JPH11249696A - Voice encoding/decoding method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To realize an auditorily natural low rate voice encoding method even though coincidence with an original sound is slightly lost. SOLUTION: When an LPC coefficient and a drive signal obtained by analyzing an input voice with an LPC analysis part 111 and a drive signal analysis part 112 are encoded, the predicted values of the LPC coefficient and the drive signal are obtained by prediction parts 141, 142, and the analysed values are compared with the predicted values related to each LPC coefficient and drive signal by an evaluation part 150. Then a degree of an effect imparted to auditory naturalness when the predicted values are used for generating a decoded voice is evaluated, and more quantization bits are allocated to the quantizing side having a larger degree of the effect between the quantization parts 121, 122 of the PC coefficient, the drive signal based on this evaluation result by a bit allocation part 160, and an LPC coefficient quantization index 1001, a drive signal quantization index 1002 and bit allocation information 1003 are transmitted.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a speech encoding method and a speech decoding method, and more particularly to a low-rate speech encoding / decoding method used for digital telephones, voice memos, and the like.

[0002]

2. Description of the Related Art In recent years, with the development of communication means such as cellular phones and the Internet, coding techniques for compressing a voice signal or a tone signal into a small amount of information for transmission or storage have been actively studied. In the time domain, CELP (Code
Excited Linear Prediction) and subband coding in the frequency domain.

[0003] Code Excited Linear Prediction (MRSc
hroeder and BSAtal, “Code Excited Linear Predic
tion (CELP): High Quality Speech at Very LoW Bit Rat
es, ”Proc. ICASSP, pp. 937-940, 1985 (Reference 1), and WSKleijin, DJ Krasinski et al.“ Improved Speec
h Quality and Efficient Vector Quantization in SEL
P, “Proc. ICASSP, pp. 155-158, 1988 (Reference 2)) and the like are CELP schemes based on linear prediction analysis.

According to the CELP system, an input speech signal is decomposed into two parameters by a linear prediction analysis, a linear prediction coefficient representing phoneme information and a prediction residual signal representing pitch and the like, and these parameters are encoded. You. The linear prediction coefficient is used as a filter coefficient of a synthesis filter composed of a recursive digital filter. On the decoding side, the original input speech signal is obtained by inputting the prediction residual signal to this synthesis filter.

In this case, in order to perform encoding at a low rate, it is necessary to represent a prediction residual signal with as little information as possible. The feature of the CELP method is that several kinds of signals called excitation signals, which are candidates for the prediction residual signal, are stored in a codebook, and each excitation signal is generated through a synthesis filter. It is configured to evaluate how close the signals are and select an excitation signal that generates a synthesized voice signal closest to the input voice signal. Accordingly, as the encoding rate is increased to increase the number of excitation signals, the waveform of the decoded audio signal approaches the waveform of the input audio signal, and as a result, a decoded audio signal close to the input audio signal is obtained.

As one of the variations of the CELP system, a multi-mode CELP is known. In the CELP method, as described above, an input speech signal is decomposed into parameters representing linear prediction coefficients (filter coefficients) and prediction residual signals (excitation signals) and encoded. The nature of the input speech is not constant, and changes from time to time, for example, in voiced and unvoiced sections. Therefore, in the multi-mode CELP system, a plurality of types of encoding systems having different bit distributions for each parameter and different codebook contents are prepared according to the characteristics of the input speech, and these are prepared according to the input speech. Efficiency can be improved by properly using them.

On the other hand, apart from the multi-mode CELP, a sub-band coding method used in MPEG audio and the like is known as a coding method for changing the bit allocation of coding according to input speech. Subband coding is a method in which an input audio signal is converted into a frequency domain, divided into a plurality of bands, and coded for each band. The number of coded bits for each band component has a large power of the component. Are distributed as much as possible. Since the masking effect is taken into account, it is not necessarily allocated only by power, but basically by adaptively allocating bits to the input voice signal whose characteristics fluctuate, the coding efficiency can be increased. I have.

[0008] The purpose of adapting the bit allocation in the multi-mode CELP system or the sub-band coding system as described above is to obtain a decoded speech that is more audible to the input speech. For this reason, it is necessary to faithfully reproduce changes in parameters (formant, pitch period, gain, etc.) and fine fluctuations included in the input voice, which further improves the coding efficiency (lower Rate).

