JPH10232697A - Voice coding/decoding method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To reproduce a high quality voice with a less bit rate. SOLUTION: An input voice is made to pass a linear predictive reverse filter 14 through, and a residual signal r(t) is obtained, and this is cut out by one pitch period at every one frame (25ms) 31, and its cut off r(t) is made so that its peak position becomes a fixed level, and this waveform is vector quantized using one cut off by one pitch periodic length from a code book consisting of code vectors that peaks are arranged. At a decoding time, the waveforms between two waveform vectors in front/behind, and between two pitch periods in front/behind are formed respectively by a linear interpolation method, and interpolated waveform vectors are connected successively by interpolated pitch periods, and a continuous exciting signal is obtained.

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, and more particularly to a speech analysis / synthesis system called a vocoder, which is 2.4 kbit / s. The present invention relates to an encoding method for realizing high-quality audio encoding at the following bit rates and a decoding method thereof.

[0002]

2. Description of the Related Art The prior art related to the present invention includes:
Linear prediction vocoder, code excitation prediction coding (CELP:
Code Excited Linear Predi
ction), mixed domain coding (Mixed Dom)
ain Coding), a representative waveform interpolation coding method (Pr
ototype Waveform Interpol
ation).

[0003] A linear prediction vocoder is 4.8 kbit / s.
The following speech coding methods in the low bit rate region have been widely used, and include a PARCOR method and a line spectrum pair (LSP) method. Details of these methods are described, for example, in Saito and Nakata, “Basics of Speech Information Processing” (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 the filter. As the excitation signal, a pitch cycle pulse train is used for voiced sounds, and white noise is used for unvoiced sounds. In a linear prediction vocoder, an excitation signal generated by a periodic pulse train or white noise is not enough to reproduce the characteristics of a speech waveform.
It is difficult to obtain synthetic speech with high naturalness.

[0004] On the other hand, in code excitation predictive coding, speech is synthesized by driving two all-pole filters representing the close 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 an error between the input speech waveform and the synthesized speech waveform is selected from among them. See Schroeder: “Code for details.
Excited Linear Prediction
n (CELP) High QualitySpee
chat Very Low Bit Rates ”
Proc. IEEE. ICASSP, pp937-94
0,1985. In the code excitation predictive coding, the reproduction accuracy has a relationship depending 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 information amount of about 4.8 kbit / s is required to obtain good voice quality.

[0005] Also, a mixed area coding method (Mixed D
In the main coding, a pitch-period waveform is extracted from a residual waveform for each frame in a voiced sound, and a difference from a previous pitch-period waveform is quantized in a time domain.
The decoder generates an excitation signal by performing linear interpolation of these waveforms in the frequency domain, and drives an all-pole filter to synthesize speech. For unvoiced sound, encoding is performed in the same manner as in code excitation prediction encoding. For details of this method, see Reference De.
Martin et al. “Mixed DomainCodi
ng of Speech at 3kb / s "Pro
c. IEEE. ICASP, PPII / 216-17
0, 1996. A feature of this method is that the length of the previous pitch period waveform is normalized to the waveform of the current frame when the difference is obtained.
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 method (Proto
type Waveform Interpolati
on Coder), the prototype waveform (Prot
Speech is synthesized by driving an all-pole filter with an excitation signal synthesized by performing linear interpolation (type Waveform). This is described in detail in the document Kleijn W.
B. “Encoding Speech Using
Prototype Waveforms "IEEE
Trans. on Speech AudioProc
essing, Vol. 1, pp 386-399 19
93. The prototype waveform is extracted from the residual waveform at a constant period, and is encoded after being subjected to Fourier transform. In this system, 3.4 k is required to obtain good quality.
It is said that an information amount of about bit / s is required.

[0007]

SUMMARY OF THE INVENTION An object of the present invention is to provide a linear predictive coding method using a noise sequence or a pitch pulse train as an excitation signal when the frequency band of an input signal is limited, such as telephone speech. , A method for achieving 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 speech, extracts a waveform corresponding to the estimated pitch period from a periodic portion of an excitation signal, and extracts the waveform corresponding to the pitch period. The code vector is determined by truncating the code vector at the pitch length so that the waveform distortion with the waveform (1) is minimized. Here, a vector obtained by convoluting the impulse response of the synthesis filter with the input pitch period waveform,
The code is selected by calculating the distance between the code vector and the vector that is similarly convolved with the code vector truncated at the pitch length. Further, when synthesizing the voice, an interpolated one of the preceding and subsequent excitation signals is returned and connected. During this interpolation,
The excitation pitch period waveform vectors of different pitch lengths are characterized in that the short vector is adjusted to the long vector and zero-padded to the difference vector.

