JPH05108098A - Speech encoding device - Google Patents

Abstract

PURPOSE:To obtain a speech of high quality even when the bit rate is low. CONSTITUTION:For the generation of a driving sound source which generates a synthesized speech having minimum distortion for a weighted input speech obtained by an auditory weighting filter 11, a sound source switch 15 selects a sound source having small distortion between a pulse sound source outputted by a pulse sound source generator 13 and a noise sound source outputted by a probable code book 14 and outputs it from an adaptive code book 12, and a driving sound source generator 16 generates the driving sound source. The pulse component and noise component of the adaptive code book 12 are separated and stored in an adaptive code book pulse component storage unit and an adaptive code book noise component storage unit and a pulse and noise component gain controller may control the gains of the pulse component and noise component to optimum gains.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice coding device used for digital communication, voice mail and the like.

[0002]

2. Description of the Related Art In recent years, a speech coding apparatus at a low bit rate of about 4.8 to 8.0 kb / s has a CELP (Code Excited Linear Prediction Coder) as shown in FIG.
) Is widely used.

A conventional speech coding apparatus of this type will be described below. In FIG. 3, reference numeral 31 is a perceptual weighting filter that performs perceptual weighting of the input voice, and outputs a weighted input voice. 32 is an adaptive codebook,
Stores past driving sound sources. Reference numeral 33 is a stochastic codebook, which holds a plurality of noise sound sources in advance. A driving sound source generator 34 generates a driving sound source signal from the adaptive codebook 32 and the stochastic codebook 33. Reference numeral 35 is a weighted synthesis filter that generates a weighted synthesized speech with the driving sound source as an input. Reference numeral 36 is a distortion minimizer that calculates the distortion of the weighted synthesized speech with respect to the weighted input speech, and outputs the long-term prediction delay and gain that minimize this distortion, and the representative vector and gain in the stochastic codebook. ..

The operation of the speech coding apparatus configured as described above will be described below. First, the perceptual weighting filter 31 obtains the weighted input voice v [n]. The CELP encoding device encodes the driving sound source e [n] that generates the weighted synthesized speech closest to the CELP encoding device. Here, in the driving sound source, e [n] is a long-term predicted signal a which is an output of the adaptive codebook 32 as shown in the following equation.
[N−L] and the vector c _I in the probabilistic codebook 33.
It consists of [n].

E [n] = β · a [n−L] + γ · c _I [n] ... (1)

In practice, it is difficult to determine both components at the same time. Usually, only the long-term predicted component is first determined from the adaptive codebook 32 by the distortion minimizer 36 (adaptive codebook search), and past driving is performed. It outputs a long-term predicted delay L indicating how far back the signal is going and an optimum gain β. Next, the distortion is minimized by the stochastic codebook 33 so as to minimize the remaining distortion (stochastic codebook search), and the selected representative codebook number I and the optimum gain γ are output.

[0007]

However, in the above-mentioned conventional speech coding apparatus, at a low bit rate of about 4.8 kb / s or less, as the driving sound source signal generation performance deteriorates, the past driving sound source becomes worse. Performance degradation also occurs in the generation of long-term prediction components from the adaptive codebook that stores signals,
In particular, there is a problem that a voiced sound having a pulsed pitch period sound source causes a remarkable deterioration in voice quality.

The present invention solves the above-mentioned conventional problems, and provides an excellent speech coding apparatus which suppresses deterioration of speech quality of voiced sound having a pulse-like pitch period sound source even at a low bit rate. The purpose is to do.

Another object of the present invention is to provide a speech coder having high speech quality by controlling the gains of pulses and noise components of a long-term predicted signal.

[0010]

In order to achieve the above object, the present invention provides, in addition to a stochastic codebook, a pulse sound source generator for generating a pulse sound source, and a pulse generated by the pulse sound source generator. A sound source switch that selects either a sound source or a noise source generated from a stochastic codebook is provided, and it is possible to suppress deterioration of voice quality in voiced sound having a pulse-like pitch period sound source even at a low bit rate. It was done.

In order to achieve another object of the present invention,
Instead of the adaptive codebook, an adaptive codebook pulse component store that generates a pulse component of the long-term prediction signal, an adaptive codebook noise component store that generates a noise component,
It is equipped with a pulse / noise component gain controller that controls the gain of the pulse component and noise component of the long-term prediction, and realizes high voice quality by controlling the gain of the pulse and noise component of the long-term prediction signal. is there.

[0012]

Therefore, according to the present invention, by selecting the optimum sound source of the pulse sound source which is the output of the pulse sound source generator and the noise sound source generated from the stochastic codebook by the sound source switching unit, low bit Even at the rate, it is possible to generate a pulse component with a voiced sound having a pulse-like pitch period sound source, and it is possible to perform encoding while suppressing deterioration of voice quality.

