JP3531780B2 - Voice encoding method and decoding method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To reproduce high quality voices using a small bit rate. SOLUTION: Inputted voices are fed into a linear prediction inverse filter 18 and residue signals r(t) are obtained. Then, the signals r(t) are segmented (31) into one pitch period for every frame (25ms), the segmented signals r(t) are normalized for a constant length (32), the peak position values of the signals r(t) are made constant (33) and the normalized waveforms of the signals r(t) are quantized. During a docoding, a waveform is made between two normalized waveform vectors before and after the subject waveform by a linear interpolation and is reproduced.

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 encoding method for digitally encoding a speech signal sequence with a small amount of information. The present invention relates to a coding method for realizing high-quality voice coding at a bit rate of 2.4 kbit / s or less, which is a domain of a voice analysis / synthesis system conventionally called a vocoder, and a decoding method thereof.

[0002]

2. Description of the Related Art The related art relating to the present invention includes:
Linear prediction vocoder, code excitation prediction coding (CELP:
Code Excited Linear Predi
ction), mixed-domain coding (Mixed-doma)
in Coding), representative waveform interpolation coding (Prot)
otype Waveform Interpolat
ion).

A linear prediction vocoder has a capacity of 4.8 kbit / s.
The following low-bit-rate speech coding methods have been widely used, and include the PARCOR method and the line spectrum pair (LSP) method. Details of these methods are described, for example, in Saito and Nakata, "Basics of Speech Information Processing" (Ohmsha Publishing). The linear predictive vocoder is composed of an all-pole filter that represents the spectral envelope characteristic of speech, and a sound source signal that drives the filter. As the driving sound source signal, a pitch cycle pulse train is used for voiced sounds, and white noise is used for unvoiced sounds. In a linear predictive vocoder, it is difficult to obtain a synthesized speech with high naturalness because a sound source driven by a periodic pulse train or white noise is not enough to reproduce the characteristics of a speech waveform.

[0004] On the other hand, in the code excitation predictive coding, speech is synthesized by driving two all-pole filters representing the close correlation and pitch correlation characteristics of the speech using a noise sequence as a driving sound source. The noise sequence is prepared in advance as a plurality of code patterns, and a code pattern that minimizes an error between the input speech waveform and the synthesized speech waveform is selected from among them. For details, see Schroeder “Co.
de-ExcitedLinear Predicti
on (CELP): High Quality Sp
tech at Very Low Bit Rate
s "Proc. IEEE. ICASPSP, pp937-
940, 1985. In the code excitation predictive coding, the reproduction accuracy depends on the number of code patterns. Therefore, if a large number of series patterns are prepared, the reproduction accuracy of the audio waveform is improved, and accordingly, the quality can be improved. However, if the bit rate of audio coding is set to 4 kbit / s or less, the number of code patterns is limited, and as a result, sufficient audio 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.

[0005] Also, mixed-domain coding (Mixed-do
In main coding, a pitch-period waveform is extracted from a residual waveform for each frame of a voiced sound, and a difference from a previous pitch-period waveform is quantized in a time domain. The decoder generates a sound source signal by performing linear interpolation of these waveforms in the frequency domain, and drives the all-pole filter to synthesize speech. For unvoiced sound, encoding is performed in the same manner as in code excitation predictive encoding. For details of this method, see Reference De.
Martin et al. “Mixed-domain Codi
ng of Speech at 3kb / s "Pro
c. IEEE. ICASP, PPII / 216-17
0, 1996. A feature of this method is that when calculating the difference, the previous pitch period waveform is normalized in length to the waveform of the current frame.
A pulse codebook and a noise codebook are used for quantization of the difference, but an information amount of about 3.5 kbit / s is required.

In addition, a representative waveform interpolation coding (Protot)
ype Waveform Interpolatio
n), the prototype waveform (Prototype W)
A sound is synthesized by driving an all-pole filter with a sound source signal synthesized by performing linear interpolation (aveform).
Details of this can be found in the document Kleijn W. B. “Encode
ing Speech Using Prototyping
e Waveforms "IEEE Trans.on
Speech Audio Processing,
Vol. 1, pp 386-399 1993. The prototype waveform is extracted from the residual waveform at a constant cycle, and is encoded after being subjected to Fourier transform. According to this method, an amount of information of about 3.4 kbit / s is required to obtain good quality.

