JPH05232996A - Voice coding device - Google Patents

Abstract

PURPOSE:To provide a voice coding device capable of effectively reproducing a voice independently of sexes at the time of decoding it. CONSTITUTION:Plural probability code books 18, 20 having respectively different contents such as male voices and female voices are prepared and pitch period ranges corresponding to respective books 18, 20 are determined. A pitch period and an LPC coefficient are extracted from an input voice 10 and a probability code book corresponding to the pitch period is selected by an evaluation device 34. Then the contents of the selected code book is used as a residual signal and passed through an LPC synthesizing filter 16 consisting of the found LPC coefficient to form a synthetic voice and an index is selected so that a difference between the synthetic voice and the input voice 10 is minimized. The selected code book is held for a fixed frame period (e.g. about 20 to 30ms) or until a different pitch period appears.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech coder for compressing and coding speech.

[0002]

2. Description of the Related Art Conventionally, in speech coding, a linear prediction analysis of speech is performed to obtain a residual signal, and the residual signal is replaced by a representative vector in a codebook to represent a representative in the codebook pointed to by an index. CELP (Code Excited) is used to compress the voice by reproducing the vector into voice and expressing it with the code with the least distortion compared to the original voice.
Linear Prediction) is well known. This CEL
Regarding P, for example, MR Schroeder, BSAtal, "CO
DE-EXCITED LINEAR PREDICTION (CELP): HIGH-QUALITY S
PEECH AT VERY LOW BIT RATES ", ICASSP '85, pp937-94
0, 1985. A probability codebook is used as the codebook used in this CELP, and white noise is normally used as its content. Probabilistic codebooks are often used with adaptive codebooks. As a typical example of conventional CELP, FIG. 3 shows an example of a speech coder that uses a probability codebook together with an adaptive codebook.

The input voice 100 is sampled at a frame length of 8 kHz, for example. Now, assuming that the frame length is 20 ms, for example, 160 samples are sent to the linear prediction analyzer 102 and the difference unit 104. Linear prediction analyzer 1
02 designates the coefficient of the synthesis filter 106 as LPC (Linear
Predictive Code) given as a set of parameters or a set of reflection coefficients.

On the other hand, the probabilistic codebook 108 and the adaptive codebook 110 store 40 samples of data for each sub-frame every 5 ms obtained by dividing the frame length into four, for example. The probability codebook 108 is connected to the probability codebook gain unit 112. The adaptive codebook 110 is also connected to the adaptive codebook gain unit 114. Then, the sum of the contents obtained by multiplying the output of each codebook by the gain is obtained by the adder 116 and input to the synthesis filter 106.

Further, the difference unit 104 outputs the original voice 100
And the voice reproduced from the codebook through the synthesis filter 106, and the sum of squares of the errors weighted by the weighting filter 118 is evaluated by the evaluator 120, and the value of the codebook 108 that minimizes the value is evaluated. The codebook is searched for the index set of 110 and the gains of the gain units 112 and 114 are adjusted.

Then, when the codebook index is determined, the contents pointed to by the index of the probability codebook 108 and the contents pointed by the index of the adaptive codebook 110 are multiplied by respective gains and summed. The obtained data is input to the buffer 122, and the contents of the adaptive codebook 110 are rewritten for each subframe based on the contents.

This work is based on the assumption that the difference between a male voice and a female voice is not noticeable at the stage of the residual signal, so the probability codebook is generally CEL.
In P, only one type was used.

[0008]

However, it is beginning to be said that, in addition to the pitch period (fundamental frequency of the voice), the voice quality of the male voice differs from that of the female voice based on the residual signal.

For example, Ito, Kohata, Takagi, "PARCO
A Study on Compressed Transmission Method of Residual Signal in R Analysis and Synthesis Method ", IEICE Technical Report SP90-51, PARCOR
Although synthesis is used, there is an announcement that a relatively good reproduction sound can be obtained for male voices and a noise feeling is attached to female voices due to the difference in energy contribution rate to the representative residual. ..