Further, according to the present invention, the pulse / noise component gain controller provided in the long-term predictor optimally controls the gains of the pulse and noise components of the long-term predictive signal, so that the voice having high voice quality is obtained. It has an effect that encoding can be performed.

[0014]

DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 shows the configuration of the first embodiment of the present invention. In FIG. 1, 11 is a perceptual weighting filter for perceptually weighting input speech, 12 is an adaptive codebook for storing past driving sound sources, 13 is a pulse sound source generator for generating pulse sound sources, and 14 is a plurality of noise sound sources. Probabilistic codebook, 15 is pulse generator 1
3 or a sound source switcher for selecting one of the sound sources generated from the probabilistic codebook 14, 16 is the sound source selected from either the pulse sound source generator 13 or the stochastic codebook 14, and the output of the adaptive codebook 12. And a driving sound source generator that generates a driving sound source from, and 17 is a weighted synthesis filter that generates a weighted synthetic speech from the driving sound source,
Reference numeral 18 denotes a distortion minimizer that calculates distortion between the weighted input speech and the weighted synthesized speech, and outputs a long-term prediction delay and a pulse position or a stochastic codebook code that minimize the distortion and their gains. is there.

Next, the operation of the first embodiment will be described. First, the perceptual weighting filter 11 obtains the weighted input voice v [n], and thereafter, the driving sound source e [n] that produces the closest weighted synthesized voice is encoded. Here, the driving sound source e [n] generated by the driving sound source generator 16
Is the long-term prediction signal a [n of the output of the adaptive codebook 12
-L], the pulse component p _M [n] generated by the pulse source generator 13 or the noise component c _I [n] from the stochastic codebook 14, and is represented by the following equation. e [n] = β · a [n−L] + γ _P · p _M [n] ··· (2) or e [n] = β · a [n−L] + γ _N · c _I [n ] (3) where p _M [n] is a pulse train having a single impulse from the position M at pitch period intervals, β, γ _P , and γ _N are long-term prediction, pulse, and gain of white noise component, respectively. Is.

The above equations (2) and (3) are both 2
It is difficult to determine the two components at the same time. First, the long-term prediction signal of the adaptive codebook 12 is determined by distortion minimization of the perceptually weighted synthetic speech, and a long-term indication indicating which part of the past drive signal is used. Prediction delay L and optimal gain β
Is output. Next, for the distortion after determining the long-term predicted signal,
Corresponding to equations (2) and (3), distortion minimization is performed by a pulse sound source or a noise sound source, and the sound source with the smaller distortion is selected as the sound source in the coding section. For example, the pulse position M and the optimum gain γ _P , and if it is a noise source, the selected noise vector number I and the optimum gain γ _N are output.

As described above, according to the first embodiment, with respect to the distortion after the determination of the long-term predicted signal, either the pulse sound source or the noise sound source having a small distortion is selected,
It is possible to generate a pulse component with a voiced sound having a pulse-like pitch period sound source, and it is possible to perform encoding while suppressing deterioration of voice quality even at a low bit rate.

FIG. 2 shows the configuration of the second embodiment of the present invention. In FIG. 2, reference numeral 21 denotes an adaptive codebook pulse component storage unit having an adaptive codebook pulse component,
Reference numeral 22 is an adaptive codebook noise component storage having a noise component of the adaptive codebook, 23 is a pulse / noise component gain controller for controlling the gain of the pulse and noise components of the long-term prediction signal, and the others are the first embodiment. It has the same configuration as. That is, 24 is a pulse source generator that generates a pulse source, 25 is a stochastic codebook that holds a plurality of noise sources, and 26 is either a pulse source generator 24 or a source generated from the stochastic codebook 25. A sound source switching unit for selecting whether or not, 27 is a drive sound source generation for generating a drive sound source from the sound source selected from either the pulse sound source generator 24 or the stochastic codebook 25 and the output of the pulse / noise component gain controller 23. , 28 is a weighted synthesis filter that generates a weighted synthesized speech from the driving sound source, 29 is a distortion between the weighted input speech and the weighted synthesized speech,
A distortion minimizer that outputs a long-term prediction delay and a pulse position or a probabilistic codebook code that minimizes the distortion and their gains, and 30 is a perceptual weighting filter that performs perceptual weighting of input speech.