[0007]

SUMMARY OF THE INVENTION An object of the present invention is to provide a linear predictive encoding method using a noise sequence or a pitch pulse train as a drive signal when the frequency band of an input signal is restricted, such as telephone speech. , A method for realizing more efficient encoding and a method for decoding the same.

[0008]

A coding method according to the present invention estimates a pitch period of an input voice, extracts a waveform corresponding to the estimated pitch period from a periodic portion of a driving sound source signal, and calculates a length of the waveform. It is characterized in that the drive signal is determined such that the waveform distortion with the normalized value is minimized. Here, a feature different from the conventional method is that the input pitch period waveform is normalized so that the length matches the fixed-length code vector and the impulse response of the synthesis filter is similarly convoluted to determine the code. It is. Further, when synthesizing the voice, an interpolated one of the preceding and following driving sound sources is returned to the pitch cycle length and connected.

[0009]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Embodiment 1 FIG. 1 shows a functional configuration of an encoding unit to which an encoding method according to the present invention is applied. This encoding unit performs the following procedure once for each frame having a length of N samples. In frame i, p of input audio signal s (t) from input terminal 11
The next linear prediction coefficient (LPC) a _j (j = 0, 1,..., P
-1) is calculated by the LPC calculation unit 12. The linear prediction coefficients are quantized by LPC quantization section 13, it is transmitted as linear prediction coefficient code I _1. For details of quantization of linear prediction coefficients, see "Speech Linear Prediction Parameter Coding Method".
(Japanese Patent Application No. 3-180819). LPC
Based on the linear prediction coefficient obtained by the calculation unit 12, the filter coefficient of the linear prediction inverse filter 14 is determined, and the residual signal r is input to the inverse filter 14 through the input audio signal s (t).
Calculate (t). The inverse filter 14 is realized by a digital filter having the following transfer characteristics.

A (z) ⁻¹ = 1 + a ₁ z ⁻¹ +... + _Ap z ^-p (1) The correlation (deformed correlation function) ρ of the residual signal obtained here is calculated by the correlation calculator 15, The delay (interval) of the maximum value ρ _max of the correlation ρ is calculated by the pitch period extracting unit 16 as the estimated pitch period p.
_i . At this time, the periodicity determination unit 17 determines whether the input audio signal s (t) is a voiced part or a non-voiced part, for example, using a threshold value θ (0.5 to 1.0) as follows.

[0011] _{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 (PARCOR) coefficients obtained by the LPC calculation section 12 is there. LPC quantization unit 1
Linear prediction coefficient code I ₁ than 3 is decoded, on the basis of the linear prediction coefficients the inverse quantization, determines the filter coefficients of the linear prediction inverse filter 18, through an input audio signal s (t) to the inverse filter 18 A residual signal r '(t) is obtained.
When the determination section 17 determines that the voice section is unvoiced, the voiceless section quantization section 19 performs quantization as shown in FIG. 2A. In the quantization unit 19, the frame is divided into S, the number of N _sub (= N / S) samples is set as a subframe, and the average power of the residual waveform r ′ (t) obtained by the inverse filter 18 in the subframe. Is calculated by the power calculator 21, and one frame thereof is vector-quantized by the vector quantizer 22 and output as the unvoiced code I ₂ . The quantization of the unvoiced part may be performed by a configuration as shown in FIG. 2B. In other words, the linear prediction synthesis filters 23 and 2 are calculated based on the quantized linear prediction coefficients from the LPC quantization unit 13.
4 is set, and the residual signal r ′ (t) from the inverse filter 18 is applied to the input sound signal s by the synthesis filter 23.
(T) is reproduced, and on the other hand, the noise code selected from the noise codebook 25 is given the gain selected from the gain codebook 27 by the gain section 26, and the noise code given the gain is combined with the synthesis filter 2.
4, a difference between the synthesized unvoiced voice and the synthesized voice from the synthesis filter 23 is calculated by a subtraction unit 28, and a noise codebook is calculated by a distortion calculation unit 29 so that the square of the difference (error) is minimized. 25, and the gain codebook 27 is selected. The noise code and gain of the gain codebook 27 of the noise codebook 25 at this time is unvoiced portion numerals I _2.