[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".
(JP-A-5-27798). The linear prediction coefficient code I ₁ from the LPC quantization unit 13 is decoded, and based on the inversely quantized linear prediction coefficient, a filter coefficient of the linear prediction inverse filter 14 is determined.
4 to calculate a residual signal r (t) through the input audio signal s (t). The inverse filter 14 has the following transfer characteristic 1 / A (z)
It is realized by a digital filter having.

A (z) ⁻¹ = 1 + a ₁ z ⁻¹ +... + _Ap z ^-p (1) The correlation (partial 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 determined by the pitch period extracting unit 16 as the estimated pitch period p _i.
And At this time, the input sound signal s
Whether (t) is a voiced or unvoiced part is determined by, for example, a threshold value θ (0.5 to 1.0) as described below.

[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 which is obtained by the linear prediction coefficient calculation unit 12 (PARCOR) It is a coefficient. Judgment unit 17
Is determined to be a voiceless part, the voiceless part quantization unit 19 performs quantization as shown in FIG. 2A. The quantization unit 19 divides the frame into S, _{sets the} number of N _sub (= N / S) samples as a subframe, and calculates the average power of the residual waveform r (t) from the inverse filter 14 in the subframe. The calculation is performed by the unit 21, and the one frame is vector-quantized by the vector quantization unit 22 and output as the unvoiced code I ₂ .

The quantization of the unvoiced portion may be performed by a configuration as shown in FIG. 2B. That is, the filter coefficients of the linear prediction synthesis filters 23 and 24 are set based on the quantized linear prediction coefficients a _j 'from the LPC quantization unit 13, and the residual signal r (t) from the inverse filter 18 is input to the input filter 23 by the synthesis filter 23. The signal s (t) is reproduced. On the other hand, the gain selected from the gain codebook 27 is given to the noise code selected from the noise codebook 25 by the gain section 26, and the noise code given the gain is subjected to speech synthesis by the synthesis filter 24. Then, the synthesized speech and the synthesis filter 2
The difference from the reproduced voice from No. 3 is obtained by the subtraction unit 28, and the distortion calculation unit 29 selects the noise code of the noise codebook 25 and the gain codebook 27 so that the square of the difference (error) is minimized.
Is performed. 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 section 17 determines that the voice section is a voiced section, the pitch cycle p _i estimated by the residual extracting section 31 is used.
, A waveform having a length of p _i is cut out from the vicinity of the center of the frame in the residual signal r (t) from the inverse filter 14. Until the correlation between the extracted residual waveform and the pulse signal becomes large, the residual waveform is circulated by the PW alignment unit 32. Here, the residual waveform r _p of one cycle waveform partial cut out is referred to as a PW (pitch cycle waveform). Estimated pitch period p _i is quantized to an integer value by rounding at the pitch period quantization unit 34, is output as a pitch period codes I _3.

The PW from the PW alignment unit 32 is vector-quantized by a PW quantization unit 35. As shown in FIG. 3A, for example, the PW quantization unit 35 passes the impulse signal to a linear prediction synthesis filter 37 whose filter coefficient is determined by the quantized linear prediction coefficient from the LPC quantization unit 13 in FIG. Thus, an impulse response h _j is obtained. An impulse response matrix H based on the impulse response h _j is
The convolution filter 39 convolves the PW from the PW alignment unit 32 in FIG. 1 to synthesize the audio signal x. On the other hand, P
The code vector from the head selected from W codebook 41, cut at the code cut-out portion 38 by the pitch period length of the pitch period extraction unit 16 is extruded in Fig. 1, with respect to the cut-out code vector c ₀ ⁱ , the gain in the gain unit 42 was taken out from the gain code book 43 g _o ^k is given, whereas, in the convolution filter 44 impulse response matrix H
Is convoluted and synthesized, an error of the synthesized audio signal with respect to the reproduced audio signal x is obtained by the subtraction unit 45, and the code vector c ₀ ⁱ of the PW codebook 41 is reduced so that the square of the error is minimized. a selection, the selection of the gain g _o ^k of the gain codebook 43 is carried out by the strain calculation unit 46. That is, the code vector and the gain are selected using the reproduced audio signal x as the target waveform vector.

