CN1275223C - A low bit-rate speech coder - Google Patents

Abstract

The invention discloses a speech coder suitable for a communication system requiring low-bit-variable-rate speech coding. It applies the SVM method to the VAD voice activation detection, which improves the correct recognition rate of the voice coder for voice detection; adopts the voice mode classification method of GSM, and merges the original four voice modes into three voice modes, so that only Two bits are used to represent the entire speech pattern. It also takes full advantage of the high coding gain of the local cosine transform, uses LCT and SVM-VAD for low-bit variable-rate speech coding, and provides a practical and high-performance low-bit variable-rate speech coder.

Description

A Low Bit Variable Rate Speech Coder

Technical field

The present invention relates to a speech coder, in particular to a speech coder suitable for a communication system requiring low bit variable rate speech coding.

Background technique

Variable bit rate (Variable Bit Rate, VBR) coding technology has been developed in recent years. Its core idea is to use different rates to encode speech jumps, steady, and silent segments, so that the average rate of VBR coding will be higher than that of the same speech quality. The FBR encoding is much lower.

The application field that can really give full play to the advantages of VBR technology is the occasion where there is no strict rate limit on the speech coding rate, but the rate "elasticity" is required, such as CDMA, VoIP, ATM, etc. At present, the wireless communication system and IP technology are developing rapidly, and will soon occupy an increasingly important position in the global communication system. To this end, ITU-T SG16 is developing a new variable rate coding standard to adapt to future packet voice communication networks (such as VoIP), IMT-2000 voice coding and high-quality low-bit-rate voice compression applications. In these applications, users can trade off between voice quality and coding rate (channel capacity), realizing the ability to have "soft" control.

A well-known example of variable bit rate is QCELP, which is a variable bit rate speech coder specified by CTIA called IS-95. So far, there are relatively many researches on CELP-based variable bit rate speech coding methods.

In voice activity detection, examples of well-known VAD methods include QCELP speech coder in IS-95 standard, EVRC in IS-127 standard, DTX mode in GSM standard and G.729 proposed by ITU-T VAD method of Annex B (G.729B).

In the past few years, there has been a lot of interest in Support Vector Machines (SVM). Experience shows that SVM generally has good performance in a large number of applications such as handwriting recognition, face recognition, and text classification. However, the application of this method in speech activation detection is rarely reported.

Low bit rate speech coding has been a major research topic over the past 20 years, resulting in the standardization of many speech coding algorithms with bit rates ranging from 16 kb/s to 2.4 kb/s. The current speech codec research focuses on high-quality speech coding at 4kb/s and lower, and recent research shows that speech coding in the frequency domain has the potential to have better quality than existing CELP-based coders. Spectral encoders are characterized by attempting to reconstruct the speech amplitude spectrum rather than recovering the speech waveform exactly. Although the above encoders based on CELP and parametric coding are widely used in low-bit-rate speech coding, most of them are limited by the assumed model accuracy, and they mainly rely on correct parameter estimation, which are often difficult to guarantee. Therefore, the robustness of these encoding methods is very poor in special environments, resulting in certain limitations in the quality of encoded speech.

The local cosine basis constructed successively by Coifman and Meyer (1991) and Auscher et al. (1992) is composed of the product of smooth, compactly supported clock function and cosine function. These localized cosine functions still retain orthogonality and have small Heisenberg products. In recent years, the theory of local cosine transform has been widely and deeply studied, but there are relatively few studies on this method in speech signal processing, especially in speech coding. However, in the article published by Malvar H.S. in 1990, it was proved that the coding gain of the LCT method in speech coding is better than that of DCT coding, and it is very close to KL transform coding. Especially compared to the DCT coding method, the "click" sound between frames is significantly reduced.

In view of the strong demand for low-bit variable-rate speech coding methods in current practical applications, as well as other previous coding methods based on models, the coding effect and the performance of the coder are often affected by the limitations of the assumed model accuracy and estimated parameter accuracy. application range.

Contents of Invention

The purpose of the present invention is to provide a practical, high-performance low-bit variable-rate speech coder by utilizing the characteristic of high coding gain of local cosine transform.

