CN1632861A - A Low Bit Rate Speech Coder - Google Patents

Abstract

This invention discloses a sound coder, which uses one flexible adjust shape function and uses the function to shape the clock function used by Donoho to get one new clock function to improve the high spectrum energy focus property. It uses sub dimensional method to the local cosine conversion parameter and uses LGB design method for each dimensional vector and uses tree structure search method from code searching.

Description

A Low Bit Rate Speech Coder

Technical field

The present invention relates to a speech coder, in particular to a low bit rate speech coder based on local cosine transform (LocalCosine Transform, LCT), suitable for use in communication systems requiring low bit rate speech coding.

Background technique

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. At present, the research focus of speech coder is on 4kb/s and lower high-quality speech coding. Although the CELP waveform encoder can still produce high-quality speech at bit rates below 6.3kb/s, when the bit rate is reduced to 4kb/s and lower, there are not enough bits to encode the details of the waveform. The system will generate a lot of quantization noise. On the other hand, parametric coding (also known as vocoder) does not try to produce a waveform similar to the original signal, but instead tries to find a set of parameters that can better represent important properties of speech perception, but they are not very useful for various special The robustness to ambient noise is poor.

However, for speech coding at bit rates of 4 kb/s and lower, recent studies have shown that speech coding in the frequency domain has the potential for better speech quality than existing CELP-based coders. Spectral encoders attempt to reconstruct the speech amplitude spectrum rather than recovering the speech waveform exactly. Although the above encoders are widely used in low-bit-rate speech coding, they are mostly limited by the assumed model accuracy, and they mainly rely on correct parameter estimation, which is often difficult to guarantee. Therefore, in special environments, the robustness of these coding methods is very poor, and the speech quality after coding has certain limitations.

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 theoretical method of local cosine transform has been widely and deeply studied. This method is widely used in image compression coding, but there are relatively few researches on speech signal processing, especially in speech coding. However, it is proved in the literature published by MalvarH.S. "Lapped transforms for efficient transform/subband coding". IEEETrans.on Acoust., Speech Signal Processing, 1990., vol.38(6), Page(s): 969-978 In speech coding, the coding gain of LCT method is better than DCT coding, and it is very close to KL transform coding. Especially compared with the DCT coding method, the "click" sound between frames is significantly reduced, and it is not necessary to use half-frame length sliding in order to reduce the abnormal "click" sound between frames like DCT transform coding. method. Therefore, the LCT method reduces the calculation amount by nearly half compared with the DCT method. In the literature "Comparison of picture compression methods: wavelet, waveletpacket and local cosine" published by Wickerhauser M.V. in 1994. Wavelets: Theory, Algorithms, and Applications, Editor (Charles K. Chui and Laura Montefusco and Luigia Puccio), Academic Press, San Diego, California, p.585～621, the comparative study of several two-dimensional image coding methods also shows that the LCT method is superior to the DCT method in terms of coding gain, and is also very close to the KL transform method. Studies have shown that the key to improving the coding gain of transform coding lies in the selection of orthogonal bases. Similarly, the key in local cosine transform coding is also the selection of local cosine quadrature bases, and the main factors affecting the selection of local cosine quadrature bases are The choice of clock function. The above few researches on applying the LCT method to speech coding are only in simple coding gain comparison, and have not really designed a feasible speech coder.

Contents of invention

The purpose of the present invention is to utilize the characteristic that local cosine transform has higher coding gain, and provide a kind of excellent low-bit-rate speech coder that is practical in local cosine transform domain.

The technical scheme that realizes the object of the present invention is: a kind of low bit rate speech coder, it is based on local cosine transform, is processed by high-pass filtering preprocessor to the original speech signal of input coder, then carries out local cosine transform (LCT) processing , it is characterized in that: the clock function bnew (n) in the described LCT transformation meets the following conditions:

ξ _[n] (n) is the shaping function adopted, which meets the condition ${\mathrm{\ξ}}_{[\mathrm{no}+1]}\stackrel{\mathrm{def}}{=}{\mathrm{\ξ}}_{\left[\mathrm{no}\right]}\left[\sin (\mathrm{\πt}/2)\right]$ and ${\mathrm{\ξ}}_{\left[0\right]}\left(t\right)\stackrel{\mathrm{def}}{=}\mathrm{\ξ}\left(t\right),$ in:

The subscript n is the number of iterations of the shaping function; the clock function takes values in the width of 1-4m.

Said clock function b _new (n) is guaranteed to be multiplied with the cosine function to form a local cosine quadrature basis.

The number n of iterations of the shaping function is 8-10.

