JP3510168B2 - Audio encoding method and audio decoding method - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a high-efficiency speech coding method and decoding method for digitally coding a speech signal sequence with a small amount of information in the processing of unvoiced sections. In particular, the present invention realizes high-quality voice encoding at a bit rate of 2.0 kbit / s or less, which is a region of a conventional voice analysis / synthesis system called a vocoder.

[0002]

2. Description of the Related Art As a prior art related to the present invention, a linear predictive vocoder and a code-excited linear predictive coding (CELP: Code Exc.
ited Linear Prediction). The linear prediction vocoder has been widely used as a speech coding method in a low bit rate region of 4.8 kbit / s or less, and includes a PARCOR method and a line spectrum pair (LSP) method. For details on these methods, see Saito,
It is described in "Basics of Speech Information Processing" by Nakata (published by Ohmsha). The linear predictive vocoder is composed of an all-pole filter that represents the spectral envelope characteristic of speech and an excitation signal that drives it. For the excitation signal, a pulse sequence is used for the voiced section and white noise is used for the unvoiced section. However, in a linear predictive vocoder, the excitation signal due to white noise is not sufficient to successfully reproduce the characteristics of the speech waveform, especially both plosive and fricative sounds, so it is difficult to obtain synthetic speech with high naturalness.

On the other hand, in code-excited linear predictive coding, speech is synthesized by driving two all-pole filters that represent the proximity correlation and pitch correlation characteristics of speech using a noise sequence as an excitation signal. The noise sequence is prepared in advance as a plurality of code patterns, and a code pattern that minimizes the error between the input speech waveform and the synthesized speech waveform is selected from among them. For details, refer to the document Schroeder: “Code-Excited Lin.
ear Prediction (CELP): High Quality Speech at Very
Low Bit Rates, ”Proc. IEEE ICASSP, pp937-940,198
It is described in 5. In code-excited linear predictive coding, the reproduction accuracy depends on the number of code patterns. Therefore, if a large number of code patterns are prepared, the reproduction accuracy of the voice waveform is increased and the quality can be improved accordingly. However, if the bit rate of voice coding is set to 4 kbit / s or less, the number of code patterns is limited, and as a result, sufficient voice quality cannot be obtained. It is said that an amount of information of about 4.8 kbit / s is required to obtain good voice quality.

[0004]

An object of the present invention is to perform linear predictive coding using a noise sequence as an excitation signal,
It is an object of the present invention to provide a method for realizing more efficient quantization of a voice waveform in an unvoiced section. Further, in the vocoder system, voiced and unvoiced discrimination of speech always includes an error, and when the voiced section is driven by white noise in which only power is matched, a remarkable deterioration in quality occurs.

[0005]

In order to solve the above-mentioned problems, the invention according to claim 1 determines a linear prediction analysis coefficient by performing a linear prediction analysis of a speech signal for each frame, and based on the linear prediction analysis coefficient. A speech coding method for determining a code obtained by quantizing a feature amount of a residual signal obtained by driving a linear prediction synthesis filter using a filter coefficient, wherein the feature amount is a periodicity of the residual signal. If the periodicity is lower than a predetermined threshold value, the residual signal is band-divided into a low-frequency component and a high-frequency component, and a noise code vector having a minimum distance to the low-frequency component is obtained. It is characterized in that a corresponding noise code is selected, and the high frequency component calculates an average power for each sub-frame forming the frame.

According to a second aspect of the present invention, in the speech coding method according to the first aspect, the calculated average power is normalized, the normalized average power is calculated, and the scale factor thereof is calculated. Characterize. According to a third aspect of the present invention, in the voice coding method according to the first or second aspect, the low-frequency component is obtained by thinning out sampling points of the low-frequency component waveform of the residual signal. It is characterized by selecting a random code corresponding to a random code vector that minimizes the distance between and.

The invention described in claim 4 is a speech decoding method for inputting an excitation source to a linear prediction synthesis filter to decode a speech signal, which is generated by the speech coding method according to claim 1. The noise code and the average power are input, the waveform of the low frequency component is decoded from the codebook based on the noise code, and the waveform of the high frequency component is the white noise that has passed through the high pass filter. Based on the converted average power, the gain is multiplied for each subframe to synthesize, and the waveforms of these two bands are added together to form an excitation source of the linear prediction synthesis filter.

