for a music like audio signal, characterised in that the encoder (200) further comprises a filter (300) for dividing the frequency band into a plurality of sub bands each having a narrower bandwidth than said frequency band, and an excitation selection block (203) for selecting one excitation block among said at least first excitation block (206) and said second excitation block (207) for performing the excitation for a frame of the audio signal on the basis of the properties of the audio signal at least in one of said sub bands. 2. The encoder (200) as claimed in claim 1, wherein said filter (300) comprises filter block (301) for producing information indicative of signal energies (E(n)) of a current frame of the audio signal at least at one sub band, and that said excitation selection block (203) comprises energy determining means for determining the signal energy information of at least one sub band. 3. The encoder (200) as claimed in claim 2, wherein at least a first and a second group of sub bands are defined, said second group containing sub bands of higher frequencies than said first group, that a relation (LPH) between normalised signal energy (LevL) of said first group of sub bands and normalised signal energy (LevH) of said second group of sub bands is defined for the frames of the audio signal, and that said relation (LPH) is arranged to be used in the selection of the excitation block (206, 207). 4. The encoder (200) as claimed in claim 3, wherein one or more sub bands of the available sub bands are left outside of said first and said second group of sub bands. 5. The encoder (200) as claimed in claim 4, wherein the sub band of lowest frequencies is left outside of said first and said second group of sub bands. 6. The encoder (200) as claimed in claim 3, 4 or 5, wherein a first number of frames and a second number of frames are defined, said second number being greater than said first number, that said excitation selection block (203) comprises calculation means for calculating a first average standard deviation value (stdashort) using signal energies of the first number of frames including the current frame at each sub band and for calculating a second average standard deviation value (stdalong) using signal energies of the second number of frames including the current frame at each sub band. 7. The encoder (200) as claimed in any of the claims 1 to 6, wherein said filter (300) is a filter bank of a voice activity detector (202). 8. The encoder (200) as claimed in any of the claims 1 to 7, wherein said encoder (200) is an adaptive multi-rate wideband codec (AMR-WB). 9. The encoder (200) as claimed in any of the claims 1 to 8, wherein said first excitation is Algebraic Code Excited Linear Prediction excitation (ACELP) and said second excitation is transform coded excitation (TCX). 10. A system for compressing audio signals, said system comprising a transmitting device (700) and a receiving device (706), wherein said transmitting device (700) is configured to transmit the compressed audio signals to a receiving device (706) via a communication network (704), wherein said transmitting device comprises an encoder (200), said encoder comprising an input (201) for inputting frames of an audio signal in a frequency band at least a first excitation block (206) for performing a first excitation for a speech like audio signal, and a second excitation block (207) for performing a second excitation for a music like audio signal, characterised in that said encoder (200) further comprises a filter (300) for dividing the frequency band into a plurality of sub bands each having a narrower bandwidth than said frequency band, that the system also comprises an excitation selection block (203) for selecting one excitation block among said at least first excitation block (206) and said second excitation block (207) for performing the excitation for a frame of the audio signal on the basis of the properties of the audio signal at least at one of said sub bands. 11. The system as claimed in claim 10, wherein said filter (300) comprises filter block (301) for producing information indicative of signal energies (E(n)) of a current frame of the audio signal at least one sub band, and that said excitation selection block (203) comprises energy determining means for determining the signal energy information of at least one sub band. 12. The system as claimed in claim 11, wherein at least a first and a second group of sub bands are defined, said second group containing sub bands of higher frequencies than said first group, that a relation (LPH) between normalised signal energy (LevL) of said first group of sub bands and normalised signal energy (LevH) of said second group of sub bands is defined for the frames of the audio signal, and that said relation (LPH) is arranged to be used in the selection of the excitation block (206, 207). 13. The system as claimed in claim 12, wherein one or more sub bands of the available sub bands are left outside of said first and said second group of sub bands. 14. The system as claimed in claim 13, wherein the sub band of lowest frequencies is left outside of said first and said second group of sub bands. 15. The system as claimed in claim 12, 13 or 14, wherein a first number of frames and a second number of frames are defined, said second number being greater than said first number, that said excitation selection block (203) comprises calculation means for calculating a first average standard deviation value (stdashort) using signal energies of the first number of frames including the current frame at each sub band and for calculating a second average standard deviation value (stdalong) using signal energies of the second number of frames including the current frame at each sub band. 16. The system as claimed in any of the claims 10 to 15, wherein said filter (300) is a filter bank of a voice activity detector (202). 17. The system as claimed in any of the claims 10 to 16, wherein said encoder (200) is an adaptive multi-rate wideband codec (AMR-WB). 18. The system as claimed in any of the claims 10 to 17, wherein said first excitation is Algebraic Code Excited Linear Prediction excitation (ACELP) and said second excitation is transform coded excitation (TCX). 19. The system as claimed in any of the claims 10 to 18, wherein it is a mobile communication device. 20. The system as claimed in any of the claims 10 to 19, wherein it comprises a transmitter for transmitting frames including parameters produced by the selected excitation block (206, 207) through a low bit rate channel. 21. A method for compressing audio signals in a frequency band, wherein said compressed audio signals are transmittable from a transmitting device (700) to a receiving device (706) via a communication network (704), in which frequency band a first excitation is used for a speech like audio signal, and second excitation is used for a music like audio signal, characterised in that the frequency band is divided in said transmitting device into a plurality of sub bands each having a narrower bandwidth than said frequency band, that one excitation among said at least first excitation and said second excitation is selected for performing the excitation for a frame of the audio signal on the basis of the properties of the audio signal at least at one of said sub bands. 22. The method as claimed in claim 21, wherein said filter (300) comprises filter block (301) for producing information indicative of signal energies (E(n)) of a current frame of the audio signal at least one sub band, and that said excitation selection block (203) comprises energy determining means for determining the signal energy information of at least one sub band. 23. The method as claimed in claim 22, wherein at least a first and a second group of sub bands are defined, said second group containing sub bands of higher frequencies than said first group, that a relation (LPH) between normalised signal energy (LevL) of said first group of sub bands and normalised signal energy (LevH) of said second group of sub bands is defined for the frames of the audio signal, and that said relation (LPH) is arranged to be used in the selection of the excitation block (206, 207). 24. The method as claimed in claim 23, wherein one or more sub bands of the available sub bands are left outside of said first and said second group of sub bands. 25. The method as claimed in claim 24, wherein the sub band of lowest frequencies is left outside of said first and said second group of sub bands. 26. The method as claimed in claim 22, 23 or 24, wherein a first number of frames and a second number of frames are defined, said second number being greater than said first number, that said excitation selection block (203) comprises calculation means for calculating a first average standard deviation value (stdashort) using signal energies of the first number of frames including the current frame at each sub band and for calculating a second average standard deviation value (stdalong) using signal energies of the second number of frames including the current frame at each sub band. 27. The method as claimed in any of the claims 21 to 26, wherein said filter (300) is a filter bank of a voice activity detector (202). 28. The method as claimed in any of the claims 21 to 27, wherein said encoder (200) is an adaptive multi-rate wideband codec (AMR-WB). 29. The method as claimed in any of the claims 21 to 28, wherein said first excitation is Algebraic Code Excited Linear Prediction excitation (ACELP) and said second excitation is transform coded excitation (TCX). 30. The method as claimed in any of the claims 21 to 29, wherein frames including parameters produced by the selected excitation are transmitted through a low bit rate channel.