In order to achieve the above object, the technical solution adopted in the present invention is: a low-bit variable-rate speech coder, which is based on local cosine transform, and the speech coder preprocesses the input original speech signal through a high-pass filter, and then inputs To the speech activation detector to detect and distinguish the active speech frame and the non-activated speech frame, and then respectively process through the LCT converter to complete the speech coding, wherein:

Described voice activation detector adopts SVM-VAD voice activation detection module, and its workflow is as follows:

① Extract the parameters of the input speech data to obtain four classification characteristic parameters of the current frame: Line Spectral Frequencies, full-band energy, low-band energy, and zero-crossing rate;

②Initialization processing: Calculate and update the above four characteristic parameters at any time according to the change of background noise when there is only background noise;

③ Differential processing: Subtract the four characteristic parameters of the above-mentioned current frame from the corresponding four characteristic parameters when the initialization indicates that the current state only has background noise, and generate the corresponding four differential characteristic parameters required for voice activation detection and classification;

④The SVM algorithm is used for voice activation detection, and the training support vector machine adopts the Sequential Minimal Optimization (SMO) method, and finally the voice is divided into two voice types: active and inactive;

⑤ Using four-step smoothing and correction algorithm for VAD discrimination smoothing;

⑥After VAD processing is performed on each frame, the inactive or active speech frame signal is output. If the background noise energy of the frame is estimated to be greater than the background noise energy threshold, the average background noise parameter processing needs to be corrected again;

Described LCT converter handles, and its method is:

1. The non-activated speech frame detected by the SVM-VAD speech activation detection module is processed according to the fractal vector dimension of the silent/background noise speech mode, and then the fractal vector is processed according to the fractal dimension of the silent/background noise speech mode. The codebook of the corresponding fractal vector is subjected to fractal vector quantization to obtain the fractal vector quantization result corresponding to the two bit lengths of the speech mode, and the gain of the speech frame of the mode is scalar quantized. According to the order of 2 bits representing the speech mode, 8 bits representing the gain, and 7 bits representing the first fractal dimension vector and the second fractal dimension vector, 3 bytes are output to represent the voice of the frame End of coding;

②For the active speech frame detected by the SVM-VAD module, it is divided into three speech modes according to the methods of unvoiced (mode 0), unvoiced (mode 1), and moderately strong voiced (mode 2), and according to the corresponding three kinds of speech modes The dimension of the fractal vector is subjected to fractal processing, and then the corresponding four fractal vectors are quantized according to the codebook of the corresponding fractal vector of the corresponding speech mode, and four different fractal vectors corresponding to the speech mode are obtained. The length bits respectively represent the quantization results of the corresponding fractal vectors; at the same time, scalar quantization is performed on the gain of the speech frame, and the speech mode will be represented according to the two bits representing the speech mode, the 8 bits representing the gain, and the slave. These bits are formed into an integer number of bytes in the order of the bits of the first fractal dimension vector to the bits of the fourth fractal dimension vector, indicating that the speech coding of the frame is completed.

The first fractal dimension vector dimension and the second fractal dimension vector dimension of the silent/background noise speech pattern are 40; the first fractal dimension vector dimension of the unvoiced, unvoiced and moderately strong voiced speech patterns , the dimension of the second fractal dimension vector and the dimension of the third fractal dimension vector are both 40, and the dimension of the fourth fractal dimension vector is both 20.

The first and second fractal-dimensional vector bit allocations of the silent/background noise voice mode are 7 bits, the third and fourth fractal-dimensional vector bit allocations are 0 bits, the gain is 8 bits, and the mode is 2 bits; The first and second fractal dimension vector bit allocations of the unvoiced voice mode are 7 bits, the third and fourth fractal dimension vector bit allocations are 8 bits, the gain is 8 bits, and the mode is 2 bits; The bit allocation of the first and second fractal-dimensional vectors in the voice mode is 11 bits, the bit allocation of the third and fourth fractal-dimensional vectors is 8 bits, the gain is 8 bits, and the mode is 2 bits; the medium-strong voiced voice mode The first and second fractal-dimensional vector bit allocations are both 8 bits, the third and fourth fractal-dimensional vector bit allocations are both 8 bits and 6 bits, the gain is 8 bits, and the mode is 2 bits.