Note that the length n of each code vector in the PW codebook 41 must be sufficiently longer than the extracted pitch length p _i (n: n < p _max ). Here, the phase of the peak of the code vector is assumed to be uniform. PW alignment unit 3 in FIG.
The pulse signal used in 2 has a vector length of n,
The phase is matched with the phase of the peak of the code vector of the PW codebook 41.

As described with reference to FIG. 3A, the PW code is determined so that the auditory weighted mean square error between the waveform obtained by combining the code vector as the excitation signal and the waveform obtained by combining the PW waveform as the excitation signal is minimized. . The following equation (3) is used for calculating the distortion scale D. D = ‖ x-g _o ^k Hc ₀ ⁱ ‖ ² (3) where, x (waveform obtained by synthesizing the PW waveform as excitation signal) is the target waveform vector, H is the linear prediction coefficients a quantized
matrix representing the impulse response of the synthesis filter 37 with _j ', c _o is the code vector, g _o represents a gain code vectors.

The target x is calculated in advance by a convolution operation in the filter 39 using the following equation (4). x = Hr _p (4) where, r _p is obtained by the vector representation of the original PW waveform before quantum. In conventional CELP coding, the H frame length used square matrix when under normal triangle N (N × N), but here the positive direction matrix pitch length p _i (p
_i × p _i ) is extended downward by (q−p _i ) rows (q × p _i ).
Use the non-square matrix of p _i ). Here is a q> p _i. H is the impulse response h of the linear prediction filter weighted by the auditory weight, h, which is truncated at the pitch length p _i
j (j = 0, 1,..., p _i -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 characteristic H ′ (z). H '(z) = 1 / A' (z) = 1 / (1 + a 1 'z -1 + ... + a p' z -p) (6) transfer characteristic of the perceptual weighting is expressed as follows.

W (z) = A (z / γ ₁ ) / A (z / γ ₂ ) (7) Here, γ ₁ and γ ₂ are parameters for controlling the degree of auditory weighting, and 0 < γ ₂ < Γ ₁ < 1 Take a value in the range.
As the matrix H used for the convolution operation in the convolution filters 39 and 44 in FIG. 3A, the expanded (q × p _i ) matrix described earlier than the impulse response h is used. Since H is extended in this manner, the target waveform vector x obtained by the calculation of Expression (4) is also

[0020]

(Equation 2) And the order is q. Here, x _j (p _i <j <q−1)
Is a component corresponding to the free response of the linear prediction filter 37, and is a zero input initial value response of the synthesis filter 37. Here, in order to reduce the calculation amount, the length of h may be truncated to a very short length (for example, about 10 samples) to form H.

The coefficient of the linear prediction inverse filter 14 and the PW quantization unit 35 are calculated using the quantized linear prediction coefficient a _j ′.
Of the synthesis filter 37 is determined. Even if H (z) is configured using the linear prediction coefficients a _j that are not quantized for both, the same quality as when H ′ (z) is used can be obtained. . Even in this case, the H matrix of the convolution filter 44 must use the quantized linear prediction coefficients a _j '.

In the selection of the PW code c ₀ , the PW code book 41
As Equation (3) is minimized among, select the code vector c ₀ ^i, calculates the ideal gains g ₀ ^i. First, D ₀ ′ in equation (9) is calculated, and a code vector c ₀ ⁱ having the maximum D ₀ ′ value is selected in a closed loop. Note that x
^T represents the transposed matrix of x. ^{_{D 0 '= (x T Hc}} 0 i) 2 / ‖Hc 0 i ‖ ² (9) calculates the ideal gains g ₀ ⁱ of the selected code vector c ₀ ⁱ is performed using equation (10).

G ₀ ⁱ = x ^T Hc ₀ ⁱ / ‖Hc ₀ ⁱ ‖ ² (10) Next, the gain g ₀ ⁱ is scalar-quantized. Since the selection of the code vector has been completed by the above procedure, the expression (3)
But to select the g ₀ ^k such that minimum. The code of the selected code vector and the code of the selected gain are represented by the PW code I.
Output as _4, further, the frame from the cycle determining unit 17 outputs a periodic code I ₅ indicating whether voiced portion or unvoiced portion. Symbols I _{1 to} I ₅ are synthesized by the multiplexer 47,
The signal is output to the transmission path or the storage unit.