When the periodicity determining unit 17 determines that the voice is a voiced part, the pitch period p _i estimated by the residual extracting unit 31 is used.
Using, cut the length of the waveform p _i from the residual signal r '(t) frame near the center in from the inverse filter 18, to normalize the length to the vector length n by stretching it. This normalization is performed by the sampling conversion unit 32 based on the sampling function of Expression (3). x (t _i ) = _{{n = −q} ^q x (nT) · sin (π / T) (t _i −nT) / {(n / T) (t _i −nT)} (3) where T Is a sampling period, and q is a value that is theoretically infinite but is truncated by a finite number.

Next, the normalized residual waveform is rotated by the alignment unit 33 until the correlation between the normalized residual waveform and the pulse signal becomes large. Here, the length of the estimated pitch period is normalized to have a length of n, and the phase of the residual waveform _rnp whose rotation is also normalized is converted to an NPW (normalized pitch period waveform).
Call. Estimated pitch period p _i is the pitch period quantizer 34
In quantized to integer values by rounding, is output as a pitch period codes I _3.

The NPW from the alignment unit 33 is vector-quantized by an NPW quantization unit 35. The NPW quantization unit 35 is, for example, as shown in FIG. 3A, the LPC quantization unit 13 in FIG.
A linear prediction synthesis filter 37 the filter coefficients are determined by linear prediction coefficients quantized more is passed through the impulse signal, the impulse response h _j is obtained, the impulse response h _j are the n sampling conversion section 38 The impulse response matrix H normalized to the length and based on the normalized impulse response h ′ _j is represented by the alignment unit 3 in FIG.
The NPW from No. 3 is convolved by the convolution filter 39 to synthesize the audio signal x. On the other hand, with respect to the code vector c ₀ ⁱ selected from NPW codebook 41, gain g _o ^k taken out from the gain codebook 43 by a gain unit 42 is provided,
On the other hand, the impulse response matrix H is convolved by the convolution filter 44 and synthesized into speech. An error of the synthesized speech with respect to the reproduced speech is obtained by the subtraction unit 45, and the square of the error is minimized. Selection of code vector c ₀ ⁱ of NPW codebook 41 and gain g _o ^k of gain codebook 43
Is selected by the distortion calculator 46.

The length of each code vector in the NPW codebook 41 is n, and the peak phases are uniform.
The cycle of the pulse signal used in the NPW alignment unit 33 in FIG. 1 is n, and the phase is matched with the phase of the peak of the code vector of the codebook 41. As described with reference to FIG.
The W code is determined such that the auditory weighted mean square error between the waveform synthesized with the code vector as the driving sound source and the waveform synthesized with the NPW waveform as the driving sound source is minimized.
The following equation (4) is used for the distance calculation of the distortion scale D of the distance.

[0016] ^{_{D = ‖x-g o k Hc}} 0 i ‖ ² (4) wherein, x represents a target (synthesized waveform NPW waveform as the driving source), H is the linear prediction coefficients a _'j quantized A matrix representing the normalized impulse response of the used synthesis filter 37, c _o is a code vector,
g _o represents the gain of the code vector.

The target x is calculated in advance by a convolution operation in the filter 39 using the following equation (5). x = _Hrnp (5) Here, _rnp is a vector representation of the original NPW waveform before the quantum. In the conventional CELP coding, a lower triangular (n × n) square matrix is usually used for H.
In order to obtain the free response obtained by passing the W waveform through the synthesis filter, the (m × n) rows are extended to the lower side by (mn) rows.
Is used. Here, m > n. For H, the impulse response h _j (j = 0, 1,..., P _i −1) of the linear prediction filter subjected to auditory weighting is sampled and converted by the sampling converter 38 and normalized in the same manner as r _np to h ′ _J (j = 0, 1,..., N−1) is used.

[0018]

(Equation 1) At this time, the linear prediction synthesis filter 37 used for calculating h _j (j = 0, 1,..., P _i -1) is realized by a digital filter having the following transfer characteristics. A (z) = 1 / (1 + a ₁ z ^-1 +... + _Ap z ^-p ) (7) The transfer characteristic of the auditory weight is expressed as follows.