The present invention has made full use of the characteristics of the SVM method, applies the SVM to the VAD detection, and improves the correct recognition rate of the speech coder for speech detection; adopts the speech pattern classification method of GSM, and combines the original four speech patterns into Three speech modes, so that only two bits are used to represent the entire speech mode.

Description of drawings

Fig. 1 is the working flow chart of the SVM-VAD voice activation module that the embodiment of the present invention provides

Fig. 2 is a schematic diagram of the frame structure of the VBR-LCT speech coder provided by the embodiment of the present invention

Detailed ways

The present invention will be further described below in conjunction with accompanying drawing and embodiment:

Example:

1. Activate voice mode division

The criteria for voice mode selection in the GSM system are as follows:

Mode=0, P _v <1.7 (voiceless).

Mode=1, P _v ≥ 1.7, P _m < 3.5 for all m, (lightly voiced).

Mode=2, 3.5≦P _m <7.0, for all m, (medium voiced).

Mode=3, P _m >7.0, for all m, (strongly voiced).

Where m=1, 2, 3, 4 represent subframes in a certain frame, wherein P _m represents the open-loop LTP prediction gain (dB) of the mth subframe, and P _v represents the open-loop LPT prediction gain (dB) of the whole frame.

Strongly voiced and moderately voiced have stronger periodicity and higher speech energy. According to speech generation models, these two speech modes have strong formants, which is a good indication that they are conducive to producing clearer voiced sounds. For frequency domain coding, there is little difference in spectral components between strongly voiced and moderately voiced, so in the embodiment of the present invention, a method of combining the strongly voiced mode and the moderately voiced mode into one mode called moderately voiced is adopted. Another reason for merging into one moderately strong voiced mode is that due to the silent frame type detected by the VAD plus the above three speech modes, only 2 bits can be used to represent the conversion between coding modes. Therefore, in this embodiment, there are only three modes for activating speech, namely mode 0, mode 1 and mode 2, respectively representing unvoiced mode, slightly voiced mode and moderately strongly voiced mode.

2. Fractal dimension vector quantization method

Roughly speaking, the first four formants of adult speech signals are located at 500Hz, 1500Hz, 2500Hz and 3500Hz respectively. This actually divides the speech signal into four important regions, and the spectra of these four regions are required to be treated differently during encoding. Therefore, in the embodiment of the present invention, the coefficients of the local cosine transform are quantized by fractal dimension when designing an encoder. For each dimension vector, the vector quantization method (LGB algorithm) proposed by Linde, Buzo and Gray in 1980 is used for codebook training. After using the LGB algorithm to generate the codebook, a tree-shaped codebook search method is used in order to improve the codebook search speed during encoding and decoding.

In the fractal quantization adopted in the embodiment of the present invention, the number of local cosine transform coefficients for each mode of the active speech frame is divided into 40, 40, 40, and 20 from low frequency to high frequency. For silent or background noise frames, only the coefficients of the first two low-frequency bands are taken, which are 40 respectively. These four vectors are called the first dimension vector, the second dimension vector, the third dimension vector and the fourth dimension vector respectively. As for the speech signal with a sampling rate of 8kHz, only keeping the spectral components below 3500Hz is enough to recover the speech signal with satisfactory quality. In order to reduce the computational complexity, only 20 coefficients are used for the fourth-dimensional vector of the active speech mode frame, while the silent or background noise frame does not use the coefficients of the upper half frequency band. Table 1 shows the fractal vector dimensions of various modes of speech frames . When inversely transforming and synthesizing the speech signal in the decoder, the 20 coefficients of the remaining highest frequency components of the active speech frame and the 80 coefficients in the high half frequency of the silence (background noise) are filled with 0.

3. Bit distribution

Different bit rate allocation strategies are adopted according to the characteristics of various active speech frames and silent (background noise) frames. Table 2 is the bit allocation table of the VBR-LCT encoder provided by the embodiment of the present invention.

