JP2508002B2 - Speech coding method and apparatus thereof - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a speech coding method and apparatus thereof, and particularly to a speech signal having a low bit rate of about 4.8 Kb / s (kilobits / second) and a high computation rate. TECHNICAL FIELD The present invention relates to a speech coding method and apparatus for quality coding.

[Conventional technology]

As a method for encoding a voice signal with a low transmission bit rate of about 4.8 Kb / s with high quality, for example, Japanese Patent Application No.
As described in 272435 (Reference 1) and Japanese Patent Application No. 60-178911 (Reference 2), a voice signal in a voiced section is expressed by a pulse train of one pitch section (representative section) and this pulse train is transmitted. It is known how to do it. According to this method, it is possible to obtain good sound quality without impairing the naturalness even at a low bit rate of about 4.8 Kb / s.

FIG. 3 is a pulse processing explanatory view showing an example of a conventional method for obtaining a pulse train of a representative section in a voiced section. Third
In the figure, (a) is a voice waveform of one frame,
(B) is an example in which the speech waveform of (a) is divided into subframes corresponding to the pitch period P'd.
And one frame is divided into sub-frame sections by utilizing the position of the _first pulse g ₁ . Further, (c) is an example in which a pulse train is obtained for each subframe, and (d) is an example of the representative section obtained by searching and an example of the pulse train in this representative section. . In this case, the representative section can be obtained by a method of searching a subframe section that can reproduce good audio in the entire frame or a method of selecting a section including a pulse having a large absolute value. As the sound source information, the position of such a representative section, that is, the position of the subframe in the case of FIG. 3 and the time information T of the start point of the subframe in the frame are transmitted.

[Problems to be solved by the invention]

However, in the above-described conventional method of this type, a large amount of calculation is required to search the pulse train or the representative section, and auxiliary information such as the position of the representative section and the start point of the subframe in the frame is required. Therefore, when trying to realize the hardware by using a general-purpose signal processor currently on the market, there is a drawback that it is difficult to realize with a simple device configuration especially for the former reason.

An object of the present invention is to eliminate the above-mentioned drawbacks and provide a speech coding method and apparatus capable of synthesizing high-quality speech even with a low transmission bit tray of about 4.8 Kb / s with a small amount of calculation and a simple device configuration. Especially.

[Means for solving problems]

On the transmission side, a discrete voice signal is input, a spectrum parameter representing a short-time spectrum envelope and a pitch parameter representing a pitch are extracted from the voice signal for each frame, and a frame section is divided into a plurality of periods equal to the period of the pitch parameter. Dividing into pitch sections, a sound source signal of one pitch section is represented by a multi-pulse train or a combination of noise and a multi-pulse train, and information representing the sound source signal, the pitch parameter and the spectrum parameter are combined and output, and the receiving side The frame is divided into pitch sections based on the pitch parameter, the excitation signal of one pitch section is restored based on the information indicating the excitation signal to restore the driving excitation signal of the entire frame, and the spectrum parameter is used. It is a configuration for synthesizing the audio signals.

A parameter calculation circuit that extracts and encodes a spectrum parameter that represents a short-time spectrum envelope and a pitch parameter that represents a pitch for each frame from an input speech signal, and a plurality of pitch intervals in which a frame interval is equal to the cycle of the pitch parameter. A drive signal calculation circuit that encodes the excitation signal by expressing the excitation signal of one pitch section by a multi-pulse train or a combination of noise and a multi-pulse train, and an output code of the parameter calculation circuit and the drive signal calculation circuit. And a multiplexer circuit for outputting the combined output code of the above.

〔Example〕

 The present invention will now be described in detail with reference to the drawings.

FIG. 1 (a) is a block diagram showing an embodiment of a voice coding method and its transmitting side according to the present invention, and FIG. 1 (b) is a voice coding method and its receiving side according to the present invention. It is a block diagram which shows one Example.