However, for example, when the same sentence is uttered twice, as can be seen from the fact that the same sound is heard by the human ear but the sound waveform level is different, the parameters included in the input sound are different. Information such as changes and small fluctuations is not necessarily required to be transmitted in terms of auditory naturalness. Furthermore, even if an auditory difference is recognized when two utterances are compared, the difference may not be so large as to cause a problem.

[0010]

As described above, conventional low-rate speech coding techniques such as multi-mode CELP and sub-band coding for adapting bit allocation to improve coding efficiency are audible. From the viewpoint of naturalness, information such as parameter changes and small fluctuations included in the input speech that are not necessarily required is faithfully reproduced, which hinders further improvement in coding efficiency.

The present invention has been made in view of such circumstances, and does not aim to obtain the same decoded speech as the input speech in an auditory manner. Coding /
It is an object to provide a decoding method.

[0012]

To solve the above-mentioned problems, a speech encoding method according to the present invention is a speech encoding method for encoding a plurality of parameters obtained by analyzing an input speech. By calculating the predicted value of each parameter from past coded data and comparing the analysis value with the predicted value for each parameter, the effect on the auditory naturalness when the predicted value is used to generate decoded speech The degree is evaluated, and based on the evaluation result, more coded bits are allocated to the parameter having a greater degree of influence, and bit allocation information indicating the coded data of each parameter and the allocation of coded bits to each parameter. Is transmitted.

[0013] In addition, the speech decoding method according to the present invention corresponding to this speech encoding method comprises, for a plurality of parameters obtained by analyzing input speech, encoding data of the parameters and a predicted value of the parameters. A bit indicating the distribution of coded bits determined such that the degree of the influence on the auditory naturalness when the decoded speech is generated is determined so that the degree of the influence is more distributed to the larger parameter based on the evaluation result. The method is characterized in that allocation information is input, and coded data is decoded according to the bit allocation information to generate a decoded speech.

Another speech encoding method according to the present invention is a speech encoding method for encoding a plurality of parameters obtained by analyzing an input speech, wherein a predicted value of each parameter is calculated by past encoding. From the data, the analysis value and the prediction value are compared for each parameter to evaluate the degree of the effect on the auditory naturalness when the prediction value is used to generate the decoded speech. The method is characterized in that only the parameter having a greater degree of influence is selected and encoded data of the analysis value of the parameter is transmitted, and selection information indicating the type of the selected parameter is transmitted.

In another speech decoding method according to the present invention corresponding to this speech encoding method, a decoded speech is generated for a plurality of parameters obtained by analyzing an input speech by using predicted values of the parameters. Input the coded data of the analysis value of the parameter having the greater degree of the influence selected based on the evaluation result of the degree of the influence on the auditory naturalness in the case of the selection, and the selection information indicating the type of the selected parameter. Then, based on the selection information, for the selected parameter, decode the encoded data of the analysis value to generate a decoded speech, and for the unselected parameter, generate the decoded speech using the predicted value as it is. Features.

As described above, according to the present invention, changes in parameters of different properties included in the input speech are predicted, and parameters that are perceptually unnatural when the predicted values are used as they are in the generation of decoded speech. On the other hand, allocate more coded bits, or allocate all coded bits, and use a smaller number of bits for parameters that are not unnatural even if they differ from the input speech. Or do not distribute any coded bits,
The prediction value is used as it is for generating the decoded speech.
Further, it is desirable that the bit allocation of each parameter to the encoded data be changed every moment depending on the characteristics of the input speech. As a result, although the consistency between the decoded speech and the original input speech is slightly sacrificed, speech encoding / decoding at a much lower rate than before becomes possible.

[0017]

Embodiments of the present invention will be described below with reference to the drawings. Note that the audio encoding / decoding method of the present invention is often realized by software using a computer, but will be described below as an audio encoding / decoding system using a block diagram. Even with such a description, it is considered that a person skilled in the art can clearly understand the procedure as the speech encoding / decoding method according to the present invention.

