JPH1020895A - Speech encoding device and recording medium - Google Patents

Abstract

PROBLEM TO BE SOLVED: To reduce an increase in the processing volume when a delayed decision method is used for all time and to attain an aurally undeteriorated encoded speech. SOLUTION: A square error minimizing control means 109 outputs evaluation information for selecting from each code book 105, 106, and 107 a candidate sequence minimizing the error between a synthesized speech through an aurally weighted synthesis filter 108 and an aurally weighted input speech. On the basis of this evaluation information, a delayed decision control means 110 calculates a degree of necessary for the later delayed decision from a speech signal level from a power analyzer 104 and the voiced and unvoiced information, and controls whether or not to execute the delayed decision.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention is used for mobile communication, and divides an input speech signal into frames of a certain section, separates the vocal tract information into sound source information, and converts vocal tract information into linear prediction parameters and sound source information. The present invention relates to a speech encoding device that expresses information using several types of codebooks.

[0002]

2. Description of the Related Art Conventionally, as this kind of speech coding apparatus,
CELP encoding at a bit rate of about 4 to 8 kbps
(Code Excited Linear Prediction coding). In the CELP method, an input audio signal is divided into audio frames of a predetermined time length, and each audio frame is analyzed by a linear prediction analyzer to calculate a linear prediction coefficient. This is a method of obtaining a synthesized speech signal by exciting a synthesis filter with a sound source signal selected from a codebook. The codebook stores the past sound source signals, stores the adaptive codebook that is used by cutting out according to the pitch period of the input signal, and the sound source signals created by learning in advance. And a noise codebook for extracting and using appropriate ones, and inputting a linear sum of these codebooks as a sound source signal to an auditory weighting synthesis filter to obtain encoded speech.

Hereinafter, this type of speech coding apparatus will be described with reference to FIG. First, the adaptive codebook 3 is set in accordance with the pitch candidate calculated by the pitch analyzer 301.
An adaptive codebook candidate is selected from the past sound source signals stored in 05. A driving excitation candidate represented by a linear sum of the selected adaptive codebook candidate and the noise codebook candidate of the noise codebook 306 is generated, and the driving excitation candidate and the linearity calculated from the input signal by the linear prediction analyzer 302 are generated. Auditory weighting synthesis filter 30 from prediction coefficients
8, a synthesized voice is obtained. A codebook candidate sequence is selected by the square error minimizing means 309 so that the error between the synthesized speech and the input speech passed through the auditory weighting filter 303 is minimized. However, since an enormous amount of calculation is required to calculate the error between the perceptually weighted input speech and the perceptually weighted synthesized speech for all combinations of each codebook, in practice, it is necessary to sequentially calculate each codebook sequentially. A method of determining an optimal codebook is used. To explain using the above configuration example, first, a candidate of the adaptive codebook 305 is determined as a first step, and a candidate of the noise codebook 306 that is an optimal combination for the candidate is selected as a second step. The gain codebook 307 determines the gain that minimizes the error as a step, thereby determining a codebook candidate sequence.
The index from the codebook, the linear prediction coefficient, and the power of the input speech from the power analyzer 304 are combined by the multiplexer 310 to output the encoded speech.

[0004]

However, in such a sequential selection method, it is not guaranteed that an optimum codebook combination is obtained, so that the PSI-CELP (RCR) which is a PDC half-rate speech coding system is used. −
27D), two candidates for the adaptive codebook, which is the first stage, are left, and candidates for the noise codebook and gain codebook are selected so that the error is minimized for each candidate. The delayed decision method in which a set of candidates for an adaptive codebook, a noise codebook, and a gain codebook with a small error is selected is applied, but the error of the combination of each candidate is minimized in such a final step. Since the delayed decision method of selecting candidate series was always applied,
Usually, in an adaptive codebook that determines candidates in the first stage, there is a problem that even if only two types of candidates are left, the processing amount is significantly increased.

The present invention solves the above-mentioned conventional problem, and determines whether or not to execute delayed decision for each subframe by using an audio level of an audio signal, an evaluation value at the time of determining a code vector, and the like. Accordingly, an object of the present invention is to provide a speech coding apparatus capable of effectively reducing the amount of processing and obtaining coded speech with no audible deterioration compared to a case where a delayed decision is always performed. .

[0006]

In order to achieve the above object, the present invention provides a conventional CELP coding apparatus which uses a speech level of a speech signal and an evaluation value at the time of determining a source codebook candidate.
It is provided with a delayed decision control means for calculating whether or not the necessity of the delayed decision thereafter is to be calculated and executed. As a result, it is possible to effectively reduce the amount of processing, and to obtain coded speech that is not audibly degraded even when compared with the case of always performing delayed decision.