Full Text

Classification of audio signals
Field of the Invention
The invention relates to speech and audio coding in which the
encoding mode is changed depending whether input signal is speech
like or music like signal. The present invention relates to an encoder
comprising an input for inputting frames of an audio signal in a
frequency band, at least a first excitation block for performing a first
excitation for a speech like audio signal, and a second excitation block
for performing a second excitation for a non-speech like audio signal.
The invention also relates to a device comprising an encoder
comprising an input for inputting frames of an audio signal in a
frequency band, at least a first excitation block for performing a first
excitation for a speech like audio signal, and a second excitation block
for performing a second excitation for a non-speech like audio signal.
The invention also relates to a system comprising an encoder
comprising an input for inputting frames of an audio signal in a
frequency band, at least a first excitation block for performing a first
excitation for a speech like audio signal, and a second excitation block
for performing a second excitation for a non-speech like audio signal.
The invention further relates to a method for compressing audio signals
in a frequency band, in which a first excitation is used for a speech like
audio signal, and second excitation is used for a non-speech like audio
signal . The invention relates to a module for classifying frames of an
audio signal in a frequency band for selection of an excitation among at
least a first excitation for a speech like audio signal, and a second
excitation fora non-speech like audio signal. The invention relates to a
computer program product comprising machine executable steps for
compressing audio signals in a frequency band, in which a first
excitation is used for a speech like audio signal, and second excitation
is used for a non-speech like audio signal.
Background of the Invention
In many audio signal processing applications audio signals are
compressed to reduce the processing power requirements when
processing the audio signal. For example, in digital communication
systems audio signal is typically captured as an analogue signal,
digitised in an analogue to digital (A/D) converter and then encoded
before transmission over a wireless air interface between a user
equipment, such as a mobile station, and a base station. The purpose
of the encoding is to compress the digitised signal and transmit it over
the air interface with the minimum amount of data whilst maintaining an
acceptable signal quality level. This is particularly important as radio
channel capacity over the wireless air interface is limited in a cellular
communication network. There are also applications in which digitised
audio signal is stored to a storage medium for later reproduction of the
audio signal.
The compression can be lossy or lossless. In lossy compression some
information is lost during the compression wherein it is not possible to
fully reconstruct the original signal from the compressed signal. In
lossless compression no information is normally lost. Hence, the
original signal can usually be completely reconstructed from the
compressed signal.
The term audio signal is normally understood as a signal containing
speech, music (non-speech) or both. The different nature of speech
and music makes it rather difficult to design one compression algorithm
which works enough well for both speech and music. Therefore, the
problem is often solved by designing different algorithms for both audio
and speech and use some kind of recognition method to recognise
whether the audio signal is speech like or music like and select the
appropriate algorithm according to the recognition.
In overall, classifying purely between speech and music or non-speech
signals is a difficult task. The required accuracy depends heavily on the
application. In some applications the accuracy is more critical like in
speech recognition or in accurate archiving for storage and retrieval
purposes. However, the situation is a bit different if the classification is
used for selecting optimal compression method for the input signal. In
this case, it may happen that there does not exist one compression
method that is always optimal for speech and another method that is
always optimal for music or non-speech signals. In practise, it may be
that a compression method for speech transients is also very efficient
for music transients. It is also possible that a music compression for
strong tonal components may be good for voiced speech segments.
So, in these instances, methods for classifying just purely for speech
and music do not create the most optimal algorithm to select the best
compression method.
Often speech can be considered as bandlimited to between
approximately 200Hz and 3400 Hz. The typical sampling rate used by
an A/D converter to convert an analogue speech signal into a digital
signal is either 8kHz or 16kHz. Music or non-speech signals may
contain frequency components well above the normal speech
bandwidth. In some applications the audio system should be able to
handle a frequency band between about 20 Hz to 20 000 kHz. The
sample rate for that kind of signals should be at least 40 000 kHz to
avoid aliasing. It should be noted here that the above mentioned values
are just non-limiting examples. For example, in some systems the
higher limit for music signals may be about 10 000 kHz or even less
than that.
The sampled digital signal is then encoded, usually on a frame by
frame basis, resulting in a digital data stream with a bit rate that is
determined by a codec used for encoding. The higher the bit rate, the
more data is encoded, which results in a more accurate representation
of the input frame. The encoded audio signal can then be decoded and
passed through a digital to analogue (D/A) converter to reconstruct a
signal which is as near the original signal as possible.
An ideal codec will encode the audio signal with as few bits as possible
thereby optimising channel capacity, while producing decoded audio
signal that sounds as close to the original audio signal as possible. In
practice there is usually a trade-off between the bit rate of the codec
and the quality of the decoded audio.
At present there are numerous different codecs, such as the adaptive
multi-rate (AMR) codec and the adaptive multi-rate wideband (AMRWB)
codec, which are developed for compressing and encoding audio
signals. AMR was developed by the 3rd Generation Partnership Project
(3GPP) for GSM/EDGE and WCDMA communication networks. In
addition, it has also been envisaged that AMR will be used in packet
switched networks. AMR is based on Algebraic Code Excited Linear
Prediction (ACELP} coding. The AMR and AMR WB codecs consist of
8 and 9 active bit rates respectively and aiso include voice activity
detection (VAD) and discontinuous transmission (DTX) functionality. At
the moment, the sampling rate in the AMR codec is 8 kHz and in the
AMR WB codec the sampling rate is 16kHz. It is obvious that the
codecs and sampling rates mentioned above are just non-limiting
examples.
ACELP coding operates using a model of how the signal source is
generated, and extracts from the signal the parameters of the model
More specifically, ACELP coding is based on a mode! of the human
vocal system, where the throat and mouth are modelled as a linear
filter and speech is generated by a periodic vibration of air exciting the
filter. The speech is analysed on a frame by frame basis by the
encoder and for each frame a set of parameters representing the
modelled speech is generated and output by the encoder. The set of
parameters may include excitation parameters and the coefficients for
the filter as well as other parameters. The output from a speech
encoder is often referred to as a parametric representation of the input
speech signal. The set of parameters is then used by a suitably
configured decoder to regenerate the input speech signal.
For some input signals, the pulse-like ACELP-excitation produces
higher quality and for some input signals transform coded excitation
(TCX) is more optimal. !t is assumed here that ACELP-excitation is
mostly used for typical speech content as an input signal and TCXexcitation
is mostly used for typical music as an input signal. However,
this is not always the case, i.e., sometimes speech signal has parts,
which are music like and music signal has parts, which are speech like.
The definition of speech like signal in this application is that most of the
speech belongs to this category and some of the music may also
belong to this category. For music like signals the definition is other
way around. Additionally, there are some speech signal parts and
music signal parts that are neutral in a sense that they can belong to
the both classes.
The selection of excitation can be done in several ways: the most
complex and quite good method is to encode both ACELP and TCXexcitation
and then select the best excitation based on the synthesised
speech signal. This analysis-by-synthesis type of method will provide
good results but it is in some applications not practical because of its
high complexity. In this method for example SNR-type of algorithm can
be used to measure the quality produced by both excitations. This
method can be called as a "brute-force" method because it tries all the
combinations of different excitations and selects afterwards the best
one. The less complex method would perform the synthesis only once
by analysing the signal properties beforehand and then selecting the
best excitation. The method can also be a combination of pre-selection
and "brute-force" to make compromised between quality and
complexity.
Figure 1 presents a simplified encoder 100 with prior-art high
complexity classification. An audio signal is input to the input signal
block 101 in which the signal is digitised and filtered. The input signal
block 101 also forms frames from the digitised and filtered signal. The
frames are input to a linear prediction coding (LPC) analysis block 102.
It performs a LPC analysis on the digitised input signal on a frame by
frame basis to find such a parameter set which matches best with the
input signal. The determined parameters (LPC parameters) are
quantized and output 109 from the encoder 100. The encoder 100 also
generates two output signals with LPC synthesis blocks 103, 104. The
first LPC synthesis block 103 uses a signal generated by the TCX
excitation block 105 to synthesise the audio signal for finding the code
vector producing the best result for the TCX excitation. The second
LPC synthesis block 104 uses a signal generated by the ACELP
excitation block 106 to synthesise the audio signal for finding the code
vector producing the best result for the ACELP excitation. In the
excitation selection block 107 the signals generated by the LPC
synthesis blocks 103, 104 are compared to determine which one of the
excitation methods gives the best (optimal) excitation. Information
about the selected excitation method and parameters of the selected
excitation signal are, for example, quantized and channel coded 108
before outputting 109 the signals from the encoder 100 for
transmission.
Summary of the Invention
One aim of the present invention is to provide an improved method for
classifying speech like and music like signals utilising frequency
information of the signal. There are music like speech signal segments
and vice versa and there are signal segments in speech and in music
that can belong to either class. In other words, the invention does not
purely classify between speech and music. However, it defines means
for categorize input signal into music like and speech like components
according to some criteria. The classification information can be used
e.g. in a multimode encoder for selecting an encoding mode.
The invention is based on the idea that input signal is divided into
several frequency bands and the relations between lower and higher