(First Embodiment) <Regarding the Encoding Side> FIG. 1 shows the configuration of a speech encoding system according to a first embodiment of the present invention. The speech encoding system includes an LPC analysis unit 111 and a drive signal analysis unit 112 that analyze an input audio signal from the input terminal 100.
, Which performs quantization of LPC coefficients and drive signals, respectively.
2. First and second delay units 131 and 132 for delaying the LPC coefficient and the quantization value of the drive signal, respectively, first and second prediction units 141 and 142 for predicting the LPC coefficient and the drive signal, respectively. The evaluator 150 evaluates the degree to which the predicted values from the units 141 and 142 are applied to the auditory naturalness when the decoded values are used as they are in the decoded speech, and the quantizers 121 and 122 based on the evaluation results of the evaluator 150.
And a bit allocation unit 160 for allocating the number of quantization bits (the number of coding bits).

Next, the operation of the speech encoding system will be described. An audio signal is input to the input terminal 100 on a frame-by-frame basis.
In LP, linear prediction analysis is performed and LP corresponding to vocal tract characteristics
The analysis value of the C coefficient is output, and the analysis value of the drive signal corresponding to the vocal cord waveform is output from the drive signal analysis unit 112.

The first quantization unit 121 uses the LPC coefficient prediction value output from the prediction unit 141 to divide the LPC coefficient analysis value from the LPC analysis unit 111 into the bit distribution unit 16.
Quantize with the number of quantization bits allocated by 0, LP
At the same time as outputting the C coefficient quantization index 1001, the quantization value of the LPC coefficient is supplied to the delay unit 131 for prediction of the next frame.

The second quantizing section 122 similarly uses the predicted value of the drive signal output from the second predicting section 142,
The analysis value of the drive signal from the drive signal analysis unit 112 is quantized by the number of quantization bits allocated by the bit allocation unit 160, and the drive signal quantization index 1002 is output. The signal is supplied to the delay unit 132 for prediction.

In the evaluation section 150, first, the LPC analysis section 11
Influence on the voice quality of the decoded speech when the predicted value of the LPC coefficient obtained by the first prediction unit 141 is used as it is as the filter coefficient of the synthesis filter based on the analysis value of the LPC coefficient obtained in Step 1 Impact on natural nature) is required.

Further, the evaluation section 150 uses the analysis value of the drive signal obtained by the drive signal analysis section 112 to input the predicted value of the drive signal obtained by the second prediction section 142 as it is to the input of the synthesis filter. The degree of influence on the sound quality of the decoded speech (effect on auditory naturalness) when used is also required.

Then, information indicating the degree of the above two kinds of influences is sent from the evaluation section 150 to the bit allocation section 160. The bit allocation unit 160 determines the allocation of the number of quantization bits of the first and second quantizers 121 and 122 based on this information so as to allocate more quantization bits to the one having a greater influence. . At the same time, the bit allocation section 160 outputs bit allocation information 10 which is information on the number of quantization bits allocated to each of the first and second quantizers 121 and 122.
03 is output.

The coded data output from the speech coding system has an LPC coefficient quantization index 100
1, drive signal quantization 1002 and bit allocation information 10
03, which are transmitted to an audio decoding system described later via a transmission path or a recording medium.

Next, an encoding method of LPC coefficients according to the present embodiment will be described. In the first delay unit 131,
The quantization value of the LPC coefficient quantized in the past frame in the first quantization unit 121 is stored. The first prediction unit 141 predicts the value of the LPC coefficient of the current frame using the quantization value of the LPC coefficient of the past frame stored in the delay unit 131. If the predicted value of the LPC coefficient is the same as the analysis value of the LPC coefficient obtained by analyzing the input audio signal in the LPC analysis section 111, there is no need to transmit information of the LPC coefficient at all. This is because the decoding side also has the same prediction algorithm, so that the same current prediction value as the encoding side can be obtained from the past quantization value.

In practice, it is rare that the predicted value of the LPC coefficient is the same as the analysis value. In many cases, a difference occurs. Therefore, it is necessary to quantize and transmit the difference. The feature is that the distribution is determined according to the difference between the number of coded bits and another parameter (in the present embodiment, the predicted value of the drive signal obtained by the same configuration as the LPC coefficient) and the importance of the input audio signal. .