As described above, one frame is set to, for example, 25 milliseconds, and a residual waveform (signal) for one pitch period is extracted therefrom, that is, for example, a fraction of one frame.
Only the part has been removed. On the other hand, even if the synthesis filter 37 is driven with the input set to zero, an output corresponding to the immediately preceding state, that is, a so-called zero input response is generated. Therefore, in CELP encoding, a target waveform vector (target) is obtained by subtracting a zero input response from an input waveform. However, in 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 a zero response rather than CELP encoding, and 1
A code vector close to the pitch period waveform including the zero input response (free response) is selected. As described above, since the waveform information is encoded only for one pitch period in one frame, it can be expressed with a smaller number of bits, and the peak position is normalized (constant phase). The number of coded bits can be reduced.

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_Five Is a demultiplexer 52 for all audio parameters.
After the meter is separated and decoded, the unvoiced / voiced parameter I
_Two , I_Four An excitation signal is generated according to. Periodic code I_Five
Is a voiceless part, the voiceless part code I_Two To the unvoiced decoding unit 53
To reproduce the excitation signal. That is, for example, as shown in FIG.
The unvoiced part code is added to the white noise from the white noise generation unit 54.
I_Two Is processed by the gain calculation unit 55 and
Generating a composite residual waveform of the section. That is, N samples of white
Noise, and generate each subframe (N_suAverage)
Power is the average power of the corresponding decoded subframe.
Calculate and multiply the gain to match the
No.

FIG. 4 shows the case where the periodic code I ₅ indicates a voiced part.
In the PW code I ₄ , the PW decoding unit 56 calculates the expression (1
As shown in 1), the PW waveform r ⁱ is decoded by multiplying the code vector c ⁱ by the gain g ⁱ . Although not shown, provided with the same as the PW codebook 41 and gain codebook 43 in FIG. 3A, code vector c ⁱ is PW codebook 41
Are cut out by the code cutout unit 60 by the pitch period length p _i from the beginning of the code vector.

R ⁱ = g ₀ ⁱ c ₀ ⁱ (11) Next, linear interpolation between the decoded PW waveform r ⁱ and the contents r ⁱ⁻¹ of the previous PW buffer 57 is performed by the linear interpolation section 58, get the PW waveform r ^in. This linear interpolation includes
For example, equation (12) is used. r ⁱⁿ (j) = (1−α) r ⁱ⁻¹ (j) + α _r ⁱ (j = 0, 1,..., p−1; 0 < α < 1) (12) where α is a waveform Represents a position in a frame of N sample lengths, p is the number of longer samples of the preceding and succeeding pitches (p _i or p _i-1 ), and r ^i-1 is the PW immediately before r ⁱ a vector, r ⁱⁿ represents a vector that can be interpolated. The remainder of the short pitch vector is padded with zeros, and interpolation is performed after matching the long vector with the long one.

That is, on the encoding side, the residual waveform is cut out only for one pitch period in each frame. Therefore,
Originally, a waveform of about one pitch cycle to several pitch cycles exists between the waveform cut out in the current frame and the waveform cut out in the previous frame. The originally linear interpolation should be present waveform decoding PW waveform r ^i-1 of the previous frame and decoding PW 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, each pitch cycle between the pitch cycle of the cut-out waveform of the previous frame and the pitch cycle of the cut-out waveform of the current frame is obtained by linear interpolation in the pitch interpolation unit 62, and the interpolation pitch cycle is used. The intermediate PW waveform obtained by the linear interpolation unit 58 is sequentially connected by the residual signal synthesis unit 64, and this is used as an excitation signal.

Although not shown in the figure, a half pitch or a double pitch cycle of PW is erroneously generated at the time of encoding.
When interpolation is performed by the above method on the decoding side when
Assuming that the pitch of the other PW is correct, the use of the interpolated waveform from the linear interpolation section 58 degrades the quality of the output sound. Therefore, if the decoded pitch cycles before and after greatly differ, for example, approximately 2: 1, the shorter of the preceding and following waveforms is repeated twice, and linear interpolation is performed using this vector and the longer waveform. Similarly, the pitch cycle of the shorter pitch cycle is doubled, and pitch interpolation with the other pitch cycle is performed.