Next, the operation of the second embodiment will be described.
To do. The closest to the weighted input speech as in the first embodiment.
Code the driving sound source e [n] that generates a weighted synthetic speech
First, the distortion of the long-term predicted signal is minimized. So
At this time, in the second embodiment, the long-term predicted signal a_S[N]
Ruth component a_P[N-L] and noise component a_NSum of [n-L]
And the long-term prediction delay L and
Component gain β_P, Β _NTo decide. a_S[N] = β_P・ A_P[N-L] + β_N・ A_N[N-L] (4)

As an example of the method of determining the long-term prediction delay L and the gains β _P and β _N of each component, first, β _P = β _N
As a long-term prediction delay L, the optimum pulse component gain β _P and noise component gain that minimize the weighted squared error E with respect to the input voice expressed by the following equation (5) are obtained. Determine β _N. E = Σ {p [n] -β P · b P [n] -β N · b N [n]} 2 → Min ··· (5) p [n]: the perceptually weighted input speech b _P [n ]: Adaptive codebook pulse component output auditory weighted synthesized speech β _P : Pulse component gain b _N [n]: Adaptive codebook noise component output weighted synthesized speech β _N : Noise component gain

By solving for β _P and β _N so that E becomes minimum in the above equation, the optimum β _P and β _N
Is calculated as follows.

[0022]

[Equation 1]

After the distortion of the long-term prediction component of the adaptive codebook is minimized, the pulse source or the noise source, which has a small distortion, is the same as the first embodiment. Select. Also, the adaptive codebook is updated separately for each of the pulse component and the noise component. If the selected sound source is a pulse sound source, it is used for updating the pulse component of the adaptive codebook. Used to update the ingredients.

As described above, according to the second embodiment, the gain of the pulse component and the noise component of the long-term prediction signal is optimally controlled by storing the adaptive codebook by separating it into the pulse component and the noise component. This has the effect of enabling speech coding with high speech quality.

[0025]

As is apparent from the above embodiment, the present invention selects the pulse component or the white noise component, which has a small distortion, with respect to the distortion after the determination of the long-term prediction component. It is possible to generate a pulse component with a voiced sound having a pitch period sound source, and it is possible to perform encoding while suppressing deterioration of voice quality even at a low bit rate.

Further, according to the present invention, the gain of the pulse component and the noise component of the long-term prediction signal can be optimally controlled by storing the adaptive codebook by separating it into the pulse component and the noise component. This has the effect of enabling speech coding with higher speech quality.

[Brief description of drawings]

FIG. 1 is a schematic block diagram of a speech encoding apparatus according to a first embodiment of the present invention.

FIG. 2 is a schematic block diagram of a speech coder according to a second embodiment of the present invention.

FIG. 3 is a schematic block diagram of a conventional CELP speech coding apparatus.

[Explanation of symbols]

 11 Auditory Weighting Filter 12 Adaptive Codebook 13 Pulse Sound Source Generator 14 Stochastic Codebook 15 Sound Source Switcher 16 Driving Sound Source Generator 17 Weighted Synthesis Filter 18 Distortion Minimizer 21 Adaptive Codebook Pulse Component Store 22 Adaptive Codebook Noise Component storage 23 Pulse / noise component gain controller 24 Pulse sound source generator 25 Stochastic codebook 26 Sound source switcher 27 Drive sound source generator 28 Weighted synthesis filter 29 Distortion minimizer 30 Auditory weighting filter

Claims (3)

[Claims] 1. A perceptual weighting filter for performing perceptual weighting of input speech in a certain section to generate a weighted input speech, an adaptive codebook for storing past driving sound sources, and a pulse sound source generator for generating pulse sound sources, A stochastic codebook that stores a plurality of noise sources in advance, a sound source switcher that selects one of the pulse sound source generator and the sound source generated from the stochastic codebook, and the pulse sound source generator or the stochastic codebook A driving sound source generator that generates a driving sound source from a sound source selected from any of the following and an output of the adaptive codebook, a weighted synthesis filter that synthesizes speech with the driving sound source generated from the driving sound source generator as an input, Calculate the sum of squares of the error of the weighted synthetic speech with respect to the weighted input speech, and minimize the long-term prediction delay, pulse A speech coder having a position or probabilistic codebook code and a distortion minimizer for selecting respective gains and outputting the code. 2. An adaptive codebook pulse component store for storing pulse components of a long-term predicted signal, an adaptive codebook noise generation store for storing noise components, and a pulse component for long-term prediction instead of the adaptive codebook. The speech coding apparatus according to claim 1, further comprising a pulse / noise generation gain controller for controlling a noise generation gain, wherein the gains of the pulse component and the noise component of the long-term prediction signal can be controlled independently. 3. The pulse / noise component gain controller determines the gains of the pulse component and the noise component of the long-term predicted signal so as to minimize the perceptual weighted squared error with respect to the input speech. 2. The audio encoding device according to 2.