The transmitting side shown in FIG.
Pitch analysis circuit 130, K parameter calculation circuit 140, pitch coding circuit 150, K parameter coding circuit 160, impulse response calculation circuit 170, autocorrelation function calculation circuit 180, weighting circuit
200, drive signal calculation circuit 220, noise memory 225, encoding circuit 23
0, multiplexer 260, etc.
The receiving side shown in (b) is provided with a demultiplexer 290, a decoding circuit 300, a noise memory 310, a pitch decoding circuit 320, a K parameter decoding circuit 330, a drive signal restoration circuit 340, an interpolation circuit 335, a synthesis filter circuit 350, and the like. Consists of

The basic processing contents of the present invention performed by the transmitting side and the receiving side are as follows.

That is, in the voiced section, the present invention represents a sound source signal by using a pulse train of this pitch section as a representative section representing the frame by using a predetermined pitch section in the frame by utilizing the periodicity of the voice signal, and the unvoiced section. In, the source signal is represented by the combination of pulse train and noise.
In this case, in the voice section, the position of the representative section is determined in advance, and the pulse train is obtained only for that section, thereby significantly reducing the amount of calculation required for pulse search and representative section search. Moreover, auxiliary information such as information indicating the position of the representative section and information indicating the start point of a subframe within a frame is unnecessary, and the information to be added to the pulse train can be increased by this amount, and high quality can be achieved with a small amount of calculation. The audio can be reproduced and the hardware is realized with a compact device configuration.

As a method for obtaining the amplitude and position of the pulse train, in addition to the methods described in the above-mentioned documents (1) and (2), for example, ANALYSIS-by-SYNTHESI.
S: A-B-S) method is known, and the details are described in B.S.ATAL.
"A new model of LPC
"Excitement for Producing Natural Sounding Speech at Low Bitrates" (A NEW MODEL OFLPC EXCITATION FOR PRODUCI
NG NATURAL SOUNDING SPEECHAT LOW BIT RATES) (PROC.ICASSP, pp 614-617,1982) (Reference 3) and the like.

On the other hand, the method of obtaining the sound source by the combination of the pulse train and the noise in the silent section is described in detail in the method of (Reference 2).

Now, in FIG. 1 (a), the audio signal X (n) is input through the input terminal 100 and accumulated in the buffer memory 110 by a preset number of samples.

The K parameter calculation circuit 140 inputs a voice signal for each preset number of samples from the buffer memory 110 and calculates a K parameter representing a spectral envelope of the voice signal. The K parameter is a PARCOR (partial autocorrelation) coefficient. The autocorrelation method is well known as a method of calculating the K parameter, and the details thereof are described in John McCall.
AKHOUL) et al. “Quantization Properties of Transmission Parameters
Linear Predictive Systems ”(QUANTIZATI
ON PROPERTIES OF TRANSMISSION PARAMETERS IN LINER
PREDICTIVE SYSTEMS) (IEEETRANS.ASS
P.pp309-321, 1983) (Reference 4) and the like.
Returning to FIG. 1 (a) again, the description of the embodiment will be continued.

The i-th order K parameter Ki output from the K parameter calculation circuit 140 is output to the K parameter encoding circuit 160, encoded based on the preset number of quantization bits, and output to the multiplexer 260 as a code lKi. Further, the K parameter encoding circuit 160 uses the K parameter decoding Ki obtained by once decoding lKi to obtain the linear prediction coefficient value a′i and outputs it to the impulse response calculation circuit 170 and the weighting circuit 200.

The pitch analysis circuit 130 uses the output of the buffer memory circuit 110 to calculate the pitch period P′d. The calculation method of P'd is, for example, RVC OX
“Real time implementation”
Of Time Domain Harmonic Scaling
Of speech "(REAL-TIME IMPLEMEN-TATION OF TIM
E DOMAIN HAR-MONIC SCALING OF SPEECH SIGNALS) (IEEE TRANS.ASSP.pp.258-272.1983)
The method described in (Reference 5) or the like can be used. In this case, it is determined whether the voice signal of the frame is voiced or unvoiced by using the autocorrelation coefficient at the position separated by the pitch period.
If unvoiced, the pitch period is set to 0.