Here, at the time of interpolation, the length of the PW is sampled.
Vector before and after the same length
(N samples) and the following equation (1) as in equation (12).
It is also possible to perform linear interpolation of two vectors based on 3)
It is possible. Sⁱⁿ(J) = (1−α) S^i-1(J) + αSⁱ(J) (j = 0, 1,..., N; 0<α<1) (13) Here, α is any one of the frames in the frame whose waveform is N samples long.
N is the normalized vector length, S
^i-1Is rⁱThe PW vector immediately before ⁱIs
rⁱAre normalized PW vectors of
SⁱⁿRepresents a normalized vector obtained by interpolation.
This SⁱⁿIs similar to the above
After the pitch period length is interpolated, they are sequentially connected.

The periodic signal I_Five Is a silent part when indicates a silent part
The synthesized sound source signal from the signal unit 53 is represented by I _Five Indicates a voiced part
Is linear using the combined excitation signal from the residual signal combining 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 I₁ The 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 predictive coefficient interpolating unit 69 calculates the linear predictive coefficient between
Linear interpolation is performed according to (12), and the
Determine the number. The interpolation of this linear prediction coefficient is
Linear interpolation may be performed every time. Note that interpolation of linear prediction coefficients
Is generally performed in the LSP area.Example 2 Implementation in case of multi-stage quantization by PW quantization section 35 in FIG.
An example PW quantizer is shown in FIG. In FIG. 6, FIG.
Corresponding parts are denoted by the same reference numerals.
In the case of child conversion, a PW codebook 71 is further provided.
Code vector c selected from W codebook 71₁ ^jIs the sign
Pitch period length p at cutout 70_iOnly cut from the beginning
On the other hand, gain section 72 selects gain codebook 43 from gain codebook 43.
Selected gain g₁ ^kAnd convolution filter 73
And the impulse response H is convolved with
Is the error signal from the subtractor 45?
Is subtracted by a subtraction unit 74, and the remainder is calculated by a distortion calculation unit 75.
And the distortion calculator 75 calculates the square of the output of the subtractor 74.
Is the code vector c of the PW codebook 71 so that
₁ ^jAnd gain g of codebook 43₁ ^kSelection and the line
Will be Also in this case, the excitation
Signal and the PW waveform as the excitation signal
Minimizes the weighted mean-squared error with the perceived waveform
The code vector c₀ ⁱ, C₁ ^j, Gain g₀ ^k,
g₁ ^kIs determined. Equation (1) is used to calculate this distortion measure.
Use 4).

[0033] ^{_{D = ‖x-g o k Hc}} 0 it -g 1 k Hc 1 j ‖ ² (14) wherein, x target (target waveform vector) calculated by formula (4) is, H is quantized A matrix representing the impulse response of the synthesis filter 37 using the linear prediction coefficients a _j ′, 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,
Equation (14) is minimized from the PW codebook 71.
Code vector c₁ ^jAnd its ideal gain g₁ ^j
And c₀ ⁱG, the ideal gain of₀ ⁱRecalculate
You. This is the code vector c_o ⁱAnd c₁ ^jVector
And performs encoding. Based on this vector orthogonalization,
For more information on vector quantization based on
Coded speech coding method ”(Japanese Patent Application No. 6-43519).
ing.

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

[0036]

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

[0037]

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

[0038]

(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-quantized. The decoding unit in this case is the same as that in FIG. 4, but uses Equation (18) for decoding the PW waveform.

[0039] ^{_{r i = g 0 i c o}} i + g 1 i c 1 i (18) In the above, the adaptive codebook as shown in Figure 5B the PW codebook (a), a fixed codebook (b), algebraic It is possible to use any of the dynamic pulse codebooks (c). 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 A codebook (2) of a conjugate structure is used as the PW quantization unit 35 in FIG.
3B is shown in FIG. 3B, and the same reference numerals are given to portions corresponding to FIG. 2B. P
A W codebook 81 is further provided. From each code vector of the PW codebook 81, a code vector having a conjugate structure with the code vector of the PW codebook 41, that is, a code vector orthogonal to each other is selected. from cut by the pitch period length p _i min, the gain that is selected from the gain codebook 43 is given by the gain section 82, and the code vector of the from the code vector and the gain section 42 that the gain is given by an adder 83 The added signals are supplied to the convolution filter 44 as excitation signals.
A waveform obtained by combining this code vector as an excitation signal and P
The PW codebooks 41 and 81 are designed to minimize the auditory weighted mean square error with the waveform synthesized from the W waveform as the excitation signal.
Are determined and their gains are determined. For calculating the distortion distance scale of this distance, the equation (14) is used similarly to the second embodiment.
Is used. The details of the encoding method using the conjugate structure codebooks 41 and 81 are described in “Multi-vector quantization method and apparatus” (Japanese Patent Application No. 63-249450).