Speech in the medium-strongly voiced mode has strong periodicity, and the speech energy is mostly concentrated in the middle and low frequency bands, so more bits need to be allocated to the middle and low frequency bands. Such speech patterns are well represented by assigning a moderate number of bits.

For the pattern speech of light voiced sound, because it is a mixture of voiced sound and unvoiced sound to a certain extent, its periodicity is not as strong as that of the pattern speech of medium and strong voiced sound, but it contains the transition part of the speech. Although the proportion of the abrupt frame in the speech is small, it contains a lot of information, so whether it can be effectively expressed will directly affect the speech quality. For this reason, this embodiment adopts a strategy of allocating a higher number of bits to the speech frames of this mode.

Unvoiced mode speech can be said to be completely composed of unvoiced sounds, so the local cosine transform spectrum of unvoiced sounds should be considered flat. In the bit allocation, the same bits are basically allocated to each frequency band, but only one bit is added to each of the two frequency bands on the high half frequency in order to enhance the unvoiced characteristics of the high frequency part.

In order to obtain speech with better naturalness, in this embodiment, the speech of the silent or background noise frame is not filled with 0. Such processing will result in a sudden change in energy between the voiced frame and the silent frame, resulting in an uncomfortable phenomenon. For this reason, certain bits are also allocated to silent or noise frames to represent them. For strong background noise or in a special environment, if the sound is misjudged as silent, then this limited bit can also represent the information of the sound to some extent, which is unique to the local cosine transform coding method The advantages.

The gain of the speech frame encoder for each mode is calculated by using the ratio of the spectral energy of the input signal to the sum of the spectral energy of the code vector searched during encoding. The quantization of the gain adopts an 8-bit scalar quantization method. The total number of bits allocated to the voice frames of various modes is an integer number of bytes, so for the encoding of each mode of voice frames, the bit error inside the frame during transmission will not cause the decoding of subsequent voice frames, which has a certain Error resistance and error correction capability.

4. SVM-VAD method

The role of VAD is to distinguish between voiced and unvoiced, which is a well-known classification problem. As with any classification problem, the parameters for classification have to be chosen, and a discriminant function has to be designed. What we choose is a group of parameters that describe signal energy and spectral components that are usually used in VAD applications. The choice of parameters is governed by each parameter's contribution to the classification result, its robustness, and its computational complexity. The parameters selected here are four differential measurement parameters obtained by the difference between the current frame parameters and the background noise sliding average parameters, namely spectral distortion, full-band energy difference, low-band energy difference, and zero-crossing rate difference.

Both the VAD algorithm and the inactive vocoder operate on digitized speech frames. For compatibility, use equal frame sizes for all methods. Figure 1 is a general flow chart of the VAD algorithm for each frame. The result of using the SVM method for VAD discrimination is partial, that is, it does not consider the short-term stationary characteristics of speech and noise. Need to use the previous adjacent frame, using a four-step smoothing and correction algorithm. If the noise level changes suddenly, using the minimum energy estimate over a long period of time, a special reset algorithm is designed to prevent the algorithm from locking into audible mode.

FIG. 2 is a schematic diagram of the frame structure of a VBR-LCT speech encoder provided by an embodiment of the present invention. The pre-processing module in Fig. 2 is high-pass filter processing for reducing low-frequency noise and DC components. Speech coder input speech signal is the speech signal of sampling rate 8kHz 16 bit PCM format. In this embodiment, voice data in wav format is used, so the level amplitude is normalized.

Transformation analysis of signals usually adopts a short-time processing method. The selection of the length of the short-term signal segment has a great influence on the analysis results. The speech signal transform coding method also involves the problem of choosing the length of the analysis window. We know that speech signals are generally weak and non-stationary signals, but they can be approximately considered to be stable in a short period of time, such as an interval of 20ms. Therefore, in order to improve the compression ratio, choose a long window as much as possible in the encoding to reduce the bit rate, but at the same time it will increase the delay of the codec. Therefore, in the selection of frame length, according to the characteristics of the speech signal, it is required to compromise the delay and bit rate of the encoder. The low-bit variable-rate encoder provided by the embodiment of the present invention requires that the frame length should not be less than 20 ms. Furthermore, the frame length of 20 ms is adopted by most encoders, and belongs to the low-medium delay coding strategy. In the speech segment with a frame length of 20 ms, the speech signal can be approximately considered to be stable, which is conducive to the orthogonal representation of the speech signal, so in this embodiment, the frame length is selected as 20 ms, that is, 160 sampling points.

