JPH0446440B2 - - Google Patents

Description

[Detailed description of the invention]

The present invention relates to an audio signal processing method that passes an audio waveform through a filter that removes its correlation to obtain a predicted residual waveform, and uses the predicted residual waveform. <Prior Art> Conventionally, there are two classes of speech coding: waveform coding and analysis and synthesis systems (vocoders). In the linear predictive coding (LPC) method, which belongs to the latter class of analysis and synthesis systems, as shown in FIG. is determined by linear predictive analysis, the voice waveform is passed through an all-zero filter (inverse filter) 12 with the opposite characteristics to obtain a predicted residual waveform, and the periodicity as a parameter characterizing this residual waveform is determined. The presence/absence (voiced/unvoiced determination), pitch period, and average power are extracted by the parameter extraction unit 13, and these and the prediction filter coefficients are sent out. On the synthesis side, instead of the predicted residual waveform, the sound source generator 14 uses a periodic pulse train in the case of voiced and a noise waveform in the case of unvoiced to drive the prediction filter 15.
A voice waveform is generated by setting a filter coefficient of 5 using the filter coefficient and outputting it to the output terminal 16. On the other hand, in adaptive coding (APC), which belongs to the former class of waveform coding, as shown in Fig. 2, after obtaining the predicted residual waveform using the same means as the LPC vocoder,
The sample value of this residual waveform is directly converted to the quantization unit 1.
7, and sends out this and the predictive filter coefficients. On the synthesis side, the residual waveform decoded by the decoder 18 is used to drive the prediction filter 15 to generate a speech waveform. The difference between these two conventional methods lies in the method of encoding the prediction residual waveform. The LPC vocoder shown in Figure 1 only needs to transmit the characteristic parameters of the residual waveform, so the bit rate is significantly lower than the APC method shown in Figure 2, which transmits the quantized value for each sample of the residual waveform. This can be reduced. but,
On the other hand, the method shown in Figure 1 inevitably suffers from quality deterioration due to replacing the residual waveform with a pulse train or noise.
There is a drawback that the quality is saturated at about s, making it impossible to provide natural voice quality, and erroneous extraction of feature parameters of the residual waveform causes quality deterioration. On the other hand, with the SPC method shown in Figure 2, by increasing the number of quantization bits of the residual waveform, it is possible to achieve audio quality that is as close as possible to the original audio, but on the other hand, when the bit rate becomes less than 16 kb/s, quantization distortion increases. The drawback was that the voice quality deteriorated rapidly. Furthermore, in the past, when changing the pitch of an audio signal or adding audio signals, there was a risk that the change would be performed at a location where energy is concentrated, and in that case, there was a drawback that the result would be unnatural. <Structure of the Invention> An object of the present invention is to provide an audio signal processing method that makes it possible to obtain naturalness when adding audio signals, for example. Another object of this invention is that the bitrate is 16kb/
An object of the present invention is to provide an audio signal processing method that can maintain relatively good audio quality even when the audio quality is less than s. According to this invention, a speech waveform is subjected to linear predictive analysis to obtain a predicted residual waveform, and this predicted residual waveform is filtered by a linear filter (phase equalization A filter coefficient of a filter) is adaptively determined from the residual waveform, and the speech waveform or the predicted residual waveform is passed through the phase equalization filter to zero phase the predicted residual waveform, that is, the phase is equalized. This phase-equalized predicted residual waveform energy concentrates in an impulse manner, and therefore, by adding, for example, a voice waveform in a portion where the energy is not concentrated, a voice waveform with good naturalness can be obtained. Furthermore, when encoding the phase-equalized speech waveform or prediction residual waveform, for example, by allocating a lot of information to the part where the energy is concentrated, it is possible to encode efficiently at 16kb/s. You can get pretty good audio quality even below. <Principle of the Invention> First, the principle of the audio signal processing method according to the present invention will be described. The sample value of the audio waveform is