The LCT coefficients of each frame after LCT transformation are first divided into fractal vector dimensions by 40, 40, 40, and 20 from low frequency to high frequency, and then four different fractal vector quantization codebooks are used to perform fractal dimension Vector quantization, the bits allocated from the first dimensional vector to the fourth dimensional vector are 12, 12, 8, and 8 bits respectively, and the gain quantization of each frame adopts 8-bit scalar quantization, according to the bits from the first dimensional vector The sequential output bits to the fourth fractal-dimensional vector bits and gain quantization bits are 48 bits, and 6 bytes are used to represent the output bit stream of each frame.

The speech coder also has a speech decoder matching it.

Due to the application of a shaping function that can be flexibly adjusted, the present invention uses this shaping function to shape the clock function adopted by Donoho to obtain a new clock function that can improve the aggregation of spectral energy; Dimensional quantization method, for each dimension vector, adopts LGB method to design codebook; codebook search in coding adopts tree structure search method, realizes a kind of excellent low-bit speech coder in local cosine transform domain. The objective parameter evaluation and informal listening test show that the coder has better naturalness and intelligibility than LPC-10e coder, and it is suitable for speech coding in various environments.

Description of drawings

Fig. 1 is the graph that the reshaping function in the speech coder of the embodiment of the present invention changes along with the number of recursions;

Fig. 2 is the bell function after the reshaping adopted in the speech coder of the embodiment of the present invention along with the increase of the number of recursions low half-frequency energy increase percentage figure (English+Chinese);

Fig. 3 is the structural representation of speech coder of the embodiment of the present invention;

Fig. 4 is the structural representation of speech decoder of the embodiment of the present invention;

Detailed ways

The technical solutions of the present invention will be further described below in conjunction with the drawings and embodiments.

Referring to accompanying drawings 3 and 4, the drawings respectively provide structural schematic diagrams of the low bit rate encoder and decoder described in this embodiment.

The key technology of the embodiment of the present invention is:

1. Obtaining the best shaped clock function

In Fig. 3, carry out high-pass filtering preprocessing to the original voice signal of input coder, then carry out LCT transform processing, in LCT transform, the clock function after the present invention adopts shaping is:

The clock function after the above shaping is obtained by the following steps:

1. Using Donoho's clock function:

When Wickerhauser MV described the local cosine transform algorithm in the monograph published in 1994, the given bell function is fixed for the given I _j and r.

The simple construction process of the clock function adopted by Donoho is given below. Suppose I _j =2m, r=m, then the width of the bell-shaped window is 4m, let

t(n)=n-0.5, 1≤n≤m.      (1)

x(n)=(1+t(n)/m)/2

Then, the bell window function used by Donoho is:

2. The structure of the shaping function is:

Let the input real-valued sequence t(n) be t(n)=[2(n-1)-m+0.5]/2m, 1≤n≤m (4)

define a real-valued continuous function

Replacing t with sin(πt/2) for the above formula, for any large fixed integer d, the d-time continuous differentiable function (ξ∈C ^d ) can be obtained. Define the following recursive function

${\mathrm{<span\; class="notranslate">\; \&xi;</span>}}_{<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; ]</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span>\stackrel{\mathrm{<span\; class="notranslate">\; def</span>}}{<span\; class="notranslate">\; =</span>}\mathrm{<span\; class="notranslate">\; \&xi;</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span>$

${\mathrm{<span\; class="notranslate">\; \&xi;</span>}}_{<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ]</span>}\stackrel{\mathrm{<span\; class="notranslate">\; def</span>}}{<span\; class="notranslate">\; =</span>}{\mathrm{<span\; class="notranslate">\; \&xi;</span>}}_{<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; ]</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; sin</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; \&pi;t</span>}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span>$

where the subscript of ξ represents the number of recursions. Through recursion, it will be seen that the derivative of ξ _[n] (t) at t=+1 and ^t =-1 is 0, which means that ξ _[n] ∈C ^2n-1 . Figure 1 shows several recursive result curves of this shaping function, where m=80.

3. Calculation of clock function after shaping:

Various shaping functions are generated by changing the number of recursions, and the bell function in (6.3) is shaped by using the shaping function ξ _[n] (t) after n times of recursion to obtain the following new clock function

The clock function in the above formula is guaranteed to be multiplied with the cosine function to form a local cosine quadrature basis.

In practical problems, it is necessary to find the best orthogonal basis on a fixed window width. That is to say, it is required to design a clock function that can be flexibly adjusted to meet the needs of practical problems. In this embodiment, the technical solution adopted is to perform de-correlation and de-redundancy on the speech signal, in order to better concentrate the spectral energy of the speech signal with a fixed frame length in several frequency bands, so as to facilitate sub-band coding. For this reason, the shaping method provided by the embodiment of the present invention is a shaping function capable of flexibly shaping the clock function adopted by Donoho, from which a shaping function suitable for frequency-domain speech coding is selected to obtain the best clock function.