W (z) = A (γ ₁ z) / A (γ ₂ z) (8) Here, γ ₁ and γ ₂ are parameters for controlling the degree of auditory weighting, and 0 < γ ₂ < γ Take a value of ₁ < 1. FIG.
The matrix H used for the convolution operation in the convolution filters 39 and 44 in A is formed by using the above-described extended m × n matrix from the normalized impulse response h _j . Since H is extended in this manner, the target x obtained by the operation of Expression (5) is also

[0020]

(Equation 2) And the order is m. _{Here, x i (n <i <} m-1) is a component corresponding to the free response of a linear prediction filter 37, a quiescent initial value response of the synthesis filter 37. In the selection of the NPW code c ₀ , the code vector c ₀ ⁱ is selected from the code vector book 41 so that the equation (4) is minimized, and the ideal gain g ₀ ⁱ is calculated. First, equation (1)
The code vector c ₀ ⁱ having the maximum value of D ′ _{0 of} (0) is selected in a closed loop.

D ′ ₀ = (x ^T Hc ₀ ⁱ ) ² / ‖Hc ₀ ⁱ ‖ ² (10) The calculation of the ideal gain g ₀ ⁱ of the selected code vector c ₀ ⁱ uses the equation (11). Do it. g ₀ ⁱ = x ^T Hc ₀ ⁱ / ‖Hc ₀ ⁱ ‖ ² (11) Next, the gain g ₀ ⁱ is scalar-quantized. Since the selection of the code vector has been completed by the above procedure, g ₀ ^k that minimizes the expression (4) is selected. The code of the selected code vector and the code of the selected gain are NPW
The frame is output as a code I ₄ , and the periodic determination unit 17 outputs a periodic code I ₅ indicating whether the frame is a voiced part or an unvoiced part. Symbols I _{1 to} I ₄ are combined by a multiplexer 47 and output to a transmission path or a storage unit.

As described above, one frame is set to, for example, 25 milliseconds, from which a residual waveform (signal) for one pitch period is extracted, that is, for example, a fraction of one frame.
Only the part has been removed. On the other hand, even when the synthesis filter 37 is driven with the input set to zero, an output corresponding to the state immediately before that, that is, a so-called zero input response occurs. Therefore, in CELP encoding, a target obtained by subtracting a zero input response from an input waveform is targeted. However, according to the present invention, since encoding is performed using only a part of one frame, the impulse response matrix of the synthesis filter 37 is extended by an amount corresponding to the zero response as compared with the CELP encoding, and the waveform for one pitch period is reduced to zero. A code vector close to this is selected, including the input response (free response). As described above, since the waveform information is encoded only for one pitch period in one frame, the number of bits is required to be smaller, the pitch period is normalized at that time, and the peak position is normalized (constant phase). Therefore, the number of coding bits can be reduced in this respect as well.

Next, an embodiment of the encoding method shown in FIG.
Decoding to which the embodiment of the decoding method of the present invention is applied
The vessel is shown in FIG. Here, the mark input to the input terminal 51
Issue I ₁~ I_FiveIs a demultiplexer 52 for all audio parameters.
After the meter is separated and decoded, the voiced and unvoiced parameters
I_Two, I_FourTo generate a driving sound source. Periodic code I
_FiveIs a voiceless part, the voiceless part code I_TwoTo the unvoiced part decoding unit 5
In step 3, the driving sound source signal is reproduced. That is, for example, as shown in FIG.
As described above, the white noise from the white noise
Sign I_TwoIs processed by the gain calculator 55.
Generate a synthetic residual waveform of the unvoiced part. In other words, N samples
A white noise is generated and each subframe (N _suLong) in
Average power is the average of the corresponding decoded subframe
Calculate and multiply the gain to match the power and drive
A motion signal.

FIG. 4 shows a case where the periodic code I ₅ indicates a voiced part.
In the NPW code I ₄ , the NPW decoding unit 56 adds the gain g ⁱ to the code vector c ² as shown in Expression (12).
Multiplied by, decoding the NPW waveform r ^i. Although not shown, the NPW codebook 41 and the gain codebook 4 in FIG.
3 is provided. r ⁱ = g ₀ ⁱ c ₀ ⁱ (12) Next, the decoded NPW waveform r ⁱ and the previous NPW buffer 5
Performs linear interpolation between the contents r ^i-1 of 7 by linear interpolation section 58, to obtain an intermediate of the NPW waveform r ^in. For this linear interpolation, for example, equation (13) is used.