(n), and the prediction coefficient obtained by linear predictive analysis of the speech waveform is a(k) (k=1, 2...P), then the sample value e(n) of the prediction residual waveform is expressed by the following equation. It will be done. e(n)= _p 〓 ^k=0 a(k)・S(nk)...(1) However, a(0)=1. Residual waveform e(n)
is the voice waveform with the spectral envelope component removed, that is, the correlation between the sample values of the voice waveform is removed, and has a flat spectrum envelope, and for voiced sounds, it has the pitch period component of the voice. Therefore, the characteristics of such a residual waveform are idealized and realized as the following pulse train. e _M (n)= _L-1 〓 ^l=0 δ(n-nl) ...(2) Here, δ(n) is Kronetzker's delta function, δ(0)=1, and δ(n) =0 (n≠0). n _l represents the pulse position, and n _l −n _l-1 corresponds to the pitch period of the voice. In other words, this pulse train e _M
In (n), there is a pulse only at the pitch position _nl , and there are zeros at the other positions. Residual waveform e(n) and pulse train
Since e _M (n) both have a flat spectral envelope and a pitch periodic component, the difference between the two waveforms is mainly for a short period of time,
In other words, this is due to the difference in phase characteristics at times less than the pitch period. Therefore, the impulse response of a linear filter with inverse characteristics of the short-term phase of the residual waveform is h
(m), the residual waveform e(n) is passed through this linear filter (phase equalization filter) to equalize the phase (zero phase), that is, the residual waveform ep(n) in which each component of the spectrum is made into the same phase. is calculated using the following formula. ep(n)= _M 〓 ^m=0 h(m)e(n-m) ...(3) This impulse response h(m) is calculated by e _P (n) and e _M
(n) by minimizing the mean square error. The mean square error is expressed by the following equation. J=1/N _N-1 〓 ⁿ⁼⁰ {e _P (n)−e _M (n)} ² ...(4) Substituting equations (2) and (3) into equation (4), h (m) and set it to zero, the impulse response h(m) can be obtained as a solution of the following simultaneous equations. _M 〓 ^k=0 v(|m−k|) h(k)= _L−1 〓 ^l=0 e(nl−m)
...(5) (m=0,1,...;M) Here, v(k) is an autocorrelation function of the residual waveform and is calculated by the following equation. v(k)= _Nk-1 〓 ⁿ⁼⁰ e(n) e(n+k) ...(6) (k=0, 1,...M) The time corresponding to the number of taps M+1 of the phase equalization filter, that is, When the response time is shorter than the pitch period, the autocorrelation function can be approximated as v(k)v ₀ δ(k) because the residual waveform has a flat spectrum. In other words, since it has a value only when k=0, in that case, equation (5) has a value only when m=k, and can be simplified as follows. h(m)=1/v _0L-1 〓 ^l=0 e(n _l −m) ...(7) Furthermore, if the analysis window length N is shorter than the pitch period, L=1 (one pulse and ), the impulse response is calculated using the following formula. h(m)=1/v ₀ e(n ₀ −m) (8) That is, the impulse response h(m) is the time axis of the residual waveform inverted with the origin at time n ₀ .
Further, assuming that the power spectrum of the residual waveform is completely white (amplitudes of all frequency components are constant), the Fourier transform of the impulse response h(m) is expressed by the following equation. (However, the gain is normalized) Here, E(k) represents the Fourier transform of the residual waveform e(n). Therefore, the Fourier transform Ep(k) of the phase-equalized residual waveform ep(n) is given by equation (3).
Ep(k)=H(k)・E(k), and E(k)=
Since |E(K)| _e ^a argE(k), the following equation is obtained by substituting equation (9) into this. From equation (10), the phase-equalized residual waveform ep(n) is a linear phase component.

Except for [Formula], the residual waveform e(n) is made to have zero phase (all spectra are made to be in the same phase). Ideally |E(k)|=E ₀ (constant)
Then, ep(n) becomes completely phaseless and has a single pulse waveform. In short, the phase equalization filter residual waveform e(n) with the filter coefficient h(m) as described above
When passed through, the energy is mainly concentrated at the pitch position, that is, the waveform becomes close to a single pulse. <First Embodiment> Next, a specific embodiment of the audio signal processing method of the present invention will be described with reference to FIG. From the input terminal 11, the sample value S(n) of the sampled audio waveform is input.
is input and supplied to the linear prediction analysis section 21 and the inverse filter section 22. The linear predictive analysis unit 21 uses linear predictive analysis from the audio waveform S(n) to calculate (1)
Calculate the prediction coefficient a(k) in the equation. The inverse filter section 22 receives the audio waveform S(n) as input.