The pitch encoding circuit 150 quantizes and encodes the pitch period P′d with a predetermined number of quantization bits,
It is output to the multiplexer 260 as ld. At this time, when the voice is unvoiced, the code corresponding to 0 is output. Also, P′d obtained by decoding is output to the drive signal calculation circuit 220.

The impulse response calculation circuit 170 inputs the prediction coefficient value a′i from the K parameter encoding circuit 160 and represents the impulse response hw representing the transfer function of the weighted synthesis filter.
Calculate (n). Here, the same method as the impulse response calculation circuit 210 described in FIG. 4 (a) of Japanese Patent Application "Japanese Patent Application No. 59-042305" (Reference 6) is used for the calculation of hw (n). be able to. Impulse response hw (n)
Is an autocorrelation function calculation circuit 180 and a cross correlation function calculation circuit 2
Output to 10.

The autocorrelation function calculation circuit 180 receives the impulse response hw (n) from the impulse response calculation circuit 170, calculates the autocorrelation function Rhh (m), and outputs it to the drive signal calculation circuit 220. Here, for the calculation of Rhh (m), for example, the above (Reference 6)
The same method as that of the autocorrelation function calculation circuit 180 described in (1) can be used.

The weighting circuit 200 sets the number of samples in the frame by x
(N) is input, the prediction coefficient ai is input from the K parameter encoding circuit 160, and xw (n) obtained by weighting x (n) is output. For this calculation, for example, the same method as the weighting circuit 410 shown in FIG.

The drive signal calculation circuit 220 first determines whether the frame is voice or unvoiced using the pitch period P'd. Then, as a sound source signal representing a voice signal, in a voiced frame, a pulse train in a predetermined pitch section (representative section) is calculated. On the other hand, in the unvoiced frame, the sound source signal is calculated by the combination of the pulse train and the noise.

FIG. 2 is a pulse processing explanatory view showing a representative section in the voiced frame in FIG. 1 (a) and a method of obtaining a pulse train within the section. (A) shows a voice waveform of one frame.
First, as shown in the divided sub-frame (b), the frame is divided into sub-frames each having a pitch period P'd. In this division, the last subframe division point Ts of the immediately preceding frame (Nl-1) is stored, and this point is P '.
Subframe division can be performed while shifting by d. When the previous frame is unvoiced, another good method can be used, such as a method of dividing the subframe from the leftmost point of the frame or a method of obtaining the rising point of the signal and dividing. Next, the pulse train of the representative section (c)
As shown in, a predetermined number of pulse trains are calculated for a predetermined pitch section. Here, the sub-frame section in the center of the frame is the representative section (sub-frame section in the figure),
Other good methods can also be used. The calculation of the pulse train in the representative section is performed as follows in order to prevent the pulses from being found adjacent to each other at the subframe boundary and to find the pulses efficiently. First, a cross-correlation function Rhx (m) is calculated for (P'd + 2L) samples (L section of (b)) in which L samples are added to both sides of this subframe section. Here, L represents the maximum delay time of the autocorrelation function Rhh (m) obtained by the autocorrelation function calculation circuit 180. Rhx (m),
For the calculation of Rhh (m), the calculation method in the cross-correlation function calculation circuit 210 and the autocorrelation function calculation circuit 180 in FIG. Rhx (m) and Rhh
A pulse train is searched using (m) to obtain a predetermined number of pulse trains in the representative section. For the pulse search, the pulse search method described in the drive signal calculation circuit 220 of FIG. 1 (a) of (Reference 2) can be referred to.
The pulse train of the representative section thus obtained is shown in FIG. 2 (c). The amplitude and position of the pulse train in the representative section are output to the encoding circuit 230. At this time, a large-amplitude pulse may be cut off depending on how the representative section is divided.Therefore, a pulse train is obtained for a section obtained by extending both sides of the representative section by several samples, and is output to the encoding circuit 230. May be.