S ⁱⁿ (j) = (1−α) S ⁱ⁻¹ (j) + αS ⁱ (j) (j = 0, 1,..., N−1; 0 < α < 1) (13) , Α are values indicating the position of the waveform in the frame having the length of N samples, S ^i-1 is a vector immediately before S ⁱ , and S ⁱⁿ is a vector obtained by interpolation. That is, on the encoding side, the residual waveform is cut out only for one pitch period in each frame. Therefore, between the waveform cut out in the current frame and the waveform cut out in the previous frame, there is originally a waveform of about one pitch period to several pitch periods. The originally linearly interpolating at should be present waveform and decoded NPW waveform r ^i-1 of the previous frame and decoding NPW waveform r ⁱ of the current frame. Α is determined according to the order of the waveform to be interpolated between the extracted waveform of the previous frame and the extracted waveform of the current frame. The pitch period codes I ₃ is decoded by a pitch decoder 59, the number of waveform is determined to be interpolated from its decoded pitch period and the frame length.

The decoding pitch period and the previous pitch buffer 6
According to the content of 1, the pitch interpolating unit 62 interpolates each pitch cycle between the pitch cycle of the cut waveform of the previous frame and the pitch cycle of the cut waveform of the current frame,
Using the pitch period and the decoding pitch period, the corresponding intermediate NPW waveform from the linear interpolation unit 58 is sampled and converted by the sampling conversion unit 63, that is, returned to the pitch period of the original voice, and successively connected by the residual signal synthesis unit 64. This is a driving sound source signal.

Although not described in the figure, the NPW corresponding to a half pitch or a double pitch period is erroneously performed during encoding.
Is extracted and when interpolation is performed by the above method, the other NPW
Assuming that the pitch is correct, the use of the interpolation waveform from the linear interpolation unit 58 degrades the quality of the output sound. Therefore,
If the decoded pitch periods before and after are greatly different, for example, approximately 2: 1, the shorter one of the preceding and following waveforms is repeated twice, and this is re-normalized to a vector of n sample length by sampling conversion. Linear interpolation is performed using the quantization vector and the longer waveform. Similarly, the pitch cycle of the shorter pitch cycle is doubled and the pitch interpolation of the other pitch cycle is performed.

The periodic signal I_FiveIs a silent part when indicates a silent part
The synthesized sound source signal from the signal unit 53 is represented by I _FiveIndicates a voiced part
Is linear using the synthesized sound source signal from the residual signal synthesis unit 64
Drives the prediction synthesis filter 65 and outputs the output sound to the output terminal 6
Get 6 Here, the linear prediction coefficient code I₁The linear predictor
The number is decoded by the numerical decoding unit 67, and the linear prediction coefficient is also
Using the contents of the coefficient buffer 68, one
Linear prediction coefficient for one cycle and one pitch cycle in the current frame
The linear prediction coefficient interpolating unit 69 calculates a value between the linear prediction coefficient and the
The coefficient of the synthesis filter 65 is obtained by performing linear interpolation according to (13).
To determine. The interpolation of the linear prediction coefficient is the same as the conventional CELP
The method used in the method may be used.Example 2 In the case where multi-stage quantization is performed by the NPW quantization unit 35 in FIG.
FIG. 6 shows the NPW quantization unit of the embodiment. In FIG. 6, FIG.
The same reference numerals are given to the portions corresponding to A,
In the case of round quantization, an NPW codebook 71 is provided.
Code vector c selected from NPW codebook 71₁ ^jTo
On the other hand, the gain selected by the gain unit 72 from the gain codebook 43
g₁ ^kIs given to the convolution filter 73,
The normalized impulse response H is convolved, which gives
The combined waveform obtained is subtracted from the error signal by the subtractor 45 to the subtracter 7.
4 and the rest is given to the distortion calculator 75.
The distortion calculation unit 75 minimizes the square of the output of the subtraction unit 74.
The code vector c of the NPW codebook 71₁ ^jof
Selection and gain g of codebook 43₁ ^kThe choice is made and
You. Also in this case, as a whole, the code vector is
And NPW waveform as driving sound source
The weighted mean square error with the waveform
Sea sign vector c₀ ⁱ, C₁ ^j, Gain g₀ ^k, G
₁ ^kIs determined. Of the distortion measure of this Euclidean distance
Equation (14) is used for distance calculation.