Here, in the present embodiment, the bit allocation method for each parameter is changed as in the conventional multi-mode CELP or the like, and a bit distribution method that minimizes the distortion to the input voice signal is selected. Bit allocation is performed so that the signal can be heard naturally even if it is perceptually different from the signal. For example, the value of the LPC coefficient may change even in a stationary section depending on the position of the analysis window. However, in the multi-mode CELP, encoding is performed so that such a difference can be faithfully reproduced. In this case, a code pattern such as ABAB... In which two different codes A and B alternately appear may occur.

In this embodiment, when a code pattern such as ABAB... Appears, all the codes are replaced with the codes A or all the codes B, so that the predicted values of the LPC coefficients are continuously transmitted. ABAB ..., AAAA ..., BBBB
It seems that there is some auditory difference between the code patterns of. However, when hearing the decoded voice generated from each code pattern, this is not a problematic difference. Considering the amount of code transmission, the code pattern of ABAB... Must keep transmitting the difference between the code A and the code B every frame, whereas the code pattern of AAAA... And BBBB. Once obtained, the same value is predicted thereafter, so that the transmission amount of the subsequent frame can be suppressed to 0, and the effect of reducing the bit rate is great. By slightly sacrificing the consistency of the decoded speech with the input speech signal as the original speech, the bit rate can be greatly reduced.

The quantization of the LPC coefficient has been described above, but the same can be said for the quantization of the driving signal. Here, the point of the present embodiment is that more coded bits are allocated to the LPC coefficient and the drive signal that have a greater auditory effect. In the transmission method of the encoded data of the LPC coefficient described above, the code pattern of ABAB.
Since a decoded voice closer to the original voice can be obtained than the code patterns of AAA... And BBBB..., The code pattern of ABAB.

Whether the code pattern of ABAB ... is transmitted,
Whether transmission is simplified to a code pattern of AAAA... Or BBBB... Depends on the degree of change of the drive signal. When the pitch period or gain of the drive signal does not change much, a large number of bits are allocated to the encoding of the LPC coefficient and transmitted as a code pattern of ABAB... To generate a decoded voice more faithful to the original voice. Becomes possible. Conversely, when the change in the drive signal is large and has a large auditory effect, many bits are assigned to the sign of the drive signal and the LPC
The coefficients are endured in a code pattern of AAAA... Or BBBB. At this time, although the consistency between the decoded speech and the original speech is somewhat sacrificed, a great advantage is that the auditory naturalness is maintained.

The LPC coefficients are representative of parameters representing vocal tract characteristics, and include various expression methods such as k parameters, LSP, and LPC cepstrum. It is also possible to express the vocal tract characteristics with parameters other than the LPC coefficient.
In the present embodiment, the LPC coefficient is used as an example of a parameter representing a vocal tract characteristic, but another parameter may be used, and this point is the same in other embodiments described below. <Regarding the Decoding Side> FIG. 2 shows a configuration of a speech decoding system corresponding to the speech encoding system of FIG. This speech decoding system includes first and second dequantizers 221 and 222 for dequantizing LPC coefficient quantization index 1001 and drive signal quantization index 1002 input to input terminals 201 and 202, respectively. Terminal 2
The bit allocation unit 210 that determines the bit allocation to the inverse quantization units 221 and 222 based on the bit allocation information 1003 input to the input unit 03 and the LPC coefficients and the drive signal inversely quantized by the inverse quantization units 221 and 222 First and second delay units 231 and 232 for delaying, respectively, first and second prediction units 24 for predicting LPC coefficients and drive signals, respectively.
1 and 242, and a synthesizing unit 250 that generates a decoded audio signal from the LPC coefficient and the driving signal dequantized by the dequantizing units 221 and 222.

More specifically, the synthesizing unit 250 is an LP to which inversely quantized LPC coefficients are given as filter coefficients.
The LPC synthesis filter is constituted by using a C synthesis filter, and by inputting the inversely quantized drive signal to the LPC synthesis filter,
Generate a decoded audio signal.