frequency bands are analysed together with the energy level variations
in those bands and the signal is classified into music like or speech like
based on both of the calculated measurements or several different
combinations of those measurements using different analysis windows
and decision threshold values. This information can then be utilised for
example in the selection of the compression method for the analysed
signal.
The encoder according to the present invention is primarily
characterised in that the encoder further comprises a filter for dividing
the frequency band into a plurality of sub bands each having a
narrower bandwidth than said frequency band, and an excitation
selection block for selecting one excitation block among said at least
first excitation block and said second excitation block for performing the
excitation for a frame of the audio signal on the basis of the properties
of the audio signal at least at one of said sub bands.
The device according to the present invention is primarily characterised
in that said encoder comprises a filter for dividing the frequency band
into a plurality of sub bands each having a narrower bandwidth than
said frequency band, that the device also comprises an excitation
selection block for selecting one excitation block among said at least
first excitation block and said second excitation block for performing the
excitation for a frame of the audio signal on the basis of the properties
of the audio signal at least at one of said sub bands.
The system according to the present invention is primarily
characterised in that said encoder further comprises a filter for dividing
the frequency band into a plurality of sub bands each having a
narrower bandwidth than said frequency band, that the system also
comprises an excitation selection block for selecting one excitation
block among said at least first excitation block and said second
excitation block for performing the excitation for a frame of the audio
signal on the basis of the properties of the audio signal at least at one
of said sub bands.
The method according to the present invention is primarily
characterised in that the frequency band is divided into a plurality of
sub bands each having a narrower bandwidth than said frequency
band, that one excitation among said at least first excitation and said
second excitation is selected for performing the excitation for a frame
of the audio signal on the basis of the properties of the audio signal at
least at one of said sub bands.
The module according to the present invention is primarily
characterised in that the module further comprises input for inputting
information indicative of the frequency band divided into a plurality of
sub bands each having a narrower bandwidth than said frequency
band, and an excitation selection block for selecting one excitation
block among said at least first excitation block and said second
excitation block for performing the excitation for a frame of the audio
signal on the basis of the properties of the audio signal at least at one
of said sub bands.
The computer program product according to the present invention is
primarily characterised in that the computer program product further
comprises machine executable steps for dividing the frequency band
into a plurality of sub bands each having a narrower bandwidth than
said frequency band, machine executable steps for selecting one
excitation among said at least first excitation and said second
excitation on the basis of the properties of the audio signal at least at
one of said sub bands for performing the excitation for a frame of the
audio signal.
In this application, terms "speech like" and "music like" are defined to
separate the invention from the typical speech and music
classifications. Even if around 90% of the speech were categorized as
speech like in a system according to the present invention, the rest of
the speech signal may be defined as a music like signal, which may
improve audio quality if the selection of the compression algorithm is
based on this classification. Also typical music signals may fall in 80-90
% of the cases into music like signals but classifying part of the music
signal into speech like category will improve the quality of the sound
signal for the compression system. Therefore, the present invention
provides advantages when compared with prior art methods and
systems. By using the classification method according to the present
invention it is possible to improve reproduced sound quality without
greatly affecting the compression efficiency.
Compared to the brute-force approach presented above, the invention
provides a much less complex pre-selection type approach to make
selection between two excitation types. The invention divides input
signal into frequency bands and analyses the relations between lower
and higher frequency bands together and can also use, for example,
the energy level variations in those bands and classifies the signal into
music like or speech like.
Description of the Drawings
Fig. 1 presents a simplified encoder with prior-art high complexity
classification,
Fig. 2 presents an example embodiment of an encoder with
classification according to the invention,
Fig. 3 illustrates an exampleof a VAD filter bank structure in AMRWB
VAD algorithm,
Fig. 4 shows an example of a plotting of standard deviation of
energy levels in VAD filter banks as a function of the
relation between low and high-energy components in a
music signal,
Fig. 5 shows an example of a plotting of standard deviation of
energy levels in VAD filter banks as a function of the
relation between low- and high-energy components in a
speech signal,
Fig. 6 shows an example of a combined plotting for both music
and speech signals, and
Fig. 7 shows an example of a system according to the present
invention.
Detailed Description of the Invention
In the following an encoder 200 according to an example embodiment
of the present invention will be described in more detail with reference
to Fig. 2. The encoder 200 comprises an input block 201 for digitizing,
filtering and framing the input signal when necessary. It should be
noted here that the input signal may already be in a form suitable for
the encoding process. For example, the input signal may have been
digitised at an earlier stage and stored to a memory medium (not
shown). The input signal frames are input to a voice activity detection
block 202. The voice activity detection block 202 outputs a multiplicity
of narrower band signals which are input to an excitation selection
block 203. The excitation selection block 203 analyses the signals to
determine which excitation method is the most appropriate one for
encoding the input signal. The excitation selection block 203 produces
a control signal 204 for controlling a selection means 205 according to
the determination of the excitation method. If it was determined that the
best excitation method for encoding the current frame of the input
signal is a first excitation method, the selection means 205 are
controlled to select the signal of a first excitation block 206. If it was
determined that the best excitation method for encoding the current
frame of the input signal is a second excitation method, the selection
means 205 are controlled to select the signal of a second excitation
block 207. Although the encoder of Fig. 2 has only the first 206 and the
second excitation block 207 for the encoding process, it is obvious that
there can also be more than two different excitation blocks for different
excitation methods available in the encoder 200 to be used in the
encoding of the input signal.
The first excitation block 206 produces, for example, a TCX excitation
signal and the second excitation block 207 produces, for example, a
ACELP excitation signal.
The LPC analysis block 208 performs a LPC analysis on the digitised
input signal on a frame by frame basis to find such a parameter set
which matches best with the input signal.
LPC parameters 210 and excitation parameters 211 are, for example,
quantised and encoded in a quantisation and encoding block 212
before transmission e.g. to a communication network 704 (Fig. 7).
However, it is not necessary to transmit the parameters but they can,
for example, be stored on a storage medium and at a later stage
retrieved for transmission and/or decoding.
Fig. 3 depicts one example of a filter 300 which can be used in the
encoder 200 for the signal analysis. The filter 300 is, for example, a
filter bank of the voice activity detection block of the AMR-WB codec,
wherein a separate filter is not needed but it is also possible to use
other filters for this purpose. The filter 300 comprises two or more filter
blocks 301 to divide the input signal into two or more subband signals
on different frequencies. In other words, each output signal of the filter
300 represents a certain frequency band of the input signal. The output
signals of the filter 300 can be used in the excitation selection block
203 to determine the frequency content of the input signal.
The excitation selection block 203 evaluates energy levels of each
output of the filter bank 300 and analyses the relations between lower
and higher frequency subbands together with the energy level
variations in those subbands and classifies the signal into music like or
speech like.
The invention is based on examining the frequency content of the input
signal to select the excitation method for frames of the input signal. In
the following, AMR-WB extension (AMR-WB+) is used as a practical
example used to classify input signal into speech like or music like
signals and to select either ACELP- or TCX-excitation for those signal
respectively. However, the invention is not limited to AMR-WB codecs
or ACELP- and TCX- excitation methods.
In the extended AMR-WB (AMR-WB+) codec, there are two types of
excitation for LP-synthesis: ACELP pulse-like excitation and transform
coded excitation (TCX). ACELP excitation is the same than used
already in the original 3GPP AMR-WB standard (3GPP TS 26.190) and
TCX is an improvement implemented in the extended AMR-WB.
AMR-WB extension example is based on the AMR-WB VAD filter
banks, which for each 20 ms input frame, produces signal energy E(n)
in the 12 subbands over the frequency range from 0 to 6400 Hz as
shown in Fig. 3. The bandwidths of the filter banks are normally not
equal but may vary on different bands as can be seen on Fig. 3. Also
the number of subbands may vary and the subbands may be partly
overlapping. Then energy levels of each subband are normalised by
dividing the energy level E(n) from each subband by the width of that
subband (in Hz) producing normalised EN(n) energy levels of each
band where n is the band number from 0 to 11. Index 0 refers to the
lowest subband shown in Fig. 3.
In the excitation selection block 203 the standard deviation of the
energy levels is calculated for each of the 12 subbands using e.g. two
windows: a short window stdshort(n) and a long window stdlong(n). For
AMR-WB+ case, the length of the short window is 4 frames and the
long window is 16 frames. In these calculations, the 12 energy levels
from the current frame together with past 3 or 15 frames are used to
derive these two standard deviation values. The special feature of this
calculation is that it is only performed when voice activity detection
block 202 indicates 213 active speech. This will make the algorithm
react faster especially after long speech pauses.
Then, for each frame, the average standard deviation over all the 12
filter banks are taken for both long and short window and average
standard deviation values stdashort and stdalong are created.
For frames of the audio signal, also a relation between lower frequency
bands and higher frequency bands are calculated. In AMR-WB+
energy of lower frequency subbands LevL from 1 to 7 are taken and
normalised by dividing it by the length (bandwidth) of these subbands
(in Hz). For higher frequency bands from 8 to 11 energy of them are
taken and normalised respectively to create LevH. Note that in this
example embodiment the lowest subband 0 is not used in these