[0007]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The invention according to claim 1 of the present invention divides an input speech signal into frames of a certain section, separates the vocal tract information into sound source information, and converts vocal tract information into linear prediction parameters, In a speech encoding device that expresses sound source information by several kinds of codebooks, a pitch analyzer that calculates pitch candidates from an input speech signal, a linear prediction analyzer that calculates a linear prediction coefficient from the input speech signal, and an input speech signal A perceptual weighting filter that performs perceptual weighting on the input signal, a power analyzer that determines the voice level of the input voice signal, and a past sound source signal that is stored and selects an adaptive codebook candidate according to the pitch candidate from the pitch analyzer. An adaptive codebook,
A noise codebook that stores the sound source signal created by learning in advance, a gain codebook that selects the gain of the adaptive codebook candidate and the noise codebook candidate, a linear prediction coefficient from the linear prediction analyzer and A perceptual weighting synthesis filter that generates a synthesized voice from the linear sum, and a candidate sequence that minimizes the error between the synthesized voice from the perceptual weighting synthesis filter and the perceptually weighted input voice from the perceptual weighting filter is selected from each codebook Means for minimizing the square error for outputting the evaluation information for performing the calculation, selecting codebook candidates from the adaptive codebook and the noise codebook based on the output evaluation information, and selecting the codebook candidate from the evaluation information and the power analyzer. Depending on the audio signal level and sound / silence information, And a delayed decision control means for controlling whether or not to execute the calculation of the necessity of the audio signal, and to reduce the processing amount and the sound quality as compared with the case of always performing the delayed decision. Synthesized speech without degradation can be obtained.

[0008] The invention described in claim 2 is the first invention.
Is a storage medium storing a program for realizing software using the signal processor, the speech encoding device described in, for example, by storing the program in a ROM or a magnetic disk, such as a personal computer The speech encoding device of the present invention can be realized by software on a general-purpose signal processing device.

Hereinafter, an embodiment of the present invention will be described with reference to FIG.
This will be described with reference to FIG. (Embodiment) FIG. 1 shows a configuration of a speech coding apparatus according to an embodiment of the present invention. In the conventional CELP coding apparatus, a delayed decision is made based on an input speech level and an evaluation value at the time of selecting a candidate in each codebook. Is added. In FIG.
Is a pitch analyzer that calculates pitch candidates from input speech,
02 is a linear prediction analyzer that calculates a linear prediction coefficient from the input voice, 103 is an auditory weighting filter that performs auditory weighting on the input voice, 104 is a power analyzer that determines the voice level of the input voice, and 105 is a past sound source signal. , And an adaptive codebook for selecting an adaptive codebook candidate in accordance with a pitch candidate from the pitch analyzer 101, a noise codebook 106 for storing a sound source signal created by learning in advance, 107 an adaptive codebook candidate and A gain codebook for selecting a gain of the noise codebook candidate; 108, an auditory weighting / synthesis filter for generating a synthetic speech from the linear prediction coefficient from the linear prediction analyzer 102 and a linear sum of the codebook candidates; From the synthesized speech from the filter 108 and the auditory weighting filter 103 Each codebook candidate sequence error is minimized between the perceptually weighted input speech 105,10
And a square error minimizing control means 110 for outputting evaluation information for selecting from the adaptive codebook 105 and the noise codebook 1 based on the output evaluation information.
06, a codebook candidate is selected, and based on the evaluation information and the audio signal level and the sound / non-sound information from the power analyzer 104, the necessity of the delayed decision is calculated and executed. The delay decision control means 111 performs each of the codebook candidates selected from the adaptive codebook 105, the noise codebook 106, and the gain codebook 107, the linear prediction coefficient from the linear prediction analyzer 102, and the power analyzer 1.
This is a multiplexer for synthesizing the audio signal level information from the output unit 04 and outputting encoded audio.

Next, the operation of the speech coding apparatus configured as described above will be described with reference to FIG. In FIG. 1, a linear prediction coefficient is calculated by a linear prediction analyzer 102 for an input speech, and an auditory weighting synthesis filter 108 is configured using the coefficient. The output of the sound source code vector is synthesized with a synthesized speech by passing through a perceptual weighting synthesis filter 108, and the one in which the difference between the synthesized voice and the target vector of which the input voice is perceptually weighted by the perceptual weighting filter 103 is minimized, It is selected by the square error minimizing control means 109. Specifically, equation (1)
The candidate of the sound source codebook that minimizes the evaluation value shown in (1) is selected. E ² = | t−g _c · H · c _j | ² (1) where t is a target vector obtained by subjecting the input speech to auditory weighting by the auditory weighting filter 103, g _c is a gain, and c _j is a sound source. The code vector, j represents a sound source code vector index, and H represents an auditory weighting filter.
To sequentially select code vectors from a plurality of codebooks, subtract the value of the code vector selected in the previous step from the target vector t, or select the code vector of the code book to be selected in the previous step. After making the code vector orthogonal, select it.
In PSI-CELP, when selecting an adaptive codebook candidate, leaving always error E ² are two smallest ones, select for each candidate, the noise codebook, the candidates of the gain codebook, the final One candidate sequence with the smallest E ² is selected.