A filtering operation as shown in equation (1) is performed and a predicted residual waveform e(n) is output. Prediction residual waveform e
(n) is supplied to the voiced/unvoiced determining section 24, the pitch position detecting section 25, and the filter coefficient number calculating section 26 in the filter coefficient determining section 23. The voiced/unvoiced determination unit 24 determines the autocorrelation function of the residual waveform e(n) using a fixed number of delay sample points, and if the maximum peak value is above a fixed threshold value, it is voiced, and if it is less than that, it is determined to be voiced or unvoiced. Make a judgment. This determination result V/UV is used to control the processing mode for determining the subsequent phase equalization filter coefficients. Since the phase equalization filter adapts to temporal changes in the phase of the residual waveform, it adapts to each pitch period in the voiced part. Now, assuming that time n is at the l-1th pitch position n _l-1 , the phase equalization filter coefficient at that time is h ^* (mn _l-1 ) (m = 0, 1...M)
Expressed as The pitch position detection section 25 detects the next pitch position nl using the pitch position nl _-1 and the filter coefficient ^* (m, _nl-1 ). FIG. 4 shows the internal configuration of the pitch position detection section 25. The residual waveform e(n) from the inverse filter section 23 is inputted from the input terminal 27, and the voiced/unvoiced determination result V/UV from the voiced/unvoiced determining section 24 is inputted from the input terminal 28. Processing mode switch 2
At step 9, the processing mode is switched in accordance with the voiced/unvoiced determination input V/UV. In the case of voiced V, the residual waveform e
(n) is input to the phase equalization filter unit 31, and the filter coefficient h ^* (m,
A convolution operation (same operation as equation (3)) is performed between n _l-1 ), and the phase-equalized residual waveform e _p (n)
is output. The relative amplitude calculation unit 33 calculates the relative amplitude of the phase-equalized residual waveform e _p (n) at time n using the following equation. The amplitude comparator 34 compares the relative amplitude m _e p (n) with a predetermined threshold value m _th and determines that m _e p (n)>m _th (n>n _l-1 ).
. . . If (12) is satisfied, the time point n is outputted to the output terminal 35 as the pitch position nl. Next, the pitch position nl is input to the filter coefficient calculation unit ²⁶ in FIG. The signal is supplied to a filter coefficient interpolation section 37 and a phase equalization filter section 31 in FIG. However, compared to equation (8), equation (13) normalizes the filter gain, and also normalizes the linear phase component (in equation (10)).

The delay in [Formula] has been corrected. In other words, as is clear from equation (10), h(m) obtained from equation (8) is delayed by M/2 samples from the actual one, so equation (13) is used. On the other hand, if the voiced/unvoiced determination result is unvoiced UV, the processing mode switch 29
(n) is input to the pitch position reset unit 36 to set the pitch position nl to the last sample time within the analysis window, and the filter coefficient calculation unit 26 sets the filter coefficient h ^* (m, nl) = 1 ( m=0)h ^* (m,
nl) = 0 (m≠0). The filter coefficient h(m, n) at each time point n is calculated by the filter coefficient interpolation unit 37 as a value smoothed using, for example, a first-order filter expressed by the following equation. h(m,n)=αh(m,m-1)+(1+α)h ^*
(m, nl) (n _l-1 <n≦nl) (14) Here, α is a coefficient that controls the rate of change of the filter coefficient and is a constant that satisfies α<1. The phase equalization filter unit 38 uses the input audio waveform S(n) of the input terminal 11 and the filter coefficient interpolation unit 37
Using the filter coefficients h(m, n), a convolution operation shown by the following equation is performed, and a phase-equalized audio waveform S _p (n) is output to the output terminal 39. S _p (n) = _M 〓 ^m=0 h (m, n) S (n - m) ... (15) <Second Example> Next, the phase-equalized speech waveform S _p (n) is Digital encoding will be explained. An example of the basic configuration of this encoding is shown in FIG. The audio waveform S(n) input from the input terminal 11 is subjected to phase equalization processing in the phase equalization processing section 41 having the configuration shown in _FIG . The encoding unit 42 outputs this phase equalized speech waveform S _p
(n) is digitally encoded and the code sequence is sent to the transmission path 43. On the receiving side, the decoding unit 44 restores the phase-equalized speech waveform S _p (n) and outputs it to the output terminal 16 . In this way, encoding and decoding are performed using the phase-equalized speech waveform S _p (n) instead of the speech waveform S(n). The audio waveform S _p (n) obtained by phase-equalizing the audio waveform S(n) is the same in quality as the original audio waveform S(n).Therefore, there is no need to transmit the filter coefficient h(m), and the phase etc. It is only necessary to reproduce the converted speech S _p (n). In particular, energy is concentrated in the residual waveform e _p (n) obtained by phase equalizing the residual waveform e (n), so
By adaptively encoding the data to give more information to that part, high-quality transmission is possible with a smaller number of bits. Various methods can be applied as the encoding method in the encoding unit 42. Here, we will discuss 3 encoding methods suitable for phase-equalized speech waveforms.
