JPH0820879B2 - Speech coding system - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a speech coding method, and more particularly to a speech coding method when band-compressing a speech waveform for digital transmission or storage.

[Conventional technology]

Various schemes have been proposed in which a speech waveform is represented by a set of a pitch period, a linear prediction coefficient representing a spectrum envelope, and a prediction error, and each of them is independently coded. A typical example is the multi-pulse method that encodes prediction residuals with multiple pulses (Reference 1: Atal and Remde “Annu Model of LPC Excitation for Producing Natural Sounding.
Speech at Lowbit Rate "ICA SSP 82, pp. 614-617, 1982.; A
tal and Remde, “A new model of LPC excitation for
producing natural−sounding speech at low bit rate
s, "ICASSP82, pp.614-617 (1982).) or a method in which the prediction residual is quantized by a tree code having normal white noise as an element (Reference 2: Schrader and Atal" Speech Coding, Efficient Block Causes "). "ICA SSP 82, 1668-16
71 pages, 1982. ; Schroeder and Atal, “Speech coding
using efficient block codes, "ICASSP82, pp.1668-167
1 (1982). ).

In these encoding methods, the amount of code assigned when encoding linear prediction coefficients and prediction residuals is deterministic, in other words, determined at the design stage of the encoder.

[Problems to be solved by the invention]

In general, in a method of coding a voice by dividing it into several constituent elements, there is a considerable degree of freedom in how to distribute the code amount assigned to each element. Therefore, the ideal allocation method considers the characteristics of the speech to be encoded and the characteristics of each component. For example, the degree of the linear prediction analysis has the number of uniform acoustic tubes that approximate the vocal tract length. Therefore, it is inefficient to assign the same code amount to the prediction coefficient for male voice and female voice having different vocal tract lengths.
In the physical sense, many orders need to be assigned to male sounds. It is also inefficient to assign the same code amount to the sound source information for voices having different pitch periods. In order to obtain high-quality reproduced speech, it is desirable to allocate a large amount of code to the prediction residual per pitch, and to allocate as much code as possible to the sound source for speech with a short pitch period, It is necessary to coordinate the code amount allocation with the above components.

When a signal whose characteristics vary depending on the speaker or the utterance content such as voice is targeted, the deterministic coding allocation as in the above-described conventional method is inefficient in terms of quality.

An object of the present invention is to perform efficient speech coding by making the amount of code distributed to each constituent element dependent on the property of the target speech in a speech coding method in which speech is divided into several constituent elements and coded. To get the scheme.

[Means for solving problems]

According to the present invention, in a speech coding system that encodes speech with a first characteristic parameter representing a pitch period and a spectrum envelope and a second characteristic parameter representing a sound source, the amount of code assigned to the second characteristic parameter. And a code amount to be assigned to the first characteristic parameter are determined by adapting to the value of the pitch period.

[Action]

It is generally known that there is a positive correlation between the vocal tract length and the pitch period. Specifically, the pitch period of speech by a speaker with a long vocal tract length is long, and the pitch period of speech by a speaker with a short vocal tract length is short. We apply this general property to the problem of code amount allocation in speech coding based on linear predictive coding.

FIG. 3 is a block diagram showing the principle of the speech coding system of the present invention. Since the vocal tract length is short for speech with a short pitch period, the code amount assigned to the linear prediction coefficient may be small, and it is better to use it for coding the prediction residual. On the other hand, for a voice with a long pitch period, since the vocal tract length is long, the degree assigned to linear prediction cannot be reduced. Instead, when viewed per pitch period, the amount of code assigned to the prediction residual is not so much compared with speech with a short pitch period. In this way, the characteristics can be expected to be improved by controlling the code amount to be assigned to the linear prediction coefficient and the prediction residual according to the value of the pitch period. These can be schematically represented as in FIG. FIG. 4 is a diagram for helping the understanding of the principle of the present invention. The principle of the present invention is to encode the linear prediction coefficient and the prediction residual with the code amount that gives the best characteristic for each pitch period. In addition, the spectrum envelope obtained by linear prediction tends to be pulled by the pitch period. In other words, even if the prediction order is increased for speech with a short pitch period, most of the orders are wasteful because they are used to represent harmonic components of the pitch period. A more efficient coding method can be obtained by allocating the code amount for the extra order to the sound source. This is realized by preparing a graph as shown in FIG. 4 in advance in the table and determining the distribution of the code amount according to the pitch period.