On the contrary, Seza, Tasaki, Nakajima, "Examination of sound source model in vocoder", Proceedings of Acoustical Society of Japan 1-
In 6-16, the noise is removed from the female voice and the male voice by noise estimation. When the noise is removed, the female voice contains a lot of vocal cord noise. Therefore, the female voice is more noisy. There are also announcements that the effect of removal is great and the subjective evaluation is improved.

Furthermore, in Kataoka, Moriya, "Low Delay Speech Coding Using Adaptive Processing of Pitch Information", Proceedings 2-6-17 of the Acoustical Society of Japan, by adaptively processing pitch gain, Although it tries to improve the voice quality, there is an announcement that this method also makes the voice of a man worse than the voice of a woman.
Therefore, it has been found that the residual signal is slightly different between the male voice and the female voice in addition to the pitch period in reality.

As described above, due to the difference between the pitch period and the residual signal, it was not possible to satisfactorily reproduce both the male voice and the female voice when the encoded voice was decoded. .. The present invention has been made in view of the above points, and an object of the present invention is to provide a voice encoding device capable of excellently reproducing voices for both male and female voices.

[0013]

In order to achieve the above object, a speech coding apparatus according to the present invention calculates a difference between a prediction value signal from a prediction filter for predictively analyzing an input speech signal and the input speech signal. A speech coding apparatus for obtaining a residual by taking the residual and quantizing the residual with a codebook, wherein a plurality of codebooks having different contents and a fundamental frequency for detecting a fundamental frequency of the speech from the input speech signal It is characterized in that it is provided with detecting means and codebook switching means for switching the plurality of codebooks in accordance with the fundamental frequency of the voice detected by the fundamental frequency detecting means.

[0014]

That is, in the speech coding apparatus according to the present invention, a plurality of codebooks having different contents, for example, probability codebooks are prepared and the range of the fundamental frequency of speech corresponding to each is determined. Then, the fundamental frequency detecting means and the linear prediction analyzing means respectively extract the pitch period and the LPC coefficient from the input speech, and the codebook switching means first selects the probability codebook according to the pitch period. After that, as in a normal speech coding apparatus, the contents of the selected probability codebook are passed through an LPC synthesis filter composed of LPC coefficients obtained from the input speech as a residual signal to create synthesized speech, Select the index that minimizes the difference between the input voices. Then, the selected probability codebook has a certain period (for example, 20 ms to
By holding the frame period of about 30 ms or until the pitch period becomes completely different), the male voice and the female voice can be well reproduced without increasing the bit rate more than necessary.

[0015]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram of the first embodiment of the present invention. The input speech 10 is input to the linear prediction analyzer 12 and the differencer 14. The linear prediction analyzer 12 provides the coefficients of the synthesis filter 16 as a set of LPC parameters or a set of reflection coefficients.

A representative vector for coding data is stored in the codebook, and the codebook includes, for example, two probability codebooks 18 and 20, a probability codebook gain unit 22, and an adaptive codebook 24. And an adaptive codebook gain unit 26.

The adder 28 adds the output of the stochastic codebook gain unit 22 and the output of the adaptive codebook gain unit 26 and outputs the result to the synthesis filter 16 and the buffer 30. The weighting filter 32 weights the output of the differentiator 14. The evaluator 34 evaluates the output of the weighting filter 32 and selects the probability codebook 18 for male voice or the probability codebook 20 for female voice according to the result.

The operation of such a configuration will be described below. The input voice 10 is sampled at a frame length of 8 kHz, for example. Now, for example, the frame length is 20
In ms, 160 samples are sent to linear prediction analyzer 12 and differencer 14. The linear prediction analyzer 12 gives the coefficients of the synthesis filter 16 as a set of LPC parameters or a set of reflection coefficients.

On the other hand, in the codebook, the frame length is
For example, data 4 for each subframe of 5 ms divided into 4
A representative vector for coding 0 samples is stored. The codebook is composed of, for example, two stochastic codebooks 18 and 20, a stochastic codebook gain unit 22, an adaptive codebook 24, and an adaptive codebook gain unit 26. Probabilistic codebooks are created, for example, by linearly predicting multiple male and female voices to obtain residual signals, so the codebook obtained from male voices is obtained from residual signals peculiar to male voices and female voices. The codebook has the characteristic of the residual signal peculiar to the female voice. Then, suppose that the probability codebook 10 is for a male voice and the probability codebook 11 is for a female voice. Assuming that the boundary between the pitch frequencies of the male voice and the female voice is 250 Hz, the codebooks for the male voice and the female voice are selected.