Encoder evaluation:

1. Objective evaluation

The data listed in Table 3 is the result of comparing the VBR-LCT speech coder provided in this embodiment with the coding standards of G.729B, GSM Half-Rate, FS1016 and FS1015. The results also illustrate the reliability of the objective evaluation method in speech encoder performance evaluation. G.729B, GSM Half-Rate, and FS1016 are all low-to-medium bit rate coding standards, and the voice quality of their coding far exceeds that of FS1015 and VBR-LCT methods. However, from these two indicators, VBR-LCT method has considerable advantages. Compared with the FS1015 coder with a similar bit rate, the SNR and PSNR of several types of speech data show that the VBR-LCT coding method provided by this embodiment is obviously higher than the SNR and PSNR of the FS1015 standard by nearly 5dB at most.

From the essential analysis of the speech coder, the VBR-LCT encoding method adopted in the present invention is carried out in the transformation domain, and its essence is the category of waveform encoding. Therefore, it is beneficial to use the two evaluation indicators of SNR and PSNR for objective evaluation, and the evaluation of the encoder by objective indicators can be used as a reference.

2. Subjective evaluation

The speech generated by the speech encoder is finally accepted by the human ear, so the quality of the speech after encoding is mainly evaluated by the auditory perception of the recipient. Here we use informal speech listening test to evaluate the speech quality.

When encoding the voice of the two-way dialogue, the average bit rate of the encoder of the VBR-LCT provided by this embodiment of the present invention is close to 1.6 kb/s. For noise-free clear speech, the reconstructed speech obtained by the VBR-LCT coder also has slight fuzziness, so the reconstructed speech like LPC-10e cannot be heard loud and loud. The speech intelligibility produced by the G.729B, GSM Half-Rate and FS1016 coding standards is not as high, but its intelligibility and naturalness are good, and significantly better than the similar bit rate LPC-10e method. The VBR-LCT coding method has strong robustness to environmental noise, and its coding distortion is not sensitive to the change of the signal, even the signals that are invalid to the G.729B, GSM Half-Rate, FS1016 and LPC-10e methods are still very stable. When using background music or other non-speech signals, the VBR-LCT coding method is significantly better than the LPC-10e method. These are entirely due to the fact that the VBR-LCT coding method belongs to waveform coding in the transform domain, so it does not depend on speech characteristic parameters such as pitch.

Table 1 voice mode Fractal vector first dimension vector second dimension vector 3rd dimension vector 4th dimension vector No sound/background noise Mode 0 (unvoiced) Mode 1 (lightly voiced) Mode 2 (moderately strong voiced) 40404040 40404040 0404040 0202020

Table 2 voice mode Fractal vector gain model bit/frame first dimension vector second dimension vector 3rd dimension vector 4th dimension vector No sound/background noise Mode 0 (voiceless) Mode 1 (light voiced) Mode 2 (moderately strong voiced 77118 77118 0888 0886 8888 2222 24404840

table 3 encoder type SNR(dB) PSNR(dB) bit rate(kb/s) G.729Anne×BGSM Half-RateFS1016FS1015(LPC-10e)VBR-LCT -0.951.240.71-3.59-0.96 15.0814.8116.7412.4715.08 85.64.82.41.6

Claims (3)