An example is shown below. Method using variable rate tree encoding The variable rate tree encoding method adaptively controls the amount of information according to the amplitude change in the temporal direction of the predicted residual waveform obtained by linear predictive analysis of the speech waveform. This is a unique encoding method. FIG. 6 shows an embodiment of an encoding system that combines phase equalization processing and variable rate tree encoding according to the present invention. Input terminal 1
The linear predictive analysis unit 21 performs linear predictive analysis on the audio waveform S(n) input from 1 to calculate the prediction coefficient ak, and the inverse filter unit 22 calculates the predictive coefficient ak.
Find the predicted residual waveform e(n) of (n). The filter coefficient determination unit 23 calculates the coefficient h(m, n) of the phase equalization filter that equalizes the short-term phase of the residual waveform e(n) as described in FIG. is set as a filter coefficient of the filter unit 38. The phase equalization filter unit 38 subjects the input audio waveform S(n) to phase equalization processing, and outputs the phase-equalized audio waveform S _p (n) to the terminal 39 . On the other hand, after the residual waveform e(n) is phase-equalized by the phase equalization filter section 45, the partial interval setting section 46 sets partial intervals for dividing the time axis according to the deviation of the residual waveform amplitude, and The calculation unit 47 calculates the power of the residual waveform in each of the set partial intervals. For example, as shown in Fig. 7, the partial interval can be set by setting the pitch position (nl) interval within the analysis window (however, only one sample point) and the pitch period T _p
Set as each interval to be divided into equal parts. The residual power u _i in the subinterval is calculated using the following formula. u _i = 1/N _Ti 〓 n∈Tie ² _p (n) ...(16) Here, T _i represents the subinterval to which sample time n belongs, and N _Ti is the number of sample points included in the subinterval. be. The bit allocation unit 48 calculates the number of information bits R(n) to be allocated to each sample time from the residual power u _i of each subinterval using the following equation. Here, is the average bit rate for the residual waveform e _p (n), N _s is the number of subintervals, w _i is the time length ratio of the subintervals, and is given by w _i =N _Ti / _NS 〓 ^j=1 N _Tj It will be done. Also, the quantization step width △(n)
is calculated by the step width calculating section 49 from the residual power u _i using the following formula. Δ(n)=Q(R(n))√ _i n 〓T _i (18) Here, Q(R(n)) is the step width of the R(n)-bit Gaussian quantizer. The number of bits R(n) and the step width Δ(n) calculated by the bit allocation section 48 and the step width calculation section 49 control the tree code generation section 51. As shown in FIG. 8, the tree code generation unit 51 has a variable rate tree structure, and has a code sequence C(n)={C(n-L), . . . , C(n-1), C
(n)}, sample values q(n) associated with each technique are output along the path determined by .
The number of techniques coming out of each node is given as 2 ^R(n) . In addition, the sample value f associated with each technique
(l,n) is given by the following equation from Δ(n) and R(n). f(l,n)=Sgn(l)|l|+0.5/2△(n), l=±1,±2...,±2 ^R(n)-1 (19) Here, Sgn( l) represents the positive or negative sign of l. Furthermore, q(n) is given as q(n)=5(l ^* ,n), where l ^* is the technique on the pass. Sample value p(n) output from tree code generation unit 51
is input to the prediction filter section 52, and the local decoded value S^ _p (n) is calculated using the following equation using an all-pole filter. S^ _p (n) = _p 〓 ^k=1 a(k) S^ _p (n-k) + q(n) ...(20) Here, a(k) is the prediction coefficient, and the linear prediction analysis section It is controlled by the output from 21. Locally decoded value S^ _p (n) and phase-equalized speech waveform S _p
(n) is the difference between both values in the subtracter 53,
It is input to the code sequence optimization section 54. The code sequence optimization unit 54 calculates the path of the tree code, that is, the code _sequence C( _n )={C(n−L ),...C(n-1)
，
C(n)}. For example, an ML algorithm is used as the optimal path search method. In the ML algorithm, in a tree code as shown in Figure 8, code sequence candidates are C _n (n) = {C _n (n - L),...