Also in this case, an adaptive codebook, fixed codebook, and 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 a plurality of types shown in FIG.
For example, a combination of an adaptive codebook and a fixed codebook may be used. The PW decoding unit 56 in FIG. 4 for the multi-stage vector quantization and the conjugate structure vector quantization prepares a codebook corresponding to the number of input code vectors, and codes corresponding to the input PW code I ₄ from these codebooks. extraction vector, respectively, and to them, each corresponding gain obtained from the gain codebook by the gain code in the input PW code I ₄ may be given respectively. The PW vectors thus decoded are each added, linearly interpolated with the added PW vector of the previous frame, and connected successively to supply a continuous signal to the synthesis filter 65.

In FIG. 1, the code I ₁ may be output after the quantization of the linear prediction coefficient is also performed for one pitch period.

[0042]

As described above, according to the coding method of the present invention, only one pitch period in one frame is coded in a voiced section. Can be reduced. In addition, since the peak positions are aligned at the time of encoding, 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. However, an intermediate PW vector is generated by interpolating between two preceding and succeeding code vectors. By generating an intermediate pitch period interpolating between the two preceding and succeeding decoding pitch periods, and subsequently linking the respective code vectors at the corresponding pitch period, a continuously changing excitation signal is obtained. Can be excited to reproduce sound.

Obtaining all of the selection of codes and the synthesis of signals at the time of decoding in the real time domain can be realized with a smaller amount of calculation than that performed in the frequency domain. In order to examine the effects of the speech encoding and decoding methods of the present invention, an analysis-synthesis speech experiment was performed under the following conditions. As input voice, 0-4k
After sampling the Hz-band sound at a sampling frequency of 8.0 kHz, a signal passed through an IRS characteristic filter corresponding to the characteristics of the telephone was used. Encoder and Decoder are Embodiment 2
(FIGS. 1, 6, and 4) were used. First, 25 ms (200 samples) is added to this input audio signal.
Each time, multiply the audio signal by a Hamming window with an analysis window length of 30 ms,
The linear prediction analysis by the autocorrelation method is performed with the analysis order set to 12, to obtain 12 prediction coefficients. The prediction coefficient is LSP
Vector quantization is performed using the Euclidean distance of the parameter.

When the state of the input speech signal is determined to be a voiced part, the obtained PW vector is converted into two noise codes c.
Vector quantization is performed using ₀ ⁱ and c ₁ ^j . The pitch determined by the partial autocorrelation method is scalar-quantized by rounding to an integer value. If the input voice signal is determined to be unvoiced, the 25 ms frame is divided into five and the average power of the residual waveform in each 5 ms subframe is calculated.
Vector quantize the two values.

The bit rate is 2.0 when there is periodicity.
0 kbit / s, 1.26 kbit when there is no periodicity
/ S, and the breakdown is as follows. Parameter Number of bits / frame Prediction coefficient (LSP) 21 Voiced / unvoiced parameter 1 Excited signal (for voiced) First-stage noise sequence 7 Second-stage noise sequence 7 Gain of noise sequence 7 Pitch period 7 Excited signal (unvoiced Case) Noise sequence 7 Speech coded under the above conditions has much higher naturalness than a conventional vocoder of the same bit rate,
Also, compared to the conventional CELP coding at the same bit rate, voice quality that is clear and has little noise is achieved.

[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 portion quantization unit 19 in FIG. 1, and FIG. 2B is a block diagram showing the unvoiced portion quantization unit 1 in FIG.
FIG. 9 is a block diagram showing another specific functional configuration of the ninth embodiment.

3A is a block diagram illustrating a specific functional configuration example of a PW quantization unit 35 in FIG. 1. FIG. 3B is a block illustrating a specific functional configuration example of the PW 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 a PW quantization unit 35 in FIG. 1 is a multi-stage vector quantization method.