In FIG. 3, 10 is an input terminal for inputting a voice waveform,
Reference numeral 11 is a pitch analysis for extracting a pitch cycle from the input speech waveform, and 12 is a code amount allocation table in which the graph of FIG. 4 is summarized in a table. The distribution of codes for the residuals is recorded. 13 is a linear prediction analysis that obtains a linear prediction coefficient from the input speech with the code amount determined by the code amount allocation table 12, and 14 is a prediction residual that encodes the prediction residual with the code amount specified in the code amount allocation table 12. This is difference encoding. Reference numeral 15 is an output terminal for outputting a code representing the encoded prediction residual.

〔Example〕

Next, an embodiment of the present invention will be described with reference to FIGS.

FIG. 1 is a block diagram showing an embodiment of a speech coding system of the present invention. A voice waveform divided at regular intervals is input to a voice input terminal 100 and supplied to a pitch analysis 110, a linear prediction analysis 130, and a buffer memory 135. Pitch analysis 110
Calculates the pitch period and pitch prediction coefficient of the input speech, and outputs them to the code amount allocation table 120, the buffer memory 135 and the multiplexer 170. Some methods for extracting the pitch period and the pitch prediction coefficient have already been studied. For example, Reference 3: Lavinar et al., “A compatitive performance study of secular pitch detection algorithm”, IEETransaction, AESP24, p.399 (1976),
(Rabiner, et al, “A comparative performance study
of several pitch detection algorithm, "IEEE Trans.,
ASSP-24, p.399 (1976)).

The code amount allocation table 120 is, as described in the section of the above operation, the order of performing the linear prediction analysis according to the value of the pitch period, the number of quantization bits when quantizing each prediction coefficient, and the prediction residual. The number of bits to do is specified. Information about the linear prediction is output to the linear prediction analysis 130, and information about the prediction residual quantization is output to the multi-pulse speech encoder 140 described as MPC and the inverse multi-pulse speech encoder (referred to as inverse MPC) 160. . Linear prediction analysis 130 is an input terminal
The speech waveform input from 100 is subjected to linear prediction analysis, and the linear prediction coefficient is determined by MPC140, inverse MPC160, and multiplexer 170.
Output to. The linear prediction analysis here has a configuration in which a PARCOR coefficient, which is one of linear prediction parameters, is obtained and quantized. The analysis and quantization are performed according to the prediction order and the number of quantization bits defined in the code amount allocation table 120. The quantization method of PARCOR coefficient is already well researched,
For example, Reference 4: Kitawaki, Itakura, Saito “Optimal Code Configuration in PARCOR-type Speech Analysis and Synthesis System” IEICE Transactions J61-A
2, pp.119-126 (Showa 53). Buffer memory 135
Supplies the voice input from the input terminal 100 to the subtractor 145 after dividing the voice for each pitch period. The subtractor 145 calculates the difference between the input voice for one pitch coming from the buffer memory 135 and the reproduced voice one pitch before coming from the pitch prediction 150, and calculates it as MP.
Supply to C140. The MPC140 multi-pulse encodes the pitch prediction error calculated by the subtractor 145.The linear prediction coefficient supplied from the linear prediction analysis 130 and the number of pulse excitation sources and the quantized bits specified by the code amount allocation table 120 are used. The sound source pulse is calculated based on the number.