calculations because it usually contains so much energy that it will
distort the calculations and make the contributions from other
subbands too small. From these measurements the relation LPH =
LevL / LevH is defined. In addition, for each frame a moving average
LPHa is calculated using the current and 3 past LPH values. After
these calculations a measurement of the low and high frequency
relation LPHaF for the current frame is calculated by using weighted
sum of the current and 7 past moving average LPHa values by setting
slightly more weighting for the latest values.
It is also possible to implement the present invention so that only one
or few of the available subbands are analysed.
Also average level AVL of the filter blocks 301 for the current frame is
calculated by subtracting the estimated level of background noise from
each filter block output, and summing these levels multiplied by the
highest frequency of the corresponding filter block 301, to balance the
high frequency subbands containing relatively less energy than the
lower frequency subbands.
Also the total energy of the current frame TotEO from all the filter blocks
301 subtracted by background noise estimate of the each filter bank
301 is calculated.
After calculating these measurements, a choice between ACELP and
TCX excitation is made by using, for example, the following method. In
the following it is assumed that when a flag is set, other flags are
cleared to prevent conflicts. First, the average standard deviation value
for the long window stdalong is compared with a first threshold value
TH1, for example 0.4. If the standard deviation value stdalong is
smaller than the first threshold value TH1, a TCX MODE flag is set.
Otherwise, the calculated measurement of the low and high frequency
relation LPHaF is compared with a second threshold value TH2, for
example 280.
If the calculated measurement of the low and high frequency relation
LPHaF is greater than the second threshold value TH2, the TCX
MODE flag is set. Otherwise, an inverse of the standard deviation
value stdalong subtracted by the first threshold value TH1 is calculated
and a first constant C1, for example 5, is summed to the calculated
inverse value. The sum is compared with the calculated measurement
of the low and high frequency relation LPHaF:
C1+(1/( stdalong -TH1))> LPHaF (1)
If the result of the comparison is true, the TCX MODE flag is set. If the
result of the comparison is not true, the standard deviation value
stdalong is multiplied by a first multiplicand M1 (e.g. -90) and a second
constant C2 (e.g. 120) is added to the result of the multiplication. The
sum is compared with the calculated measurement of the low and high
frequency relation LPHaF:
M1* stdalong +C2 If the sum is smaller than the calculated measurement of the low and
high frequency relation LPHaF, an ACELP MODE flag is set. Otherwise
an UNCERTAIN MODE flag is set indicating that the excitation method
could not yet be selected for the current frame.
A further examination is performed after the above described steps
before the excitation method for the current frame is selected. First, it is
examined whether either the ACELP MODE flag or the UNCERTAIN
MODE flag is set and if the calculated average level AVL of the filter
banks 301 for the current frame is greater than a third threshold value
TH3 (e.g. 2000), therein the TCX MODE flag is set and the ACELP
MODE flag and the UNCERTAIN MODE flag are cleared.
Next, if the UNCERTAIN MODE flag is set, the similar evaluations are
performed for the average standard deviation value stdashort for the
short window than what was performed above for the average standard
deviation value stdalong for the long window but using slightly different
values for the constants and thresholds in the comparisons. If the
average standard deviation value stdashort for the short window is
smaller than a fourth threshold value TH4 (e.g. 0.2), the TCX MODE
flag is set. Otherwise, an inverse of the standard deviation value
stdashort for the short window subtracted by the fourth threshold value
TH4 is calculated and a third constant C3 (e.g. 2.5) is summed to the
calculated inverse value. The sum is compared with the calculated
measurement of the low and high frequency relation LPHaF:
C3+(1 /( stdashort-TH4)) > LPHaF (3)
If the result of the comparison is true, the TCX MODE flag is set. If the
result of the comparison is not true, the standard deviation value
stdashort is multiplied by a second multiplicand M2 (e.g. -90) and a
fourth constant C4 (e.g. 140) is added to the result of the multiplication.
The sum is compared with the calculated measurement of the low and
high frequency relation LPHaF:
M2* stdashort+C4 If the sum is smaller than the calculated measurement of the low and
high frequency relation LPHaF, the ACELP MODE flag is set.
Otherwise the UNCERTAIN MODE flag is set indicating that the
excitation method could not yet be selected for the current frame.
At the next stage the energy levels of the current frame and the
previous frame are examined. If the rate between the total energy of
the current frame TotEO and the total energy of the previous frame
TotE-1 is greater than a fifth threshold value TH5 (e.g. 25) the ACELP
MODE flag is set and the TCX MODE flag and the UNCERTAIN MODE
flag are cleared.
Finally, if the TCX MODE flag or the UNCERTAIN MODE flag is set
and if the calculated average level AVL of the filter banks 301 for the
current frame is greater than the third threshold value TH3 and the total
energy of the current frame TotEO is less than a sixth threshold value
TH6 (e.g. 60) the ACELP MODE flag is set.
When the above described evaluation method is performed the first
excitation method and the first excitation block 206 is selected if the
TCX MODE flag is set or the second excitation method and the second
excitation block 207 is selected if the ACELP MODE flag is set. If,
however, the UNCERTAIN MODE flag is set, the evaluation method
could not perform the selection. In that case e either ACELP or TCX is
selected or some further analysis have to be performed to make the
differentiation.
The method can also be illustrated as the following pseudo-code:
if (stdalong SET TCX_MODE
else if (LPHaF > TH2)
SET TCX_MODE
else if ((C1+(1/( stdalong -TH1))) > LPHaF)
SETTCX MODE
else if ((M1* stdalong +C2) SET ACELP_MODE
else
SET UNCERTAIN_MODE
if (ACELP_MODE or UNCERTAIN_MODE) and (AVL > TH3)
SET TCX_MODE
if(UNCERTAIN_MODE)
if (stdashort SET TCX_MODE
else if ((C3+(1/( stdashort -TH4))) > LPHaF)
SET TCX_MODE
else if ((M2* stdashort+C4) SETACELP_MODE
else
SET UNCERTAIN_MODE
if(UNCERTAIN_MODE)
if ((TotEO/TotE-1)>TH5)
SET ACELP_MODE
if (TCX_MODE || UNCERTAIN_MODE))
if (AVL > TH3 and TotEO SET ACELP_MODE
The basic idea behind the classification is illustrated in Figures 4, 5 and
6. Fig. 4 shows an example of a plotting of standard deviation of
energy levels in VAD filter banks as a function of the relation between
low and high-energy components in a music signal. Each dot
corresponds to a 20 ms frame taken from the long music signal
containing different variations of music. The line A is fitted to
approximately correspond to the upper border of the music signal area,
i.e., dots to the right side of the line are not considered as music like
signals in the method according to the present invention.
Respectively, Fig. 5 shows an example of a plotting of standard
deviation of energy levels in VAD filter banks as a function of the
relation between low and high-energy components in a speech signal.
Each dot corresponds to a 20 ms frame taken from the long speech
signal containing different variations of speech and different talkers.
The curve B is fitted to indicate approximately the lower border of the
speech signal area, i.e., dots to the left side of the curve B are not
considered as speech like in the method according to the present
invention.
As can be seen in figure 4, most of the music signal has quite small
standard deviation and relatively even frequency distribution over the
analysed frequencies. For the speech signal plotted in figure 5, the
tendency is other way around, higher standard deviations and more
low frequency components. Putting both signals into the same plot in
figure 6 and fitting curves A, B to match the borders of the regions for
both music and speech signals, it is quite easy to divide the most of the
music signals and the most of the speech signals into different
categories. The fitted curves A, B in the figures are the same than
presented also in the attached pseudo-code above. The pictures
demonstrate only a single standard deviation and low per high
frequency values calculated by long windowing. The pseudo code
contains an algorithm, which uses two different windowings, thus
utilising two different versions of the mapping algorithm presented in
Figures 4, 5 and 6.
The area C limited by the curves A, B in Figure 6 indicates the
overlapping area where further means for classifying music like and
speech like signals may normally be needed. The area C can be made
smaller by using different length of the analysis windows for the signal
variation and combining these different measurements as it is done in
our pseudo-code example. Some overlap can be allowed because
some of the music signals can be efficiently coded with the
compression optimised for speech and some speech signals can be
efficiently coded with the compression optimised for music.
In the example presented above the most optimal ACELP excitation is
selected by using analysis-by-synthesis and the selection between the
best ACELP-excitation and TCX-excitation is done by pre-selection.
Although the invention was presented above by using two different
excitation methods it is possible to use more than two different
excitation methods and make the selection among them for
compressing audio signals. It is also obvious that the filter 300 may
divide the input signal into different frequency bands than presented
above and also the number of frequency bands may be different than
12.
Figure 7 depicts an example of a system in which the present invention
can be applied. The system comprises one or more audio sources 701
producing speech and/or non-speech audio signals. The audio signals
are converted into digital signals by an A/D-converter 702 when
necessary. The digitised signals are input to an encoder 200 of a
transmitting device 700 in which the compression is performed
according to the present invention. The compressed signals are also
quantised and encoded for transmission in the encoder 200 when
necessary. A transmitter 703, for example a transmitter of a mobile
communications device 700, transmits the compressed and encoded
signals to a communication network 704. The signals are received from
the communication network 704 by a receiver 705 of a receiving device
706. The received signals are transferred from the receiver 705 to a
decoder 707 for decoding, dequantisation and decompression. The
decoder 707 comprises detection means 708 to determine the
compression method used in the encoder 200 for a current frame. The
decoder 707 selects on the basis of the determination a first
decompression means 709 or a second decompression means 710 for
decompressing the current frame. The decompressed signals are
connected from the decompression means 709, 710 to a filter 711 and
a D/A converter 712 for converting the digital signal into analog signal.
The analog signal can then be transformed to audio, for example, in a
loudspeaker 713.
The present invention can be implemented in different kind of systems,
especially in low-rate transmission for achieving more efficient
compression than in prior art systems. The encoder 200 according to
the present invention can be implemented in different parts of
communication systems. For example, the encoder 200 can be
implemented in a mobile communication device having limited
processing capabilities.
It is obvious that the present invention is not solely limited to the above
described embodiments but it can be modified within the scope of the
appended claims.