1. A low-bit variable rate speech encoder, after the input original speech signal is preprocessed by a high-pass filter, it is input to a speech activation detector to detect and distinguish active speech frames and inactive speech frames, and then pass through local cosine transformers respectively Process, complete speech coding, it is characterized in that: Described voice activation detector adopts support vector machine-voice activation detection module, and its workflow is as follows: ① Extract the parameters of the input voice data to obtain four classification feature parameters of the current frame, the line spectrum frequency, full-band energy, low-band energy, and zero-crossing rate; ②Initialization processing: Calculate and update the above four characteristic parameters at any time according to the change of background noise when there is only background noise; ③ Differential processing: Subtract the four characteristic parameters of the above-mentioned current frame from the corresponding four characteristic parameters when the initialization indicates that the current state only has background noise, and generate the corresponding four differential characteristic parameters required for voice activation detection and classification; ④Using support vector machine algorithm for voice activation detection, the training support vector machine adopts the sequential minimum optimization method, and finally divides the voice into two voice types: active and inactive; ⑤Adopt four-step smoothing and correction algorithm for speech activation detection, discrimination and smoothing; ⑥ After the voice activation detection processing is performed on each frame, the inactive or active voice frame signal is output. If the estimated background noise energy of the frame is greater than the background noise energy threshold, it is necessary to re-correct the average background noise parameter processing; Described local cosine transformer process, its method is: ①For inactive speech frames detected by the support vector machine-speech activation detection module, perform fractal processing according to the fractal vector dimension of the silent/background noise speech mode, and then the fractal vectors are respectively divided according to the silent/background noise speech mode The codebook of the corresponding fractal vector of the corresponding fractal vector is carried out fractal vector quantization, obtains the fractal vector quantization result that the length of two bits corresponding to this voice mode is 7 bits, and scalar quantizes the gain of this mode voice frame at the same time, According to the order of 2 bits representing the voice mode, 8 bits representing the gain, and 7 bits representing the first fractal dimension vector and the second fractal dimension vector, 3 bytes are output to represent the frame Speech encoding ends; ②For activated speech frames detected by the Support Vector Machine-Voice Activation Detection module, divide them into three speech modes according to unvoiced, unvoiced, and moderately strong voiced sounds, and perform fractal dimensioning according to the fractal vector dimensions of the corresponding three speech modes Then, the corresponding four fractal-dimensional vectors are quantized according to the codebook of the corresponding fractal-dimensional vector of the corresponding speech mode, and four different-length bits corresponding to the speech mode are obtained to represent the corresponding fractal-dimensional vectors respectively. Dimensional vector quantization result; Carry out scalar quantization to the gain of this speech frame at the same time, will represent two bits of speech pattern, 8 bits of expression gain and according to representing the bit of the first fractal dimension vector of this speech pattern The order of the bits from the bit to the fourth fractal dimension vector forms these bits into an integer number of bytes to output, indicating that the speech coding of the frame is completed. 2. low bit variable rate speech coder according to claim 1, is characterized in that: the first fractal dimension vector dimension of described silent/background noise speech pattern, the second fractal dimension vector dimension are 40, The third and fourth fractal dimension vector dimensions are all 0; the first fractal dimension vector dimension, the second fractal dimension vector dimension and the third fractal dimension vector dimension of the unvoiced, unvoiced and moderately strong voiced speech patterns The number is 40, and the dimension of the fourth fractal dimension vector is 20. 3. The low bit variable rate speech encoder according to claim 1, characterized in that: the first and second fractal dimension vector bit allocations of the described silent/background noise speech pattern are 7 bits, and the third and fourth The fractal-dimensional vector bit allocation is 0 bit, the gain is 8 bits, and the mode is 2 bits; the first and second fractal-dimensional vector bit allocations of the unvoiced voice mode are 7 bits, and the third and fourth fractal-dimensional vector bits are The allocation is 8 bits, the gain is 8 bits, and the mode is 2 bits; the first and second fractal-dimensional vector bit allocations of the unvoiced and voiced speech modes are both 11 bits, and the third and fourth fractal-dimensional vector bit allocations are both 8 bits, the gain is 8 bits, and the mode is 2 bits; the first and second fractal-dimensional vector bit allocations of the medium-strong voiced speech mode are both 8 bits, and the third and fourth fractal-dimensional vector bit allocations are both 8 and 6 bits, 8 bits for gain, and 2 bits for mode.

Images