On the other hand, in the unvoiced frame, the sound source signal is obtained by combining the pulse train and noise. For this, the method described in the above (Reference 2) or the like can be referred to. In addition to this method, the frame is divided into several sub-frames for each fixed section, the pulse train is obtained only for one of these sections, and the sound source signal is represented by noise in the remaining sub-frame sections. May be. As the subframe section for obtaining the pulse train, it is possible to select a subframe section in which the power of the original speech signal or the prediction residual signal that predicts the original speech signal is large.

When the pulse train is input, the encoding circuit 230 encodes the amplitude and position of the pulse train, and outputs these codes to the multiplexer 260. Here, as the pulse train encoding method, for example, the same method as the encoding circuit 230 described in the above (Document 6) can be used.

When the pulse train and noise information is input, the pulse train is encoded using the same method as described above,
For noise, the amplitude and phase are encoded with a predetermined number of bits and the code is output to the multiplexer.

The multiplexer circuit 260 is a K-parameter encoding circuit 1
The code lki of 60, the code ld of the pitch coding circuit 150, and the code of the coding circuit 230 are input, and these are combined and output from the transmission side output terminal 270.

This is the end of the description of the transmitting side of the audio encoding method according to the present invention.

Next, the configuration of the receiving side of the audio encoding method according to the present invention will be described with reference to FIG. 1 (b).

The demultiplexer 290 separates a code representing the K parameter, a code representing the pitch period, and a code representing the excitation information among the codes input from the receiving side input terminal 280, and separates them into the K parameter decoding circuit 330, Output to the decoding circuit 320 and the decoding circuit 300.

The K parameter decoding circuit 330 decodes the K parameter and outputs the decoded value K′i to the interpolation circuit 335.

The pitch decoding circuit 320 decodes the pitch period P′d and outputs it to the drive signal restoration circuit 340 and the interpolation circuit 335.

The decoding circuit 300 decodes the sound source information and restores the drive signal
Output to.

The drive signal restoration circuit 340 uses the pitch period decoded value P′d, and if it is a value other than 0, determines that it is a voiced frame and a pulse train can be used as a sound source, and the drive signal calculation circuit on the transmission side. The frame is divided into subframes of pitch period P'd using the same method as 220. Then, a pulse train is generated using the received pulse information for the sub-frame section represented by the position of the predetermined representative section in the frame. Next, the pulse train is interpolated using the pulse train of the representative section and the pulse train of the adjacent frame to generate a sound source pulse train of one frame to restore the driving sound source signal and output it to the synthesis filter circuit 350.

For this interpolation method, the same method as the drive signal restoration circuit 340 described in (Reference 6) can be used.

On the other hand, when the combination of the pulse train and the noise is used as the sound source, the same process as the drive signal restoration circuit 340 described in (Reference 2) is performed to obtain the drive sound source signal and output it to the synthesis filter circuit 350.

The interpolation circuit 335 interpolates the decoded K parameter for each pitch cycle, and outputs the interpolated K parameter to the synthesis filter circuit 350.

The synthesis filter circuit 350 inputs the driving sound source signal and the interpolated K parameter, and outputs the synthesis filter circuit 250 on the transmission side.
Performs the same operation as that of the synthesized voice signal X for one frame.
Calculate (n) and output from the output terminal 360 on the receiving side.

The above is the explanation of the receiving side of the audio encoding method according to the present invention.

The above-described embodiment is merely one embodiment of the present invention, and various modifications thereof can be considered.