[0029] ^{_{D = ‖x-g o k Hc}} 0 i -g 1 k Hc 1 j ‖ ² (14) wherein, x represents a (5) targets determined by the equation, H is the linear prediction coefficients a quantized Synthesis filter 3 using _i '
A matrix representing the normalized impulse response of 7
c _o and c ₁ represent code vectors, and g _o and g ₁ represent gains of the respective code vectors.

First, as described with reference to FIG.
Stage c_oAnd its ideal gain g_o ⁱIs determined. next,
From the code vector book 71, the expression (14) is minimized.
Such a code vector c₁ ^jAnd its ideal gain g
₁ ^jAnd c₀ ⁱG, the ideal gain of₀ ⁱRecalculate
Calculate. This gives the code vector c_o ⁱAnd c₁ ^jNo
Performs vector orthogonalization and performs encoding. This vector orthogonalization
For more information on vector quantization based on
Orthogonalized speech coding method "(Japanese Patent Application No. 6-43519).
Have been.

For the selection, the code vector c ₁ ⁱ having the maximum value of D ₁ ′ in equation (15) is selected in a closed loop.

[0032]

[Equation 3] Calculation of the ideal gain g ₁ ^j of the selected code vector is performed using Expression (16).

[0033]

(Equation 4) Further, the ideal gain g ₀ ⁱ is recalculated using the equation (17).

[0034]

(Equation 5) Since the selection of the code vector has been completed by the above procedure, (g ₀ ^k g ₁ ^k ) that minimizes the expression (14) is selected, and vector quantization is performed. The decoder in this case is the same as that in FIG. 4, but uses equation (18) for decoding the NPW waveform.

[0035] In ^{_{r i = g 0 i c o}} i + g 1 i c 1 i (18) above, the adaptive codebook as shown in Figure 5B the codebook (a), a fixed codebook (b), algebraic Any of the pulse codebooks (c) can be used. The adaptive codebook (a) is a residual waveform in the past, and the algebraic pulse codebook (c) can be generated each time according to rules. Embodiment 3 FIG. 3B shows an embodiment in which quantization is performed using a codebook (two) having a conjugate structure as the NPW quantization unit 35 in FIG. 1, and parts corresponding to those in FIG. is there. An NPW codebook 81 is further provided. Each code vector of the NPW codebook 81 and the code vector of the NPW codebook 41 have a conjugate structure with each other, that is, are orthogonal to each other. The gain selected from the book 43 is provided, and the code vector to which the gain has been provided and the code vector from the gain unit 42 are added by the addition unit 83 and provided to the convolution filter 44 as a drive excitation signal. N is set so that the auditory weighted mean square error between the waveform obtained by combining the code vector as the driving sound source signal and the waveform obtained by combining the NPW waveform as the driving sound source signal is minimized.
Each code vector of the PW codebooks 41 and 81 and its gain are determined. Equation (14) is used in the distance calculation of the distance distortion scale as in the second embodiment. For details of the encoding method using the conjugate structure codebooks 41 and 81, see "Multiple Vector Quantization Method and Apparatus Thereof" (Japanese Patent Application No. 63-2).
49450).

Also in this case, an adaptive codebook, a fixed codebook, and an algebraic pulse codebook as shown in FIG. 5B can be used as the codebook. In the above description, when a plurality of codebooks are used, from the plurality of types shown in FIG. 5B,
For example, a combination of an adaptive codebook and a fixed codebook may be used. The NPW decoding unit 56 in FIG. 4 for multi-stage vector quantization and conjugate structure vector quantization prepares codebooks corresponding to the number of input code vectors, and codes based on the input NPW code I ₄ from these codebooks. Fetch the vectors, and input them to the input N
PW code I ₄ in gain encoded by each corresponding gain obtained from the gain codebook may be given respectively. The thus decoded NPW vectors may be linearly interpolated, and may be further subjected to sampling conversion to obtain continuous signals and may be added to each other to be supplied to the synthesis filter 65 as a residual signal.

The unvoiced part quantization unit 19 shown in FIG. 1 extracts a noise vector from the noise codebook, and selects a noise vector so as to minimize the square of the error between the noise vector obtained by giving the gain and the input residual signal. And the gain to be given to it. In FIG. 1, the linear prediction inverse filter 14 may be omitted, and the output of the linear prediction inverse filter 18 may be input to the correlation calculator 15. Pitch period detection accuracy is inverse filter 1
4 is better. In FIG. 1, the linear prediction coefficient code I ₁ may be output for one pitch period.