Next, referring to FIG.
The control procedure in the control unit 110 will be described.
For simplicity of explanation, the sound source code book
Block 105, noise codebook 106, gain codebook
And the candidates are determined in this order.
And In addition, the number of adaptive codebook 105 candidates is
The noise code book 106 and the gain code
The codebook 107 corresponds to one adaptive codebook candidate.
The sole decision shall be made. First, in step 201, voice
Determine the level. Generally when the audio level is low
Has a low sound quality, so it is important to have high sound quality.
You don't have to. From this, the threshold is set to ε _sThe sound
Voice level is ε_sIf greater, go to step 204,
In other words, two adaptive codebook candidates are left,
At step 202, the adaptive codebook
Only one complement shall be left. Select Step 202
In step 203, the noise code book candidate
In step 208, each of the gain codebook candidates is
Select one by one to determine the sound source codebook candidate sequence.
You.

Next, the case where step 204 is selected will be described. The selection evaluation value calculated when the two adaptive codebook candidates left in step 205 are selected is compared. Here, it is assumed that E ² shown in Expression (1) is a selected evaluation value, and the evaluation value with the smallest evaluation value is selected as a first candidate, a second candidate,. At this time, if the selection evaluation value of the two candidates is within a certain difference, it is not possible to determine which candidate is finally selected. Meanwhile, 2
If the selection evaluation value of the two candidates has a large difference,
Considering that the fitness of the first candidate is sufficiently high, only the first candidate of the adaptive codebook is left in step 207. In the case of step 207, noise codebook candidates and gain codebook candidates are uniquely determined as in the case of step 202. In the case of step 206, a noise codebook candidate and a gain codebook candidate are uniquely determined for each adaptive codebook candidate, and a candidate sequence with a small error is finally selected. In this way, it is controlled whether or not the necessity of the delayed decision after that is calculated based on the audio level of the audio signal and the evaluation value at the time of determining the sound source codebook candidate, and whether or not to execute the calculation.

In the above description, an example was described in which the speech level and the adaptive codebook candidate selection evaluation value were used as the delayed decision selection criterion. However, other speech parameters were used, or the speech pattern and the delayed The same effect can be expected by learning a decision execution rate or the like and using a method of controlling by a neural network.

[0014]

As described above, according to the present invention, the conventional CEL
By providing the delayed decision control means in the P encoder, there is an effect that the speech quality can be improved without significantly increasing the processing amount.

[Brief description of the drawings]

FIG. 1 is a block diagram of a speech encoding device according to an embodiment of the present invention.

FIG. 2 is a flowchart for explaining the operation of a delayed decision control means in the embodiment of the present invention.

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

[Explanation of symbols]

 Reference Signs List 101 Pitch analyzer 102 Linear prediction analyzer 103 Auditory weighting filter 104 Power analyzer 105 Adaptive codebook 106 Noise codebook 107 Gain codebook 108 Auditory weighting synthesis filter 109 Square error minimization control means 110 Delayed decision control means 111 Multiplexer

Claims (2)

[Claims] 1. A speech coding system which divides an input speech signal into frames of a certain section, separates the vocal tract information and sound source information, expresses the vocal tract information by linear prediction parameters, and expresses the sound source information by several kinds of codebooks. A pitch analyzer that calculates pitch candidates from an input audio signal, a linear prediction analyzer that calculates a linear prediction coefficient from the input audio signal, an auditory weighting filter that performs auditory weighting on the input audio signal, A power analyzer that calculates the audio level of the signal, an adaptive codebook that accumulates past sound source signals and selects an adaptive codebook candidate according to pitch candidates from the pitch analyzer, and a sound source signal created by learning in advance A gain codebook that selects the gains of the accumulated noise codebook, adaptive codebook candidates and noise codebook candidates, and linear A perceptual weighting synthesis filter that generates a synthesized voice from a linear prediction coefficient from the prediction analyzer and a linear sum of each codebook candidate; a synthesized voice from the perceptual weighted synthesis filter; and a perceptually weighted input voice from the perceptual weighting filter. Squared error minimizing control means for outputting evaluation information for selecting a candidate sequence that minimizes the error from each codebook, and a codebook from an adaptive codebook and a noise codebook based on the output evaluation information. Delayed decision control that selects candidates and calculates whether or not to execute the subsequent delayed decision based on the evaluation information and the audio signal level and sound / non-sound information from the power analyzer. And a speech encoding device. 2. A storage medium storing a program for realizing the speech encoding device according to claim 1 by software using a signal processor.