..., C _n (n-1), C _n (n)} (m=1, 2, ...
M′) is the error evaluation value d at each node.
(m, n) is the sample value S _p (n) given for the code sequence candidate C _n (n) and the input sample value S _p
The squared error between the time series of (n) is calculated using the following formula. d(m,n)= _M 〓 ^t=nL {S _p (t)−S^ _p (t)} ^Second , evaluate the evaluation value d from among the M′ code sequence candidates.
Select the code sequence C _n (n) for which (m, n) is the minimum, and select the code sequence C _n (n
-L) is determined as the optimal code. Candidate code sequence at time n+1 C _n (n+1)={C _n (n+1−
L),...C _n (n), C _n (n+1)} is d(m, n)
After selecting M code sequences C _n (n) in descending order of the value of , all possible codes C (n+1) at time point n+1 are added to each code sequence to give a sequence. The above processing is performed sequentially at each point in time,
At time n, the optimal code C(n
-L) is output. Note that the symbol * in Figure 8 is
Indicates a null sign. In the encoding method in this embodiment, the prediction coefficient ak output from the linear prediction analysis unit 21 as auxiliary information together with the code C(n) of the residual waveform, and the period T _p of the subinterval output from the subinterval setting unit 46
and the position T _d , and the partial section residual power u _i output from the power calculation section 47 are multiplexed by the multiplexing section 55 and then sent to the transmission path 43 . On the receiving side, after demultiplexing each piece of information in a demultiplexing unit 56, a residual waveform generating unit 57 calculates a decoded value q(n) of the residual waveform according to the code sequence, and the decoded value q(n) The prediction filter 15 is driven using the driving sound source information to restore the speech waveform S _p (n) and output it to the output terminal 16 . By equalizing the phase of the residual waveform e(n), it is made into a pulse, that is, the energy is increased, and by allocating a large number of bits to that part and increasing the number of tree codes, it can be made efficiently at a small bit rate. Information can be transmitted. Method using multipulse coding The basic principle of multipulse was described by Atal in 1982.
International Conference on Acoustics and Speech Signals (Proceedinhg)
proposed in ICASSPpp.614-617). In this method, the predicted speech residual waveform is represented by multiple pulse trains, and the temporal position and intensity of each pulse are minimized to minimize the error between the speech waveform synthesized using this multi-pulse residual waveform and the input speech waveform. This method determines the In this method, the speech waveform itself is directly encoded, but in the embodiment of the present invention, the speech waveform after being phase-equalized is input and multipulse encoding is performed. FIG. 9 shows an embodiment of an encoding method that combines phase equalization processing and multi-pulse encoding. The linear prediction analysis unit 21 calculates a prediction coefficient for the sample value S(n) of the audio waveform input from the input terminal 11, and the prediction inverse filter unit 22 calculates the prediction residual waveform of the audio waveform S(n). demand. Next, the filter coefficient determination unit 23 uses the residual waveform e(n) to determine the phase equalization filter coefficient n(m,
n) and pitch position nl. The filter coefficient of the phase equalization filter section 38 is set to h (m, n), and the phase equalization filter section 38 inputs the audio waveform S _N
is phase-equalized, and the subtracter 53 calculates the difference between the output and the multi-pulse encoded value S^ _p (n).