[0038]

As described above, according to the encoding method of the present invention, only one pitch period in one frame is encoded in a voiced section. Can be reduced. In addition, during encoding, the vector length is normalized and the peak positions are aligned, so that there is no phase information of the waveform, and the number of encoded bits can be further reduced.

According to the decoding method of the present invention, only information for one pitch period in one frame is input in a voiced section, but a code vector is generated by interpolating between two preceding and succeeding normalized code vectors. Similarly, a pitch period is generated by interpolating between the preceding and following two decoding pitch periods, and then each code vector is expanded and contracted to the corresponding pitch period and connected to form a connected drive signal, thereby producing a synthesized signal. The sound can be reproduced by driving the filter.

In order to examine the effects of the speech encoding / decoding method of the present invention, an analysis / synthesis speech experiment was performed under the following conditions. As the input sound, a sound in a band of 0 to 4 kHz, which was sampled at a sampling frequency of 8.0 kHz and then passed through an IRS characteristic filter, was used. The encoder and decoder having the configuration of the second embodiment (FIGS. 1, 6 and 4) were used. First, 25 ms (200 ms)
For each sample), a voice signal is multiplied by a Hamming window having an analysis window length of 30 ms, and a linear prediction analysis is performed by an autocorrelation method with an analysis order of 12, thereby obtaining 12 prediction coefficients. The prediction coefficient is vector-quantized using the Euclidean distance of the LSP parameter.

If the state of the input speech signal is determined to be a voiced part, the residual waveform having the length of the estimated pitch period before quantization is represented by
Assuming that q = 20 and n = 120, 12
Sampling conversion is performed on an NPW vector having a length of 0 samples, and this NPW waveform is vector-quantized using two noise code vectors c ₀ ⁱ and c ₁ ^j . The pitch determined by the partial autocorrelation method is scalar-quantized to an integer value.

If the input voice signal is determined to be unvoiced, the average power of the residual waveform in each 5 ms subframe is calculated by dividing the 25 ms frame into five, and the five values are vector quantized. The bit rate is 2.08 kbit / s when there is periodicity, and 1.2 when there is no periodicity.
4 kbit / s, and the breakdown is as follows.

Parameter Number of bits / frame Prediction coefficient (LSP) 21 Voiced / unvoiced parameter 1 Driving sound source (if voiced) First-stage noise sequence 7 Second-stage noise sequence 7 Noise sequence gain 8 Pitch period 7 Driving sound source (In the case of unvoiced) noise sequence 8 Speech coded under the above conditions has much higher naturalness than a conventional vocoder of the same bit rate, and also has a higher quality than a conventional CELP coding of the same bit rate. The voice quality was also clear and less noise.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a functional configuration example of an encoder to which an embodiment of an encoding method according to the present invention is applied.

2A is a block diagram showing a specific functional configuration of the unvoiced part quantization unit 19 in FIG. 1; FIG.
FIG. 9 is a block diagram showing another specific functional configuration 9.

3A is a block diagram illustrating a specific functional configuration example of an NPW quantization unit 35 in FIG. 1; FIG. 3B is a block illustrating a specific functional configuration example of the NPW quantization unit 35 in the case of conjugate structure vector quantization; FIG.

FIG. 4 is a block diagram showing a functional configuration example of a decoder to which the embodiment of the decoding method according to the present invention is applied;

5A is a block diagram illustrating a specific functional configuration of the unvoiced section decoding unit 53 in FIG. 4, and FIG. 5B is a diagram illustrating examples of various codebooks used in the present invention.

FIG. 6 is a block diagram showing an example of a functional configuration when an NPW quantization unit 35 in FIG. 1 is a multi-stage vector quantization method.