A fifth aspect of the present invention is a voice decoding method for inputting an excitation source to a linear prediction synthesis filter to decode a voice signal, which is generated by the voice encoding method according to the second aspect. The normalized average power and the scale factor and the noise code are input, the waveform of the low frequency component is decoded from the codebook based on the noise code, and the waveform of the high frequency component is the normalized average power and the scale coefficient. It is characterized in that the waveforms of these two bands are added together to form the excitation source of the linear prediction synthesis filter.

The invention according to claim 6 is the invention according to claim 4 or 5.
In the speech decoding method described in (1), the sampling points of the low-frequency component waveform of the residual signal generated in the speech coding method are thinned to form a low-frequency component, and the distance to this low-frequency component is minimized. It is characterized in that the selected noise code corresponding to the noise code vector is input, and the sample points thinned out by sampling conversion based on the noise code are recalculated. According to the present invention, by using the above configuration, the noise sequence in the unvoiced section of the speech is band-divided, so that the low frequency band is used in combination with the waveform coding and the high frequency band is used in the quantization of the average power. It is characterized by efficient quantization to improve quality.

[0010]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Embodiment 1 FIG. 1 shows a functional configuration of an encoding unit to which the quantization method of the present invention is applied. The encoder performs the following procedure once for each frame having a length of N samples. In the frame i, the p-th order linear prediction coefficient (LPC) α _j (j = 0,1, ..., p-1) of the input speech signal s (t) from the input terminal TI1 is calculated by the linear prediction coefficient calculation unit 1 Calculate with. This linear prediction coefficient α is quantized by the linear prediction coefficient quantizer 2 and sent as a linear prediction coefficient code I ₁ . For details of the quantization of the linear prediction coefficient α, refer to “Speech Linear Prediction Parameter Coding Method” (Japanese Patent Application No. 3-180819).
No .: JP-A-5-27798). The linear prediction coefficient code I ₁ from the linear prediction coefficient quantization unit 2 is decoded, the filter coefficient of the linear prediction inverse filter 3 is determined based on the decoded linear prediction coefficient α ′, and the linear prediction inverse filter 3 Residual signal r through input voice signal s (t)
get (t). The linear prediction inverse filter 3 is realized by a digital filter A (z) having the following transfer characteristic.

A (z) = 1 + α ₁ z ⁻¹ + ... + α _p z ^-p (1) The correlation calculation unit 4 calculates the correlation (partial correlation function) ρ of the residual signal r obtained here. , The maximum value of the correlation ρ is ρ _max . At this time, the periodicity determination unit 5 determines whether the input voice signal s (t) is a voiced portion or an unvoiced portion by a threshold value θ (0.5 to 1.0) as described below, and the periodicity code I Output _{2 k 1/2 + ρ max>} θ; voiced portion _{k 1/2 + ρ max <} θ; unvoiced portion (2) where, k ₁ is the first-order partial autocorrelation which is obtained by the linear prediction coefficient calculator 1 (PARCOR) coefficients .

When the periodicity determining section 5 determines that the section is a voiced section,
Switch the switch SW1 to the voiced section quantizer 6 to change the residual signal.
No. r (t) is quantized by the voiced section quantizer 6 and encoded.
Output I ₃Is output. For details on voiced interval quantization,
Is described in "Voice coding method" (Japanese Patent Application No. 11-108161).
Has been done. When the periodicity determination unit 5 determines that the unvoiced section
Switches the switch SW1 to the unvoiced interval quantizer 7
The difference signal r (t) is quantized by the unvoiced section quantizer 7,
Encoded output I_FourIs output.

FIG. 2 shows details of the unvoiced section quantizer A.
On the basis of the linear prediction residual r from the linear prediction inverse filter 3, the average power calculation unit 9 divides the frame in the time direction by n _sfr to calculate the linear prediction residual for each subframe unit of length N _sfr = N / n _sfr . Calculate the average power p (element of the matrix representation P of the average power sequence). The following formula is used for this calculation.

[0014]

[Equation 1] At this time, 0 <i <n _fsr (n _fsr : the number of subframes in one frame). The average power P obtained here is
One frame of the average power is converted to the average power quantizer 10
The vector is quantized by and output as unvoiced part code I _4-1 .
To calculate the selection scale d when selecting the vector quantization code,
The following equation (4) is used.