is input to. The encoded value S^ _p (n) is the multi-pulse signal e^e output from the multi-pulse generator 61
By passing (n) through the prediction filter 62, it is calculated using the following formula. S^ _p (n) = - _p 〓 ^k=1 akS^ _p (n-k) + e(n) where e^e (n) is the pulse time t _i . It is expressed by the following equation, where the pulse amplitude is m _i . e^e(n)=δ(t-t _i ) The pulse time point t _i and the pulse amplitude m _i are calculated as the waveform value S _p (n) and S^ _p in the pulse time point calculation section 58 and the pulse amplitude calculation section 59, respectively. The average power of the difference between (n)
It is determined to minimize P _e . In the algorithm presented in the above-mentioned paper,
Given t _i and m _i , the lth pulse position tl
is all possible time points (where t _l ≠t _i (i=1...l
−1)), the pulse amplitude m _i at which the average power P _e is the minimum is determined by the least squares method, and P _e
is determined as the point at which is the minimum. This procedure is performed sequentially starting from l=1 until l=q, and all pulse times and amplitudes are determined. This algorithm requires a large amount of processing to calculate the pitch point. However, in this embodiment, in order to reduce the amount of processing, the pitch position n _i determined during the phase equalization process is
(i=1, 2...q'), the first q' pulse points are determined as t _i =n _i (i=1, 2...q'). The prediction coefficient ak, pitch time point (position) t _i and pitch amplitude m _i are multiplexed by a multiplexer 55 and sent out;
On the receiving side, after demultiplexing these signals using a demultiplexer 56, a multipulse generating section 63 generates a multipulse signal, which is passed through a prediction filter 15 to obtain an encoded signal output at a terminal. Speech analysis and synthesis system using pulsed residual waveform In this example, in the sample value time series of the predicted residual waveform whose phase has been equalized by the phase equalization process described above, the sample value at the pitch position is left,
By setting the other sample values to zero, the predicted residual waveform is made into a pulse, and a synthesized speech is generated by driving a prediction filter using this pulse train as a driving sound source. That is, as shown in FIG. Sample value S of the audio waveform input from the input terminal 11
(n), the linear prediction analysis unit 21 calculates the prediction coefficient ak
After calculating, the predicted residual waveform e(n) of the speech waveform S is obtained by the predicted inverse filter 22. Next, the filter coefficient determination unit 23 calculates the phase equalization filter coefficient h(m, n), the voiced/unvoiced determination value V/UV, and the pitch position n _l from the residual waveform e(n). After the residual waveform e(n) is phase equalized by the phase equalization filter section 64, the residual waveform e(n) is subjected to phase equalization in the pulse processing section 65.
Let the sample value of the phase equalized residual waveform e _p (n) at the pitch position n _l be m _l =e _p (n _l ) (l=1, 2...L). Here L is the number of pitch positions within the analysis window. The sample value m _l is quantized by a quantizer 67 using a quantization step width Δ given from a quantization step width calculating section 66 . The multiplexing unit 55 outputs the quantized output C(n), the pitch position n _l , the prediction coefficient ak,
The voiced/unvoiced judgment value V/UV and the residual power v are multiplexed and sent. The demultiplexer 56 demultiplexes the
The voiced part 68 dequantizes the quantized output C(n),
From this and the pitch position and n _l , the pulse train e^ _p (n) = _L 〓 ^l=1 m _l
Make 〓(n−n _l ). In the silent section 69, white noise whose power is equal to v is used as a driving sound source. The switch 71 is controlled according to the voiced/unvoiced judgment value V/UV to supply the output of the voiced part 68 for voiced V and the output of the unvoiced part 69 for unvoiced UV to the prediction filter 15 as driving sound source information, and to generate synthesized speech S. ^ _p (n) output terminal 16
Output to. <Effects> As described above, the audio signal processing method according to the present invention adaptively equalizes the phase of the short-time phase characteristic of the predicted residual waveform according to its temporal change, thereby improving the amplitude of the residual waveform. It has the effect of increasing the temporal concentration, thereby making it possible to detect the pitch period and pitch position of the audio waveform, and for example, to shorten the time by removing parts where energy is not concentrated, or to reduce the time to zero. Even if the pitch of the speech waveform is changed by inserting it for a longer time, the naturalness can be maintained, and the coding efficiency can be greatly improved. The audio quality when only phase equalization processing is performed is equivalent to 7.6-bit logarithmic companding PCM,
Waveform distortion due to this processing is hardly perceptible.
Therefore, even if a phase-equalized audio waveform is input to be encoded, no quality deterioration occurs at the input stage. Furthermore, if the phase-equalized audio waveform can be correctly reproduced, high-quality audio can be obtained by using the phase-equalized audio waveform as a driving sound source signal. In all of the encoding methods shown in the above embodiments, the encoding efficiency is improved by increasing the degree of temporal concentration of the amplitude of the speech prediction residual waveform. In variable rate tree encoding, information is allocated temporally according to the deviation of the waveform amplitude, and by increasing the deviation through phase equalization, the effect of information allocation becomes larger and the coding efficiency improves. When encoded with 1 bit and 1 sample (approximately 10 kb/s), the SN ratio of encoded speech is 19.0 dB, which is 4.4 dB better than when phase equalization processing is not included. In terms of quality, the quality equivalent to 5.5-bit PCM is improved to the quality equivalent to 6.6-bit PCM. Since 7-bit PCM has no quality problems, relatively high quality can be obtained in this example even at a bit rate of 16 kb/s or less. In multi-pulse encoding, the residual waveform is converted into pulses through phase equalization processing, so the multi-pulse representation is more suitable, and the residual waveform can be generated with fewer pulses than when using the conventional input audio itself. I can express it. In addition, many of the pulse positions in multi-pulse encoding match the pitch positions in this phase equalization process, so using this pitch position information simplifies the pulse position determination process in multi-pulse encoding. can do. The performance of multi-pulse encoding when the number of pulses is 20 (equivalent to 1-bit 1-sample encoding, approximately 10 kb/s) is the SN ratio for direct audio input.