We Claim:
1. An encoder (200) for compressing audio signals in a transmitting device (700), said transmitting device (700) being configured to transmit said compressed audio signals to a receiving device (706) via a communication network (704), said encoder (200) comprising an input (201) for inputting frames of an audio signal in a frequency band, at least a first excitation block (206) for performing a first excitation for a speech like audio signal, and a second excitation block (207) for performing a second excitation for a music like audio signal, characterised in that the encoder (200) further comprises a filter (300) for dividing the frequency band into a plurality of sub bands each having a narrower bandwidth than said frequency band, and an excitation selection block (203) for selecting one excitation block among said at least first excitation block (206) and said second excitation block (207) for performing the excitation for a frame of the audio signal on the basis of the properties of the audio signal at least in one of said sub bands.
2. The encoder (200) as claimed in claim 1, wherein said filter (300) comprises filter block (301) for producing information indicative of signal energies (E(n)) of a current frame of the audio signal at least at one sub band, and that said excitation selection block (203) comprises energy determining means for determining the signal energy information of at least one sub band.
3. The encoder (200) as claimed in claim 2, wherein at least a first and a second group of sub bands are defined, said second group containing sub bands of higher frequencies than said first group, that a relation (LPH) between normalised signal energy (LevL) of said first group of sub bands and normalised signal energy (LevH) of said second group of sub bands is defined for the frames of the audio signal, and that said relation (LPH) is arranged to be used in the selection of the excitation block (206, 207).
4. The encoder (200) as claimed in claim 3, wherein one or more sub bands of the available sub bands are left outside of said first and said second group of sub bands.
5. The encoder (200) as claimed in claim 4, wherein the sub band of lowest frequencies is left outside of said first and said second group of sub bands.