4. Determination of the best bell-shaped function:

The embodiment of the present invention will involve the practical problem of transform domain speech coding, and what needs to be solved is to determine how many times of recursion the shaping function formed after shaping the bell function adopted by Donoho is the most suitable bell function. In this embodiment, the frame length is 20ms, the frequency band of the voice signal with a sampling rate of 8kHz is divided into high and low frequency bands, and the purpose of the shaping clock function is to require the spectral energy to be concentrated in the low half frequency band with a large amount of information as much as possible It is convenient for subsequent coding to optimize the allocation of bit numbers to the spectral coefficients of the high and low half-bands.

Referring to accompanying drawing 2, embodiment of the present invention adopts English and Chinese speech to carry out test and obtains along with the change of recursion number, utilizes the clock function after shaping than adopts Donoho's clock function to carry out the spectrum energy of lower half frequency band after local cosine transform. The percentage increase in total spectral power. It can be seen from FIG. 2 that the spectral energy increases the most when recursing 9 times. Therefore, in the embodiment of the present invention, a shaping function with 9 recursive times is selected for shaping. Although the proportion of spectral energy increase is small, it shows that adjusting the appropriate clock function can change the degree of spectral energy aggregation, which facilitates the optimization of bit allocation during encoding.

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, requiring us to treat the spectra of these four regions differently when encoding. For the encoding of the parameters in the transform domain, the split vector quantization (Splitted Vector Quantization) method is mostly used. Therefore, in the embodiment of the present invention, the designed encoder adopts the split dimension quantization method for the coefficients of the local cosine transform. Codebook training is performed separately for each dimension vector. 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.

During fractal quantization, the number of transformation coefficients of each dimension vector is divided into 40, 40, 40, and 20 from low frequency to high frequency. We refer to these four vectors as the first dimension vector, the second dimension vector, the third dimension vector and the fourth dimension vector. 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, the fourth dimension vector only uses 20 coefficients. When inversely transforming and synthesizing the speech signal in the decoder, the remaining 20 coefficients of the highest frequency components are filled with 0.

In the embodiment of the present invention, the bit allocation is 12, 12, 8, and 8 bits allocated to the vectors of each dimension from low frequency to high frequency. The speech encoder gain is calculated by taking the ratio of the spectral energy of the input signal to the sum of the spectral energies of the four code vectors searched for during encoding. The quantization of the gain adopts an 8-bit scalar quantization method. Table 1 shows the total bit allocation per frame of the encoder designed in the embodiment of the present invention.

Speech coder input speech signal is the speech signal of sampling rate 8kHz 16 bit PCM format. This embodiment uses voice data in wav format, so the level amplitude is normalized. The system has no special requirements on the type of speech, and is suitable for speech coding in various languages.

To the evaluation of encoder described in the embodiment of the present invention:

1. Objective evaluation

Other standardized encoders used for testing and comparing with the encoder described in the embodiment of the invention include G.729 Annex B (G.729B), GSM Half-Rate, FS1016, FS1015 (LPC-10e). The parameters used in objective evaluation are Signal to Noise Ratio (SNR) and Peak Signal to Noise Ratio (PSNR):

$\mathrm{<span\; class="notranslate">\; SNR</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 10</span>}\mathrm{<span\; class="notranslate">\; lo</span>}{\mathrm{<span\; class="notranslate">\; g</span>}}_{\mathrm{<span\; class="notranslate">\; 10</span>}}\frac{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; )</span>}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; e</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; )</span>}$

Here σ _x ² is the mean square of the speech signal, and σ _e ² is the mean square of the difference between the original speech signal and the reconstructed speech signal.

$\mathrm{<span\; class="notranslate">\; PSNR</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 10</span>}{\mathrm{<span\; class="notranslate">\; log</span>}}_{\mathrm{<span\; class="notranslate">\; 10</span>}}\frac{{\mathrm{<span\; class="notranslate">\; NX</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; -</span>\stackrel{<span\; class="notranslate">\; ~</span>}{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}$

Here N is the length of the reconstructed signal, X is the absolute maximum value in the signal x of length N,

is the sum of squares of the difference between the original signal and the reconstructed signal.

It is well known that objective evaluation of encoded speech signals sometimes yields puzzling results. Even if the speech encoded by one coder has a high SNR, sometimes it may not be of higher speech quality than the speech produced by another coder with a low SNR. On the contrary, the same holds true. Therefore, the objective parameter evaluation cannot be used as the main index of speech coder performance evaluation, it can only be used as an auxiliary evaluation.