Next, the procedure for selecting the probability codebook will be described. The probability codebook is selected for each frame, and pitch information obtained by the adaptive code search is used for this purpose.

That is, in order to create the adaptive codebook 24, a buffer 30 of about 150 samples is usually prepared. The buffer 30 stores the residuals of the synthetic signal for 150 samples, which is traced back from one subframe. The buffer 30 is composed of, for example, a shift register and is updated by discarding old 40 samples and inputting new 40 samples every 5 ms period. Each data of the adaptive codebook 24 is stored in the buffer 3
The sample in 0 is created by extracting and repeating the assumed pitch period, for example, 20 to 147 sampling periods.

Subsequently, each representative vector in the adaptive codebook 24 is passed through the synthesis filter 16 via the adaptive codebook gain unit 26 and the adder 28 to obtain the voice reproduced from the codebook. Next, in the differentiator 14, the voice reproduced from this codebook and the original voice (input voice 1
0) and the square of the error weighted by the weighting filter 32 is evaluated by the evaluator 34, and the code having the minimum value is obtained. Then, the pitch frequency corresponding to the index of the obtained probability codebook is determined, and if it is 250 kHz or less, the probability codebook 18 for male voice is selected, and if it exceeds, the probability codebook 20 for female voice is selected.

Subsequently, the selected probability codebook 18
Or search for 20. That is, each codeword in the probability codebook is sequentially passed through the synthesis filter 16 via the probability codebook gain unit 22 and the adder 28, and then the adaptive codebook gain unit 26 is applied to the codeword selected by the adaptive codebook 24. The optimum gain is given by
And the difference between the synthesized speech obtained through the synthesis filter 16 and the original speech is obtained by the differentiator 14. This value is weighted by the weighting filter 32, the squared error is evaluated by the evaluator 34, and the code that minimizes the evaluated value is obtained.

The value obtained by giving the optimum gain by the stochastic codebook gain unit 22 to the code vector selected from the stochastic codebook in this way is added by the adder 28 to the adaptive codebook sound source previously selected, and the synthesis filter 1
By passing through 6, encoded voice is obtained. With the LPC coefficient or the reflection coefficient obtained by the linear prediction analyzer 12,
Adaptive codebook 24 and probability codebook 18 or 20
The code vector index and the optimum gain value selected from are recorded or transmitted as compressed data by a recording unit or a transmission unit (not shown).

In the above method, when decoding compressed data, the pitch period is known by referring to the adaptive codebook 24, and it is not necessary to encode the information indicating which probability codebook of male voice or female voice is selected. ..

Next, a second embodiment of the present invention will be described with reference to the block diagram of FIG. The second embodiment is characterized in that a pitch detector 36 is provided. The other structure is the same as that of the first embodiment, and therefore, the same portions are denoted by the same reference numerals as those in FIG. 1 and their description is omitted. The input voice 10 is sampled for each frame, and the synthesis filter coefficient is obtained by the linear prediction analyzer 12.

The adaptive code book 24 is created by changing the pitch period according to each of the male voice and the female voice. That is,
In the case of a male voice, it is created in a period of 36 to 147 samples. On the other hand, in the case of a female voice, it is created every half sample period in the range of 20 to 40 samples.

In this way, by creating the adaptive codebook 24 in accordance with the characteristics of the pitch period of the male voice and the female voice,
The codebook size can be reduced. For this reason,
The calculation of the codebook search becomes easy. Further, in the case of a female voice in which a minute difference in pitch period affects the sound quality, it is possible to improve the sound quality by making the accuracy finer than the sampling cycle.