For example, in the drive signal calculation circuit 220, various voices in the unvoiced section are satisfactorily represented, and in order to realize a good transition between the unvoiced section and the voiced section, the sound source is represented by a combination of a pulse train and noise. In this case, the number of pulses may be adaptively changed from 0 (that is, only noise) to several as a combination of a pulse train and noise. In this case, it is necessary to transmit information indicating the number of pulses (for example, about 2 bits per frame).

In addition to the method described in the present embodiment, various methods can be used as the method of calculating the pulse train. For example, a method of adjusting the amplitude of a pulse obtained in the past every time one pulse is obtained can be used. For details of this method, refer to the paper by Ono et al. Entitled "Examination of Source Pulse Search Method in Multi-Pulse Driven Speech Coding Method" (Proceedings of the Acoustical Society of Japan, 157, 1983) (Reference 7). You can

Also, as another method of calculating the noise source, for example,
A method is known in which a noise signal is generated for each subframe, and noise is selected so as to minimize the error power between the signal synthesized from the noise signal and the voice signal in the subframe section. For more information on this method, see STOCHASTIC CODING OF SPEECH by BSATAL et al., “Stochastic Coding of Speech Signals at Very Low Bit Rate”.
SIGNALS AT VERY LOW BIT RATES) (PRO
C., ICC84.pp.1610-1613, 1984) (reference 8) and the like can be referred to. As another method, a method of obtaining the amplitude and phase of a noise source from a prediction residual signal obtained by predicting a voice signal is known. Since this method does not need to synthesize a voice signal, the amount of calculation can be reduced, but the quality deteriorates. For details of this method, see a paper by Oyama entitled "Linear Predictive Analysis and Synthesis Method by Driving Sound Source Modeling Residual with Noise" (Proceedings of Acoustical Society of Japan, Showa 59).
165-166) (Reference 9).

Further, when the pulse train was obtained, the K parameter value was constant within the frame (that is, the characteristics of the synthesis filter did not change within the frame), but the pulse was obtained while smoothly changing the K parameter value for each subframe. May be. Specifically, the value of the K parameter is interpolated for each subframe using the values of the K parameters of the preceding and succeeding frames, this value is converted into a prediction coefficient, and the weighting circuit 200,
It outputs to the impulse response calculation circuit 170 and calculates a cross-correlation function and an autocorrelation function using the impulse response of the representative section to obtain a pulse train. In this way, a temporally smooth spectrum change can be obtained, and higher quality speech can be synthesized.

The pulse train and the K parameter may be interpolated in synchronization with the pitch cycle using a typical pitch section as a reference, or may be a predetermined pitch section (for example, a pitch section near the center of the frame) as a reference. May be interpolated as

Further, the same process as the K parameter can be performed on the pitch period.

As the interpolation method of these parameters, methods other than linear interpolation can be considered. For example, logarithmic interpolation or the like can be considered for the pulse train and the pitch period. In addition, the interpolation of the parameters of the synthesis filter can be performed, for example, by using linear prediction coefficients (however,
In this case, it is necessary to check the stability of the filter), a logarithmic cross-sectional area function, a method of interpolating a formant parameter or an autocorrelation function, and the like can be used. These concrete methods are BS ATAL (BSATAL)
"SP-EECH ANALYSIS AND SYNTHESIS by SPEECH ANALYSIS AND SYNTHESIS BY LINEAR BRIDISION OF THE SPEECH WAVE"
BY LINEAR PREDICTION OF THE SPEECH WAVE) (J.ACOUST.SOC.AM..pp637-655,1971) (Reference 10)
Etc.) can be referred to.

Further, as a method of selecting the representative pitch section, another method can be used.

In the present embodiment, the analysis of the K parameter and the calculation of the sound source pulse train are performed with the frame length being fixed, but the frame length may be variable. In such a case, the frame length can be shortened in the voice change section and the frame length can be increased in the stationary section, so that the transmission bit rate can be reduced.