D = ‖P−c _i ‖ ² (4) At this time, P is the vector of the average power sequence, and c _i is the code vector to be evaluated. In addition, the linear prediction residual signal r is filtered by the low-pass filter calculation unit 11 having a cutoff frequency f _c Hz to obtain a linear prediction residual of only the low band.
Calculate r _l . The cutoff frequency f _c Hz is 500 Hz to 1000 Hz.

Next, vector quantization of the low band prediction residual is performed. The residual signal r is the distance calculation unit for calculating the distance between the low-frequency prediction residual signal r _l generated via the low-pass filter calculation unit 11 and the noise code vector c (= c _i ) from the noise codebook 12. Calculation is performed using the following equation (5) in 13. d = ‖r _l −c _i ‖ ² (5) Here, r _l and c _i are the low-frequency prediction residual waveform and the matrix representation of the code vector to be evaluated, respectively.

Then, the distance calculation unit 13 selects a noise code corresponding to the vector c having the minimum distance, and the code of the selected code vector is output as I _4-2 . Next,
FIG. 5 shows a decoder A to which the embodiment of the decoding method corresponding to the embodiment of the encoding method shown in FIG. 2 is applied. Here, the codes I _{1 to} I ₄ input to the input terminal TI2 (that is,
After demultiplexing 29, all voice parameters are separated and decoded, and then a voiced section sound source synthesis unit 21
And unvoiced / voiced parameters in the unvoiced section sound source synthesis unit 22
An excitation signal is generated by I ₃ and I ₄ .

When the switch SW2 is switched by the periodic signal I ₂ to indicate the unvoiced section, the synthetic excitation signal from the unvoiced section sound source synthesis unit 22 is output, and when I ₂ indicates the voiced section, the voiced section sound source synthesis unit 21 outputs the synthesized excitation signal. The linear predictive synthesis filter 23 is driven by using the synthetic excitation signal e, and the output voice is obtained at the output terminal TO2. Here, the linear prediction coefficient code I ₁ is decoded by the linear prediction coefficient decoding unit 20 and output to the linear prediction synthesis filter 23.

As a method of decoding the voiced section, for example, the method described in "Voice coding method" (Japanese Patent Application No. 11-108161) is used. FIG. 6 shows details of the unvoiced section sound source synthesis unit A. In the unvoiced section, first, the average power decoding unit 24 decodes the average power sequence p _i based on the unvoiced part code I _4-1 . Next, the white noise generated by the white noise generation unit 25 is multiplied by the average power p in the multiplier 26, and the processing of the high pass filter calculation unit 27 of the cutoff frequency f _c Hz is performed, and the average power p Gain adjustment is performed so that each coefficient p _i is multiplied by f _c / 4000 to obtain a sequence p _i ^~, and the unvoiced high-frequency drive sound source signal e _h is generated. In this process, signals up to f _c Hz are
This is because the remaining f _c Hz to 4000 Hz signal power is approximately f _c / 4000 times p _i because it is generated more.

Further, the unvoiced section low frequency drive excitation e _l is decoded from the random codebook 28 based on the code I _{4-2 of the} code vector. Next, the unvoiced high-frequency driving sound source signal e _h and the unvoiced section low-frequency driving sound source e _l are added together by the adder 35, and are input to the linear prediction synthesis filter as the unvoiced section driving sound source e to obtain the output speech. Embodiment 2 FIG. 3 shows an unvoiced section quantizing unit B of an embodiment when the average power in FIG. 1 is normalized and quantized.

The average power sequence p _i obtained by the average power calculation unit 9 from the residual signal r is normalized based on the following equation. at ^{_{p i ~ = p i / s}} p (6) Here, the scale factor s _p is,

[0022]

[Equation 2] Calculate more. The scale factor s _p is the scale factor quantization unit
The scalar quantization in 16 and the normalized sequence p _i ^~ are vector-quantized in the normalized average power quantization unit 15, and the respective codes I _4-3
And I _4-4 are output. At this time, the following equation (8) is used to calculate the distance measure d when selecting a code for vector quantization of P 1 ^to P.

D = ‖P ^to −c _i ‖ ² (8) Here, P ^to is a matrix representation of the average power sequence, and c _i is a code vector to be evaluated. Decoder B corresponding to this embodiment
FIG. 7 shows decoding of the average power sequence p _i in. Each I
_4-3 and I _4-4 are decoded by the scale factor decoding unit 31 and the normalized mean power decoding unit 30, multiplied by the p _i ^~ and s _p at the multiplier 32,
Recalculate the average power sequence p _i .