Claims (15)

[Claims] An excitation for driving a linear prediction synthesis filter of a linear prediction coefficient obtained by analyzing a voice signal for each frame longer than a pitch period of the voice, and a filter coefficient based on the linear prediction coefficient. In a speech coding method for expressing speech characteristics by using a signal, voiced / unvoiced section discrimination is performed for each frame, and if the frame is a voiced section, the speech signal is calculated from a residual signal obtained by performing linear prediction inverse filtering. , A residual signal vector having a pitch period length is extracted, and the extracted residual signal vector is circulated so as to increase the correlation with a predetermined reference signal vector to obtain a target residual vector. A target waveform vector is obtained from the residual vector through the synthesis filter, and a plurality of predetermined code vectors are truncated at the pitch period length. And a synthesized waveform vector obtained by performing voice synthesis using the synthesis filter as the excitation signal, and selecting the code vector that minimizes distortion of the waveform of the synthesized waveform vector with respect to the target waveform vector, and performs quantization coding. Is determined. 2. The method according to claim 1, wherein the target waveform vector is generated by adding a non-square matrix extended downward to obtain a free response of the filter to the target triangular square matrix based on the impulse response of the synthesis 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, and the non-square matrix is created by truncating the impulse response with a small number of samples. The speech encoding method according to claim 2. 4. The speech encoding method according to claim 1, wherein the plurality of predetermined code vectors have the same peak position. 5. If the voiced / unvoiced section determination is a voiceless section, the synthesis filter is excited by a noise vector selected from a noise codebook, and a noise vector that minimizes distortion between an output signal and an input speech signal is obtained. 5. The speech coding method according to claim 1, wherein the residual signal is quantized by selecting. 6. If the voiced / unvoiced section determination is a voiceless section, one frame is divided into a plurality of subframes, and the average power of each of the subframes is vector-quantized for each frame. 5. The speech encoding method according to claim 1, wherein encoding is performed. 7. A linear prediction synthesis filter having a filter coefficient obtained by inputting a linear prediction coefficient code coded for each frame and a quantization code of an excitation signal and decoding the linear prediction coefficient code. In a speech decoding method for synthesizing an output speech by driving with a decoded signal of the quantization code of the excitation signal obtained from the codebook, if the frame is a voiced section, the speech decoding method according to the quantization code of the excitation signal The code vector extracted from the codebook is truncated at the pitch period length obtained by decoding the input pitch period code, the two code vectors before and after the truncation are interpolated, and the interpolation between the decoded two pitch periods is performed. And generating an excitation signal having the frame length by sequentially connecting the interpolated code vectors in accordance with the interpolated pitch period. Encoding method. 8. The speech decoding method according to claim 7, wherein said interpolation is weighted linear interpolation. 9. The interpolation is characterized in that a shorter vector length of the two preceding and succeeding code vectors is zero-filled with a difference portion so as to match the longer vector length, and then linear interpolation is performed. The speech decoding method according to claim 7. 10. The interpolation according to claim 1, wherein the lengths of the two preceding and succeeding code vectors are normalized to a constant vector length, linear interpolation is performed, and thereafter, the pitch period length is returned to the pitch period length before the normalization. 9. The speech decoding method according to 7 or 8. 11. When the pitch periods obtained from the preceding and succeeding two frames are significantly different from each other by about twice the pitch period length of the other, the interpolation corrects the shorter pitch length by a factor of two and replaces the shorter vector with the shorter vector. 11. The speech decoding method according to claim 7, wherein interpolation is performed after repetition. 12. The filter coefficient of the synthesis filter according to claim 7, wherein the two preceding and succeeding decoded linear prediction coefficients are linearly interpolated simultaneously with the interpolation of the waveform vector to obtain a filter coefficient of the synthesis filter. Audio decoding method. 13. The speech decoding method according to claim 7, wherein the decoded linear prediction coefficients are obtained by linear interpolation for each subframe. 14. When the frame is a voiceless section, a quantized code of an input excitation signal is decoded to obtain a voiceless residual waveform, and the synthesis filter is driven by the voiceless residual waveform. 14. The audio decoding method according to claim 7, wherein: 15. If the frame is an unvoiced section, each power of a plurality of subframes obtained by dividing one frame from the power codebook is applied to each sub-output of the synthesis filter excited by generated white noise. 14. The speech decoding method according to claim 7, wherein the output speech is obtained by matching the average power of the frames.