Table 2 is the result of comparing the speech coder (FBR-LCT) of the present embodiment with G.729B, GSM Half-Rate, FS1016 and FS1015 coding standards. The results also illustrate the reliability of the objective evaluation method in speech encoder performance evaluation. G.729B, GSM Half-Rate and FS1016 all belong to low-medium bit-rate coding standards, and the voice quality of their coding far exceeds that of FS1015 and LCT coding methods, but from these two indicators, LCT method has a considerable high advantage. Compared with the FS1015 coder with the same bit rate, it shows that the SNR and PSNR of the LCT coding method are obviously higher than the SNR and PSNR of the FS1015 standard by nearly 5dB at most.

The encoding method adopted by the encoder in the embodiment of the present invention is performed in the transform domain, which is essentially the category of waveform encoding. Therefore, it is beneficial to use the two evaluation indicators of SNR and PSNR for objective evaluation. Therefore, objectively speaking, evaluating the encoder from several objective indicators cannot explain the problem, and can only 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. Informal speech listening tests are generally used to evaluate speech quality.

For noise-free clear speech, the speech reconstructed by the LCT coding method (FBR-LCT) adopted in the embodiment of the present invention has slight fuzziness, so the loud speech reconstructed like LPC-10e cannot be heard. The voice clarity produced by the G.729B, GSM Half-Rate and FS1016 coding standards is not as high, but its intelligibility and naturalness are good, and it is obviously better than the LPC-10e method with the same bit rate. The LCT coding method has strong robustness, 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 FBR-LCT coding method is significantly better than the LPC-10e method. These are entirely due to the fact that the LCT coding method belongs to waveform coding in the transform domain, so it does not depend on speech characteristic parameters such as pitch. On the contrary, G.729B, GSM Half-Rate, FS1016 and LPC-10e are based on speech source-filter generation model and estimation of linear prediction parameters, and are particularly sensitive to the accuracy of parameter estimation. The local cosine transform-based low bit rate encoder described in the present invention can also be realized by software simulation.

Table 1

Fractal vector gain frame

1st dimension vector 2nd dimension vector 3rd dimension vector 4th dimension vector (bits) (bits)

12 12 8 8 8 8 48

Table 2

English Chinese Chinese + background music bit rate

Encoder class

SNR(dB) PSNR(dB) SNR(dB) PSNR(dB) SNR(dB) PSNR(dB) (kb/s)

G.729 Annex -0.95 15.08 -1.46 18.32 -1.18 15.58 8

GSM Half-Rate -1.24 14.81 -0.82 19.46 -0.74 16.09 5.6

FS1016 0.71 16.74 1.37 21.63 1.27 18.09 4.8

FS1015(LPC10e) -3.59 12.47 -2.65 17.64 -1.80 15.02 2.4

FBR-LCT -0.44 15.08 0.26 20.54 -1.07 15.75 2.4

Claims (5)

1. a low bit rate speech coder, it is based on local cosine transform, the original speech signal of input coder is processed by high-pass filter preprocessor, then carries out local cosine transform processing, it is characterized in that: described local cosine The clock function b _new (n) in the transformation meets the following conditions: ξ _[n] (n) is the shaping function adopted, which meets the condition ${\mathrm{\&xi;}}_{[\mathrm{no}+1]}\stackrel{\mathrm{def}}{=}{\mathrm{\&xi;}}_{\left[\mathrm{no}\right]}\left[\sin (\mathrm{\&pi;t}/2)\right]$ and ${\mathrm{\&xi;}}_{\left[0\right]}\left(t\right)\stackrel{\mathrm{def}}{=}\mathrm{\&xi;}\left(t\right),$ in: The subscript n is the number of iterations of the shaping function; the clock function takes values in the width of 1-4m. 2. A low-bit-rate speech coder according to claim 1, characterized in that: said clock function b _new (n) is guaranteed to be multiplied with a cosine function to form a local cosine-orthogonal basis. 3. A low bit rate speech coder according to claim 1, characterized in that: the number of iterations n of said shaping function is 8-10. 4. a kind of low bit rate speech coder according to claim 1, is characterized in that: the local cosine transform coefficient of each frame after local cosine transform, first press respectively 40,40,40, 20 Divide the dimension of the fractal vector, and then use four different fractal vector quantization codebooks to perform fractal vector quantization. The bits allocated from the first dimension vector to the fourth dimension vector are respectively 12, 12, and 8 , 8 bits, the gain quantization of each frame adopts 8-bit scalar quantization, and the output bits are 48 bits according to the order from the first fractal-dimensional vector bit to the fourth fractal-dimensional vector bit and gain quantization bit, using 6 The bytes represent the output bitstream for each frame. 5. A low bit rate speech coder according to claim 1, characterized in that: said speech coder also has a speech decoder matching it.

Images