Embodiment 3 The low-pass linear prediction residual r _l , which is the output of the low-pass filter operation unit 11 of FIG. 3, is reduced in time by shrinking the quantization operation amount, and the calculated average power is used. FIG. 4 shows the unvoiced section quantizer C when the quantization efficiency is improved. The low-frequency linear prediction residual r _l obtained in the same manner as in the above example is
In the time contraction unit 17, signals are decimated and contracted in the time direction, and the obtained signal r _l ^~ is converted into a noise code vector that minimizes the distance by using the normalized noise codebook 18 and the distance calculation unit 19. The corresponding random code is selected, the code of the code vector is generated, and vector quantization is performed.

If the speech signal s is sampled at a sampling frequency of 8000 Hz, this low-frequency linear prediction residual r
_{Since l} (f _c : 1000Hz to 500Hz) has only 1/8 to 1/4 of the total amount of information, even if 3 to 7 samples are thinned out from the sampling theorem and represented by 1 sample, the original low The region linear prediction residual r _l can be obtained, and the low region linear prediction residual can be quantized with a small amount of calculation. Next, in order to improve the quantization efficiency, the scale w _i is calculated for each subframe based on p _i obtained by the average power calculation unit 9, and the white noise w _i is calculated for each subframe. The white noise is weighted by multiplying w _i by a multiplier for each subframe.

The scale w _i is obtained from the following equation (9).

[0027]

[Equation 3] The following equation (10) is used to calculate the distance between the low-frequency prediction residual signal r _l and the codebook vector c (= c _i ). d = ‖r _l ^~ −wc _i ‖ ² (10) where r _l ^~ , c _i , and w are the time-reduced low-frequency prediction residual, the code vector to be evaluated, and the matrix representation of the scale w _i sequence, respectively. is there. Then, the code I _{4-5 of the} code vector selected from the normalized noise codebook 18 is output.

A decoder C corresponding to this embodiment is shown in FIG. The low band residual signal decoded by the code I _{4-5 of the} code vector is subjected to sampling conversion in the time expansion unit 34, and the decimated intermediate points are calculated in the encoder. Sampling conversion is performed based on the following equation (11).

[0029]

[Equation 4] Where x (nT _s ) is the original signal, x (t) is the desired midpoint,
T _s is the sampling period. In the signal obtained above,
Gain adjustment processing is performed so as to match the average power sequence p _i , and the unvoiced low-frequency drive sound source signal e _l is generated.

[0030]

As described above, according to the unvoiced section speech coding method of the present invention, compared to the method of quantizing only the power information in the unvoiced section simply by white noise used in the conventional vocoder. Quantization efficiency is improved. In addition, voiced and unvoiced discrimination, which has been a problem, separates the voice into low and high frequencies, so that even if the high frequency is driven by white noise, if the waveform of the low frequency component is encoded as it is, The periodicity is maintained to some extent. Therefore, even if an error occurs in voiced / unvoiced discrimination, there is an advantage that the quality is not significantly deteriorated.

In order to investigate the effect of the speech coding / decoding of the present invention, an analysis-synthesis speech experiment was conducted under the following conditions. As input voice, the voice of 0-4kHz band is sampled frequency
IR corresponding to the characteristics of the telephone after sampling at 8.0kHz
What passed the S characteristic filter was used. The encoder having the configuration of the second embodiment was used. First, add 25ms to this signal
For every (200 samples), a speech signal is multiplied by a Hamming window with an analysis window length of 30 ms, and a linear prediction analysis is performed by the autocorrelation method with the analysis order of 12th order to obtain 12 prediction counts. The prediction coefficient is vector-quantized using the Euclidean distance of the LSP parameter. 7 bits are allocated to I ₁ and I ₂ and I ₃ in the above-mentioned embodiment, and I _4-3 , I in the aspect of the third embodiment.
7 bits are allocated to _4-4 and I _4-6 , and the number of bits in the unvoiced section is 21 bits in any of the modes. Furthermore, 10 was used for n _sfr .

Under the above conditions, an improvement in auditory quality was recognized as compared with the case where the unvoiced section was simply encoded using 21-bit white noise information.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of an encoder.

FIG. 2 is a block diagram showing the configuration of an unvoiced section quantization unit A.

FIG. 3 is a block diagram showing the configuration of an unvoiced section quantization unit B.