6. The encoder (200) as claimed in claim 3, 4 or 5, wherein a first number of frames and a second number of frames are defined, said second number being greater than said first number, that said excitation selection block (203) comprises calculation means for calculating a first average standard deviation value (stdashort) using signal energies of the first number of frames including the current frame at each sub band and for calculating a second average standard deviation value (stdalong) using signal energies of the second number of frames including the current frame at each sub band.
7. The encoder (200) as claimed in any of the claims 1 to 6, wherein said filter (300) is a filter bank of a voice activity detector (202).
8. The encoder (200) as claimed in any of the claims 1 to 7, wherein said encoder (200) is an adaptive multi-rate wideband codec (AMR-WB).
9. The encoder (200) as claimed in any of the claims 1 to 8, wherein said first excitation is Algebraic Code Excited Linear Prediction excitation (ACELP) and said second excitation is transform coded excitation (TCX).
10. A system for compressing audio signals, said system comprising a transmitting device (700) and a receiving device (706), wherein said transmitting device (700) is configured to transmit the compressed audio signals to a receiving device (706) via a communication network (704), wherein said transmitting device comprises an encoder (200), said encoder comprising an input (201) for inputting frames of an audio signal in a frequency band at least a first excitation block (206) for performing a first excitation for a speech like audio signal, and a second excitation block (207) for performing a second excitation for a music like audio signal, characterised in that said encoder (200) further comprises a filter (300) for dividing the frequency band into a plurality of sub bands each having a narrower bandwidth than said frequency band, that the system also comprises an excitation selection block (203) for selecting one excitation block among said at least first excitation block (206) and said second excitation block (207) for performing the excitation for a frame of the audio signal on the basis of the properties of the audio signal at least at one of said sub bands.