Since many methods of obtaining a sound source pulse in the multi-pulse method are already known, they are omitted here. For example, reference 5: Ozawa, Araseki, Ono “Study on multi-pulse driven speech coding method”, IEICE Communication System Study Group, CS
Details on 82-161 (57). The number of sound source pulses specified by the linear prediction analysis 130 is determined by the method of Literature 5, and the positions and amplitudes of the determined pulses are quantized with the code amount also specified by the linear prediction analysis 130. Here, the obtained sound source pulse parameter is output to the inverse MPC 160 and the multiplexer 170. The inverse MPC 160 reproduces the parameters of the sound source pulse input from the pitch prediction 150 as a sound source pulse train according to the regulation of the code amount allocation table 120, and reproduces pitch prediction from the reproduced sound source pulse train and the linear prediction coefficient input from the linear prediction analysis 130. The error is calculated. The reproduction pitch prediction error is output to the adder 147. The adder 147 outputs the reproduced sound one pitch before output from the pitch prediction 150 and the inverse MPC160.
The playback pitch prediction error signal output from is added to calculate the playback voice for the current pitch, and pitch prediction 15
Supply to 0. The pitch predictor 150 is a memory for storing one pitch of the reproduced voice, and the subtractor 145 and the adder 147 multiply the reproduced voice for the pitch period analyzed by the pitch analysis 110 by the pitch prediction coefficient analyzed by the pitch analysis 110. And output to.
The multiplexer 170 communicates the code from the output terminal 180 by combining the parameter representing the sound source pulse output from the MPC 140, the linear prediction coefficient analyzed by the linear prediction analysis 130, the pitch prediction coefficient extracted by the pitch analysis 110, and the pitch period. Output to the road.

FIG. 2 shows the segmental SNR measured by changing the code amount assigned to the linear prediction coefficient and the excitation pulse for each pitch period in the configuration of this embodiment. In FIG.
The code amount to be allocated to the linear prediction coefficient and the excitation pulse for the pitch period of 3.8 msec and 7.5 msec is 2.3, 2.0 and 1.7 k, respectively.
The vertical axis shows the segmental SNR obtained by varying b / s and 4.5, 4.95, 5.4 kb / s. That is, if the assigned code amount is changed for each pitch period, the encoding characteristic can be improved, and the effect of this embodiment can be confirmed.

〔The invention's effect〕

As described above, the present invention assigns a voice to a second feature parameter in a voice encoding system that encodes a voice with a first feature parameter representing a pitch period and a spectrum envelope and a second feature parameter representing a sound source. By determining the code amount and the code amount assigned to the first feature parameter by adapting to the value of the pitch period, each feature parameter is adapted to the value of the pitch period when encoding the spectrum envelope and the prediction residual. The amount of code assigned to is changed. Using the well-known property that there is a positive correlation between the pitch period and vocal tract length due to the characteristics of the speech generation process, the code amount is assigned according to the characteristics of the speech to be encoded. For example, since more efficient code allocation can be realized as compared with the conventional deterministic code amount allocation, the present invention has the effect of improving the encoding characteristics.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of a voice coding system of the present invention, FIG. 2 is a diagram for explaining the effect of the present invention, and FIG.
FIG. 1 is a block diagram showing the principle of the speech coding system of the present invention,
FIG. 4 is a diagram for helping understanding of the principle of the present invention. 10,100 …… Input terminal, 11,110 …… Pitch analysis, 12,120…
… Code amount allocation table, 13,130… Linear prediction analysis, 14…
… Predictive residual coding, 15, 180… Output terminal, 135… Buffer memory, 140… MPC, 145… Subtractor, 147… Adder, 150… Pitch prediction, 160… Inverse MPC, 170… … Multiplexer.

Claims (1)

[Claims] 1. A speech coding system for coding speech with a first characteristic parameter representing a pitch period and a spectrum envelope and a second characteristic parameter representing a sound source, and a code amount assigned to the second characteristic parameter. A speech coding method, wherein the code amount to be assigned to the first characteristic parameter is determined by adapting to the value of the pitch period.