(Second Embodiment) <On the Encoding Side> FIG. 3 shows a speech encoding system according to a second embodiment of the present invention. This speech coding system removes the bit allocation unit 160 in the speech coding system of the first embodiment shown in FIG.
The first and second quantization units 121,
Coding parameter selection control section 170 that determines which of the LPC coefficient from step 122 and the quantization index of the drive signal is to be selected as the coding parameter to be transmitted
Any of the LPC coefficients from the quantization units 121 and 122 and the quantization index of the drive signal (LPC coefficient / drive signal quantization index) is controlled by the encoding parameter selection control unit 170 as the encoding parameter 1005. The configuration is such that the switch 180 is replaced with an extraction switch 180.

The coding parameter selection control section 170 outputs coding parameter selection information 1004 indicating which of the LPC coefficient and the quantization index of the drive signal has been selected.

In the first embodiment, the method of changing the bit allocation of the LPC coefficient and the drive signal to the coded data in accordance with the LPC coefficient and the predicted value of the drive signal has been described. Push all the coded bits to one of the LPC coefficient and the drive signal. As a result, transmission of encoded data of LPC coefficients or transmission of encoded data of a drive signal changes for each frame. <Regarding the Decoding Side> FIG. 4 shows a speech decoding system corresponding to the speech encoding system of FIG. In this speech decoding system, according to the coding parameter selection information 1004 input to the input terminal 204, the coding parameter (LPC coefficient / drive signal quantization index) input to the input terminal 205 is inversely quantized by the inverse quantization units 221 and 222. And input. That is, the LPC coefficient quantization index is input to the inverse quantization unit 221, and the drive signal quantization index is input to the inverse quantization unit 222. The subsequent operation is the same as in FIG.

As described above, according to the present embodiment, in the frame in which the encoded data of the LPC coefficient is transmitted on the decoding side, the predicted value is used as it is for the drive signal, and the frame in which the encoded data of the drive signal is transmitted is used. Then, since the decoded speech is generated by using the predicted value as it is as the LPC coefficient, the consistency of the decoded speech with the original speech is lost as compared with the first embodiment, and a lower rate encoding is possible. Becomes

(Third Embodiment) <On the Encoding Side> FIG. 5 shows a speech encoding system according to a third embodiment of the present invention. In this speech coding system, the analysis unit 110 in the first embodiment is replaced by an analysis unit 190 including a pitch period analysis unit 191 and a pitch waveform analysis unit 192, and the LPC analysis unit 1
An LPC quantization unit 320 that quantizes the analysis value of the LPC coefficient 11 and the LPC coefficient is provided outside the analysis unit 310.

Further, in this embodiment, the first and second quantizers 321 and 322 for quantizing the pitch period and the pitch waveform, respectively, and the first and second quantizers for delaying the pitch period and the quantization value of the pitch waveform, respectively. 2 delay units 331,
332, first and second prediction units 341 and 342, which predict the pitch period and the pitch waveform, respectively, and the prediction unit 34
Evaluating unit 35 that evaluates the degree to which auditory naturalness is imparted when the predicted value from 1,342 is used to generate decoded speech.
There is provided a bit allocation unit 360 that allocates the number of quantization bits to the quantization units 321 and 322 based on 0 and the evaluation result of the evaluation unit 350.

Next, the operation of the speech coding system according to this embodiment will be described. An audio signal is input to the input terminal 100 on a frame-by-frame basis. In synchronization with this, a linear prediction analysis is performed, and LPC coefficients and a drive signal are separated and encoded. That is, the LPC analysis section 111 outputs an analysis value of the LPC coefficient, and the analysis section 190 separates the drive signal into a pitch cycle and a pitch waveform for one pitch by the pitch cycle analysis section 191 and the pitch waveform analysis section 192, respectively. You. The analysis value of the LPC coefficient output from the LPC analysis unit 111 is quantized by the LPC quantization unit 320, and the LPC coefficient quantization index 1001 is transmitted.

The first quantization section 321 uses the bit cycle allocation section 360 to distribute the pitch cycle analysis value from the pitch cycle analysis section 191 using the pitch cycle prediction value output from the prediction section 341. Quantize with the number of quantization bits,
At the same time as outputting the pitch period quantization index 1011, the quantization value of the pitch period is supplied to the delay unit 331 for prediction of the next frame.