11. The system as claimed in claim 10, wherein said filter (300) comprises filter block (301) for producing information indicative of signal energies (E(n)) of a current frame of the audio signal at least one sub band, and that said excitation selection block (203) comprises energy determining means for determining the signal energy information of at least one sub band.
12. The system as claimed in claim 11, wherein at least a first and a second group of sub bands are defined, said second group containing sub bands of higher frequencies than said first group, that a relation (LPH) between normalised signal energy (LevL) of said first group of sub bands and normalised signal energy (LevH) of said second group of sub bands is defined for the frames of the audio signal, and that said relation (LPH) is arranged to be used in the selection of the excitation block (206, 207).
13. The system as claimed in claim 12, wherein one or more sub bands of the available sub bands are left outside of said first and said second group of sub bands.
14. The system as claimed in claim 13, wherein the sub band of lowest frequencies is left outside of said first and said second group of sub bands.
15. The system as claimed in claim 12, 13 or 14, wherein a first number of frames and a second number of frames are defined, said second number being greater than said first number, that said excitation selection block (203) comprises calculation means for calculating a first average standard deviation value (stdashort) using signal energies of the first number of frames including the current frame at each sub band and for calculating a second average standard deviation value (stdalong) using signal energies of the second number of frames including the current frame at each sub band.
16. The system as claimed in any of the claims 10 to 15, wherein said filter (300) is a filter bank of a voice activity detector (202).
17. The system as claimed in any of the claims 10 to 16, wherein said encoder (200) is an adaptive multi-rate wideband codec (AMR-WB).