At the beginning of one frame, the pitch detector 36 finds the autocorrelation function of 150 samples to find the pitch period. If the pitch period is, for example, in the male voice range, the adaptive codebook 24 is created corresponding to the voice,
Also, the adaptive codebook 24 for male voice is selected. At the same time, a flag indicating which probability codebook is used is recorded at the beginning of the frame. For example, when the probability codebook 18 for male voice is selected, the flag is set, and when the probability codebook 20 for female voice is selected, the flag is not set. This codebook identification information is 2
Only 1 bit of information is required every 0 ms, and the total is 5
Only an increase of 0 bps. It should be noted that a method used for other pitch extraction may be used for the pitch extraction.

The sum of the contents of the index of the codebook determined in the above processing is passed through the synthesis filter 16, and the difference unit 1
In step 4, the difference between the original voice and the voice reproduced from the codebook is calculated, the square of the error weighted by the weighting filter 32 is evaluated by the evaluator 34, and the codebook is obtained so as to obtain the index set having the minimum value. To explore. When the index of the codebook is determined in this way, the contents indicated by the index of the probability codebook indicated by the flag and the contents indicated by the index of the adaptive codebook 24 are respectively supplied to the gain units 22 and 26. Then, the sum obtained by multiplying the gains of the above by the adder 28 is input to the buffer 30, and the adaptive codebook 24 is created for each subframe based on the data.

Hereinafter, compressed data can be obtained by encoding in the same procedure as in the first embodiment described above. FIG. 2B shows the data coded by the second embodiment. Reference numeral 38 in the figure is a flag indicating which of the above probability codebooks has been used.

In the second embodiment, it is necessary to record in the compressed data the data (flag 38) indicating whether the probability code data of male voice or female voice is used. However, the adaptive codebook and the probability codebook are used. Any of these can be coded more efficiently by adaptively switching.

Next, a third embodiment of the present invention will be described.
In the third embodiment, not only is the probability codebook switched between male and female voices, but more types are provided, and the probability codebook is selectively switched according to the pitch cycle. By this means, it is possible to reflect a finer residual characteristic that is adapted to the pitch period and improve the coding efficiency. As a result, the size of each probability codebook can be reduced, and the search becomes easy.

For example, if eight kinds of probability codebooks are prepared, 3 bits are required for each frame as information for selecting the probability codebook. However, it is possible to reduce this information by utilizing the characteristic that the pitch period does not change frequently. That is, a variable-length code such as a Huffman code is used to identify the probability codebook, a short (for example, 1-bit) code is assigned when the probability codebook is not changed, and the frequency is reduced when the probability codebook is changed. Try to assign a long code. With this method, it is possible to shorten the average code length.

Although the first to third embodiments have been described on the assumption that the probability codebook and the adaptive codebook are used, the present invention can be applied to other methods using the probability codebook. is there.

[0036]

As described in detail above, according to the present invention,
By classifying the probability codebooks according to the pitch period and having a plurality of probability codebooks, it is possible to provide a speech coding apparatus capable of reproducing speech satisfactorily for both males and females.

[Brief description of drawings]

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

FIG. 2A is a block configuration diagram of a speech encoding apparatus according to a second embodiment of the present invention, and FIG. 2B is a second configuration of the present invention.
FIG. 6 is a diagram showing data encoded according to the example of FIG.

FIG. 3 is a block configuration diagram of a conventional speech encoding device.

[Explanation of symbols]

10 ... Input speech, 12 ... Linear prediction analyzer, 14 ... Difference calculator, 16 ... Synthesis filter, 18, 20 ... Stochastic codebook, 22 ... Stochastic codebook gainer, 24 ... Adaptive codebook, 26 ... Adaptive codebook gain Unit, 28 ... Adder, 30 ... Buffer, 32 ... Weighting filter, 34
... Evaluator, 36 ... Pitch detector, 38 ... Flag.

Claims (1)

[Claims] 1. A speech coding apparatus for obtaining a residual by obtaining a difference between a predicted value signal from a prediction filter for predictively analyzing an input speech signal and the input speech signal, and quantizing the residual by a codebook. In each of the plurality of codebooks having different contents, a fundamental frequency detecting unit for detecting a fundamental frequency of the voice from the input voice signal, and a fundamental frequency of the voice detected by the fundamental frequency detecting unit, A speech coding apparatus comprising: a codebook switching means for switching a plurality of codebooks.