11.3dB, 15.0dB for phase equalized audio,
Phase equalization processing improves the SN ratio by 3.7dB. Furthermore, in terms of quality, the quality equivalent to 4.5-bit PCM is improved to that equivalent to 6-bit PCM by phase equalization processing. Conventionally, when the bit rate becomes 16 kb/s or less, the voice quality deteriorates rapidly, but even when this multi-pulse encoding is applied, quite good voice quality can be obtained even at a bit rate of 10 kb/s. Note that h ^* (m, n _l ) may be used as the filter coefficient of the phase equalizable filter section 38 and the filter coefficient interpolation section 37 may be omitted. Further, each of the above-mentioned parts may be constructed by independent hardware or a microprocessor, or a plurality of parts may be combined by one microprocessor or electronic computer.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing the basic configuration of a conventional linear predictive analysis and synthesis method, FIG. 2 is a block diagram showing the basic configuration of a conventional adaptive predictive coding method, and FIG. 3 is a block diagram showing the basic configuration of a conventional adaptive predictive coding method. A block diagram showing an example of the configuration of the adaptive phase equalization processing method, FIG. 4 is a block diagram showing the internal configuration of the pitch position detection section 25, and FIG. 5 shows the basic configuration of speech encoding using phase equalization processing. 6 is a block diagram showing a configuration example of variable rate tree encoding using phase equalization processing, FIG. 7 is an explanatory diagram regarding the method of setting subintervals, and FIG. Fig. 9 is a block diagram showing a configuration example of multi-pulse encoding using phase equalization processing; Fig. 10 is an explanatory diagram showing the structure;
The figure is a block diagram showing an example of the configuration of a speech analysis and synthesis system using a pulsed residual waveform. 11...Input terminal, 21...Linear prediction analysis section,
22... Inverse filter section, 24... Voiced/unvoiced determination section, 25... Pitch position detection section, 26... Filter coefficient calculation section, 37... Filter coefficient interpolation section, 3
8... Phase equalization filter section, 39... Output terminal,
41... Phase equalization processing section, 45... Phase equalization filter section. 46... Partial interval calculation unit, 47... Power calculation unit, 48... Bit allocation unit, 49...
Step width calculation unit, 51...Tree code generation unit, 52
...Prediction filter section, 53...Subtractor, 54...
Code sequence optimization unit.

Claims (1)

[Scope of Claims] 1. Means for obtaining a sample value of a predicted residual waveform using a prediction filter that removes correlation between sample values of a voice waveform, and the predicted residual waveform or the voice waveform is supplied. a phase equalizing filter for reducing the phase of the predicted residual waveform to zero phase; means for determining a phase equalizing filter coefficient having a phase characteristic opposite to that of the predicted residual waveform from the predicted residual waveform; and means for adapting the phase equalization filter according to temporal changes in the phase of the predicted residual waveform. 2. The input audio signal is phase-equalized by the phase equalization filter, the phase-equalized output is encoded using a digital encoding method, and the result is directly used as an encoded audio output. Audio signal processing method. 3 The digital encoding method described above is a variable rate tree encoding method, and in order to control the number of branches (number of bits) emerging from each node of the tree code and the output sample value of the tree code assigned to each technique. 3. The audio signal processing method according to claim 2, wherein the necessary information is extracted from the prediction residual signal obtained during the phase equalization process. 4. The digital encoding system is a multi-pulse encoding system, and some of the pulse positions are determined as pitch positions obtained in the phase equalization process, as set forth in claim 2. audio signal processing method. 5. The digital encoding method is a means of creating a pulse train in which only sample values at pitch positions are left and other sample values are set to zero in the sample value time series of the phase-equalized predicted residual waveform. The audio signal processing method according to claim 2, wherein synthesized speech is obtained by driving a prediction filter using the pulse train as a driving sound source of voiced sound.