The autocorrelation function can also be calculated from the power spectrum, as is well known in the field of digital signal processing. The cross-correlation function can also be calculated from the cross power vector. For these correspondences, see AVOPPENHEIM
"Digital signal processing" by "DIGITAL SIGNAL PRO
Since it has been described in detail in Chapter 8 such as a book titled "CESSING" (reference 11), the description thereof is omitted here.

All of the above can be easily implemented without impairing the spirit of the present invention.

As described above, according to the present invention, as the sound source signal,
In the voiced section, the periodicity of the voice signal is used to search for a pulse train in a predetermined pitch section, and a one-frame sound source signal is represented using this, so the amount of calculation required for pulse search and search for the representative section is calculated. It can be significantly reduced. Moreover, since it is not necessary to send auxiliary information for indicating the position of the representative section and the start point of the subframe in the frame, information can be assigned to the pulse train accordingly. Therefore, a very high quality voice can be synthesized with a small amount of calculation even at a low bit rate, and there is an effect that hardware implementation is extremely easy.

On the other hand, in the unvoiced section, since the sound source signal is represented by the combination of the pulse train and the noise, there is an effect that various consonant sound waveforms and transient speech waveforms can be represented extremely well. Further, even if the voiced / unvoiced discrimination is mistaken, the deterioration of the sound quality is significantly reduced.

[Brief description of drawings]

FIG. 1 (a) is a block diagram showing a configuration of an embodiment of a voice encoding method and a transmitting side of the apparatus according to the present invention.
FIG. 1B is a block diagram showing the configuration of an embodiment of the receiving side of the voice encoding method and apparatus according to the present invention, and FIG.
FIG. 3A is a pulse processing explanatory view showing a representative section in a voiced frame and a method of obtaining an intra-section pulse train, and FIG. 3 is a pulse processing explanatory view showing an example of a conventional method for obtaining a pulse train of a representative section in a voiced section. is there. 110 …… Buffer memory, 130 …… Pitch analysis circuit, 140
... K parameter calculation circuit, 150 ... pitch coding circuit, 160 ... K parameter coding circuit, 170 ... impulse response calculation circuit, 180 ... autocorrelation function calculation circuit, 220 ...
… Drive signal calculation circuit, 225 …… Noise memory, 230 …… Encoding circuit, 260 …… Multiplexer, 290 …… Demultiplexer, 300 …… Decoding circuit, 310 …… Noise memory, 320 ……
Pitch decoding circuit, 330 ... K parameter decoding circuit, 335 ...
Interpolation circuit, 340 Drive signal restoration circuit, 350 Synthesis filter circuit.

Claims (2)

(57) [Claims] 1. A transmission side inputs a discrete voice signal, extracts a spectrum parameter indicating a short-time spectrum envelope and a pitch parameter indicating a pitch for each frame from the voice signal, and a frame section is cycled by the pitch parameter. Is divided into a plurality of pitch intervals equal to and a sound source signal of one pitch interval is represented by a multi-pulse train or a combination of noise and a multi-pulse train, and information representing the sound source signal, the pitch parameter and the spectrum parameter are output in combination. Then, on the receiving side, the frame is divided into pitch sections based on the pitch parameter, the excitation signal of one pitch section is restored based on the information indicating the excitation signal to restore the driving excitation signal of the entire frame, and Synthesizing the speech signal using spectral parameters Voice coding method for the butterflies. 2. A parameter calculation circuit for extracting and coding a spectrum parameter representing a short-time spectrum envelope and a pitch parameter representing a pitch for each frame from an input voice signal, and a frame section having a period equal to the pitch parameter. A drive signal calculation circuit that encodes the excitation signal by dividing the excitation signal into a plurality of pitch intervals and expressing the excitation signal of one pitch interval by a multi-pulse train or a combination of noise and multi-pulse train, and an output code of the parameter calculation circuit. And a multiplexer circuit for outputting the combination with the output code of the drive signal calculation circuit.