──────────────────────────────────────────────────続 き Continued on front page (58) Field surveyed (Int.Cl. ⁷ , DB name) G10L 19/12

Claims (12)

(57) [Claims] A drive for driving a linear prediction synthesis filter of a linear prediction coefficient obtained by the analysis and a filter coefficient based on the linear prediction coefficient by analyzing the voice signal for each frame longer than the pitch period of the voice. In a speech coding method for expressing speech characteristics by signals, voiced / unvoiced section discrimination is performed for each frame, and if the frame is a voiced section, the residual signal obtained by performing linear predictive inverse filtering on the voice signal is used. Extracting a residual signal vector having a pitch period length, normalizing the extracted residual signal vector length to a predetermined length, and comparing the normalized residual signal vector with a predetermined reference signal vector. In order to increase the correlation, a target residual vector is obtained by circulating through the residual signal vector element, and the target residual vector is sent to the synthesis filter. A target waveform vector is obtained by synthesizing the target signal with a voice signal, and a selected one of a plurality of predetermined code vectors is used as a drive signal to synthesize a voice by the synthesis filter to obtain a synthesized waveform vector, and the target waveform of the synthesized waveform vector is obtained. A speech code characterized by selecting the code vector that minimizes the distortion of the waveform with respect to the vector, determining a quantization code, and determining the code by quantizing the residual signal if the determination is an unvoiced section. Method. 2. The method according to claim 1, wherein the target waveform vector is generated by lowering a non-square matrix obtained by expanding a lower triangular matrix based on an impulse response of the synthesis filter to obtain a free response of the filter. 2. The speech encoding method according to claim 1, wherein a convolution operation is performed on a difference vector, and the non-square matrix is convolved with the selected code vector to generate the composite waveform vector. 3. An impulse response is obtained by passing an impulse through a synthesis filter having a filter coefficient based on the linear prediction coefficient, the impulse response is normalized to a vector length having the predetermined length, and the length is changed. 3. The speech encoding method according to claim 2, wherein the non-square matrix is created by an impulse response. 4. The driving signal for obtaining the composite waveform vector by a weighted linear sum of code vectors selected from a plurality of codebooks, respectively. Voice encoding method. 5. The driving signal for obtaining the composite waveform vector by a weighted linear sum of code vectors respectively selected from a plurality of codebooks having a conjugate structure. The speech encoding method according to any one of the above. 6. The quantization of the unvoiced section is performed by exciting the synthesis filter with a noise vector selected from a noise codebook and selecting a noise vector that minimizes distortion between an output signal and an input speech signal. 2. The method according to claim 1, wherein
6. The speech encoding method according to any one of claims 1 to 5. 7. The quantization of the unvoiced section is performed by dividing one frame into a plurality of sub-frames and vector-quantizing the average power of each sub-frame for each frame. 6. The speech encoding method according to any one of claims 1 to 5. 8. A linear prediction unit encoded for each frame.
Numeric code, periodic code and,Input the quantization code of the drive signal.
Filter obtained by decoding the linear prediction coefficient code
A linear prediction synthesis filter having coefficients is
Driving with the decoded signal of the quantization code to synthesize the output sound
In the audio decoding method If the periodic code indicates a voiced section, the drive
Two code vectors before and after decoding the quantized code of the motion signal
ofWeighted linearInterpolation and input pitch circumference
Of the two pitch periods before and after decoding the period codeWeighted linear
Interpolate, According to the interpolated pitch period, the interpolated code
By expanding and contracting the vector length of the
Generate motion signalAnd The two pitch periods before and after the decoding differ greatly from each other.
If the above, interpolation of the above code vector is a short pitch period length
After repeating the corresponding code vector twice,
Is normalized to the length of the code vector of
Interpolation between the vector and the other code vector, and
The interpolation of the pitch period is also doubled for the shorter one,
The correction between the switch cycle andSoundVoice decoding
Method. 9. before and after the two above decoded speech decoding method of claim 8 Symbol mounting the linear prediction coefficients the linear interpolation and obtains the filter coefficient of the synthesis filter. 10. If the periodic code indicates an unvoiced section, decoding the input unvoiced part code to obtain an unvoiced residual waveform, and driving the synthesis filter with the unvoiced residual waveform. The speech decoding method according to claim 8 or 9 , wherein: 11. The decoding of the unvoiced code is performed by matching the power of each subframe of the generated white noise with the power of a plurality of subframes obtained by dividing one frame from a power codebook. 11. The speech decoding method according to claim 10 , wherein a residual waveform is obtained. 12. The speech decoding method according to claim 10 , wherein the decoding of the unvoiced part code extracts a noise vector from a noise codebook to obtain the speech residual waveform.