18. The system as claimed in any of the claims 10 to 17, wherein said first excitation is Algebraic Code Excited Linear Prediction excitation (ACELP) and said second excitation is transform coded excitation (TCX).
19. The system as claimed in any of the claims 10 to 18, wherein it is a mobile communication device.
20. The system as claimed in any of the claims 10 to 19, wherein it comprises a transmitter for transmitting frames including parameters produced by the selected excitation block (206, 207) through a low bit rate channel.
21. A method for compressing audio signals in a frequency band, wherein said compressed audio signals are transmittable from a transmitting device (700) to a receiving device (706) via a communication network (704), in which frequency band a first excitation is used for a speech like audio signal, and second excitation is used for a music like audio signal, characterised in that the frequency band is divided in said transmitting device into a plurality of sub bands each having a narrower bandwidth than said frequency band, that one excitation among said at least first excitation and said second excitation is selected for performing the excitation for a frame of the audio signal on the basis of the properties of the audio signal at least at one of said sub bands.
22. The method as claimed in claim 21, wherein said filter (300) comprises filter block (301) for producing information indicative of signal energies (E(n)) of a current frame of the audio signal at least one sub band, and that said excitation selection block (203) comprises energy determining means for determining the signal energy information of at least one sub band.
23. The method as claimed in claim 22, wherein at least a first and a second group of sub bands are defined, said second group containing sub bands of higher frequencies than said first group, that a relation (LPH) between normalised signal energy (LevL) of said first group of sub bands and normalised signal energy (LevH) of said second group of sub bands is defined for the frames of the audio signal, and that said relation (LPH) is arranged to be used in the selection of the excitation block (206, 207).

24. The method as claimed in claim 23, wherein one or more sub bands of the available sub bands are left outside of said first and said second group of sub bands.
25. The method as claimed in claim 24, wherein the sub band of lowest frequencies is left outside of said first and said second group of sub bands.
26. The method as claimed in claim 22, 23 or 24, wherein a first number of frames and a second number of frames are defined, said second number being greater than said first number, that said excitation selection block (203) comprises calculation means for calculating a first average standard deviation value (stdashort) using signal energies of the first number of frames including the current frame at each sub band and for calculating a second average standard deviation value (stdalong) using signal energies of the second number of frames including the current frame at each sub band.
27. The method as claimed in any of the claims 21 to 26, wherein said filter (300) is a filter bank of a voice activity detector (202).
28. The method as claimed in any of the claims 21 to 27, wherein said encoder (200) is an adaptive multi-rate wideband codec (AMR-WB).
29. The method as claimed in any of the claims 21 to 28, wherein said first excitation is Algebraic Code Excited Linear Prediction excitation (ACELP) and said second excitation is transform coded excitation (TCX).
30. The method as claimed in any of the claims 21 to 29, wherein frames including parameters produced by the selected excitation are transmitted through a low bit rate channel.

Documents:

http://ipindiaonline.gov.in/patentsearch/GrantedSearch/viewdoc.aspx?id=r0rEFrJ1BeJwnuPt+LTuRQ==&loc=+mN2fYxnTC4l0fUd8W4CAA==

« Previous Patent

Next Patent »

Patent Number

279527

Indian Patent Application Number

4817/DELNP/2006

PG Journal Number

04/2017

Publication Date

27-Jan-2017

Grant Date

24-Jan-2017

Date of Filing

23-Aug-2006

Name of Patentee

NOKIA TECHNOLOGIES OY

Applicant Address

KEILALAHDENTIE 4, FIN-02150 ESPOO, FINLAND.

Inventors:

#	Inventor's Name	Inventor's Address
1	VAINIO, JANNE	KOKONKATU 15, FI-33960 PIRKKALA, FINLAND
2	MIKKOLA, HANNU	IPPISENKATU 15, FI-33300 TAMPERE, FINLAND
3	OJALA, PASI	NEIDONKALLIONTIE 9, FI-02400 KIRKKONUMMI, FINLAND
4	MAKINEN, JARI	ETUNIITYNKATU 4 AS. 2, FI-33580 TAMPERE, FINLAND

PCT International Classification Number

G10L 19/14

PCT International Application Number

PCT/FI2005/050035

PCT International Filing date

2005-02-16

PCT Conventions:

#	PCT Application Number	Date of Convention	Priority Country
1	20045051	2004-02-23	Finland