FIG. 4 is a block diagram showing the configuration of an unvoiced section quantization unit C.

FIG. 5 is a block diagram showing the configuration of a decoder.

FIG. 6 is a block diagram showing a configuration of an unvoiced section sound source synthesis unit A.

FIG. 7 is a block diagram showing a configuration of an unvoiced section sound source synthesis unit B.

FIG. 8 is a block diagram showing a configuration of an unvoiced section sound source synthesis unit C.

[Explanation of symbols]

1 Linear prediction coefficient calculator 2 Linear prediction coefficient quantizer 3 Linear prediction inverse filter 4 Correlation calculator 5 Periodicity judgment section 6 Voiced section quantizer 7 Unvoiced section quantizer 8 multiplexer 9 Average power calculator 10 Average power quantizer 11 Low-pass filter calculation unit 12 noise codebook 13,19 Distance calculator 14 Average power sequence normalization unit 15 Normalized average power quantizer 16 Scale coefficient quantizer 17 hours contraction 18 Normalized random codebook 20 Linear prediction coefficient decoder 21 Voiced section sound source synthesizer 22 Voiceless section sound source synthesis section 23 Linear predictive synthesis filter 24 Average power decoding unit 25 White noise generator 27 High-pass filter calculator 28 Random codebook 29 Demultiplexer 31 Scale factor decoding unit 34 hours extension

─────────────────────────────────────────────────── --Continued from the front page (56) References JP-A-10-232697 (JP, A) JP-A-9-16198 (JP, A) JP-A-5-150800 (JP, A) Special Table 2001-525079 (JP, A) Yusuke Hiwasaki, et al., 2 kbit / s speech coding method based on pitch waveform, Proceedings of the 1997 Spring Research Presentation Meeting of the Acoustical Society of Japan, March 17, 1997, 3-7-15, p. ． 287-288 (58) Fields surveyed (Int.Cl. ⁷ , DB name) G10L 11/06 G10L 19/08 G10L 19/12 JISC file (JOIS)

Claims (6)

(57) [Claims] 1. A residual signal obtained by driving a linear prediction synthesis filter using a filter coefficient based on the linear prediction analysis coefficient by linearly analyzing a speech signal for each frame to obtain a linear prediction analysis coefficient. A voice encoding method for determining a code quantized a feature amount, wherein as the feature amount, the periodicity of the residual signal is determined, and when the periodicity is lower than a predetermined threshold, the residual error The signal is band-divided into a low-frequency component and a high-frequency component, a noise code corresponding to a noise code vector having a minimum distance to the low-frequency component is selected, and the high-frequency component is a subframe that constitutes the frame. A speech coding method, characterized in that an average power is calculated for each. 2. The speech coding method according to claim 1, wherein the calculated average power is normalized, the normalized average power is calculated, and a scale coefficient thereof is calculated. . 3. The speech encoding method according to claim 1, wherein a low-frequency component is obtained by thinning out sampling points of the low-frequency component waveform of the residual signal, and a distance from the low-frequency component is A speech coding method characterized by selecting a random code corresponding to a minimum random code vector. 4. A speech decoding method for driving a linear predictive synthesis filter with an excitation source to decode a speech signal, wherein a noise code and an average power generated by the speech coding method according to claim 1 are obtained. The waveform of the low-pass component is decoded from the codebook based on the noise code, and the waveform of the high-pass component is the white noise that has passed through the high-pass filter. A speech decoding method, characterized in that each subframe is multiplied by a gain for synthesis, and waveforms of these two bands are added together to form an excitation source of a linear prediction synthesis filter. 5. A speech decoding method for driving a linear predictive synthesis filter with an excitation source to decode a speech signal, comprising: a normalized average power and scale generated by the speech encoding method according to claim 2. The coefficient and the noise code are input, the waveform of the low frequency component is decoded from the codebook based on the noise code, and the waveform of the high frequency component is synthesized by multiplying the normalized average power and the scale coefficient. A speech decoding method, characterized in that the waveforms of the two bands are added to form an excitation source of a linear prediction synthesis filter. 6. The speech decoding method according to claim 4, wherein low-frequency components are obtained by thinning out the sampling points of the low-frequency component waveform of the residual signal generated in the speech encoding method. A speech decoding method characterized by inputting a selected noise code corresponding to a noise code vector having a minimum distance to a low-frequency component, and recalculating sample points thinned out by sampling conversion based on the noise code. .