The second quantizing unit 322 similarly uses the predicted value of the pitch waveform output from the second predicting unit 322 to transmit the analyzed value of the pitch waveform from the pitch waveform analyzing unit 192 to the bit distribution unit. Quantize with the number of quantization bits allocated by 360 and generate a pitch waveform quantization index 1012
, And supplies the quantized value of the pitch waveform to the delay unit 332 for prediction of the next frame.

The evaluation section 350 first uses the pitch period predicted value obtained by the first prediction section 321 as it is for the generation of decoded speech based on the pitch period analysis value obtained by the pitch period analysis section 191. In this case, the degree of influence of the decoded speech on the voice quality is determined. In addition, the evaluation unit 350 uses the predicted value of the pitch waveform obtained by the second prediction unit 322 to generate a decoded speech based on the analysis value of the pitch waveform obtained by the pitch waveform drive signal analysis unit 312. The degree of influence on the sound quality of the decoded speech in that case is also obtained.

Then, information indicating the degree of the above two types of influence is sent from the evaluation section 350 to the bit allocation section 360. The bit allocation unit 360 determines the allocation of the number of quantization bits of the first and second quantizers 321 and 322 based on this information so as to allocate more quantization bits to the one having a greater influence. . At the same time, the bit allocation section 360 outputs bit allocation information 10 which is information on the number of quantization bits allocated to each of the first and second quantizers 341 and 342.
13 is output.

The coded data output from the speech coding system has an LPC coefficient quantization index of 100
1, pitch cycle quantization index 1011, pitch waveform quantization index 1012 and bit allocation information 1
013, which are transmitted to a later-described audio decoding system via a transmission path or a recording medium.

In the first embodiment, the example in which the bit allocation of the encoded data is changed with respect to the LPC coefficient and the drive signal has been described. In the present embodiment, the bit allocation is changed with respect to the pitch period and the encoded data of the pitch waveform. It is like that.

First, the pitch period will be described. The pitch period has a flat section with little change, a section with a large change such as a sharp rise and fall, and a slight fluctuation even in a flat section. Conventionally, such a change in pitch cycle has been faithfully encoded.

On the other hand, in the present embodiment, it is not intended to always match the pitch period of the input voice signal,
Predict pitch patterns that sound audibly natural. If the prediction is appropriate, there is no need to transmit information on the pitch period. This is because the decoding side also has the same prediction algorithm, so that the same prediction value as the encoding side can be generated from past quantized values. If there is an auditory problem that deviates from the prediction, more coded bits are allocated and a pitch cycle closer to the pitch cycle of the input speech signal is transmitted. The same applies to the pitch waveform.

Further, in the present embodiment, more coded bits are allocated to those having a greater auditory effect on the pitch period and the pitch waveform. When the pitch period is substantially constant, on the other hand, if the pitch period changes but the pitch waveform does not change, more coded bits are allocated to the pitch period. By doing so, the matching of the pitch period with the original sound is somewhat sacrificed, but the drive signal can be encoded at a low rate. <Regarding the Decoding Side> FIG. 6 shows the configuration of a speech decoding system corresponding to the speech encoding system of FIG. This audio decoding system has input terminals 201, 211, 212
LPC coefficient quantization index 1 respectively input to
001, the pitch period quantization index 1011, the inverse quantization units 420, 421 and 422 for inversely quantizing the pitch waveform quantization index 1012, and the inverse quantization unit 4 based on the bit allocation information 1013 input to the input terminal 213.
Bit allocation unit 4 that determines bit allocation to 21, 422
10. First and second delay units 431 and 432 for respectively delaying the pitch cycle and the pitch waveform dequantized by the inverse quantization units 421 and 422, and the first and second delay units for predicting the pitch cycle and the pitch waveform, respectively. 2 prediction units 441, 4
42, and a synthesis unit 450 that generates a decoded speech signal from the LPC coefficient, pitch period, and pitch waveform dequantized by the dequantization units 402, 421, and 422.

More specifically, the synthesizing section 450 is an LP to which inversely quantized LPC coefficients are given as filter coefficients.
A decoded speech signal is generated by inputting a drive signal generated from the dequantized pitch cycle and pitch waveform to this LPC synthesis filter.

(Fourth Embodiment) FIG. 7 shows a fourth embodiment of the present invention.
1 shows a speech encoding system according to an embodiment. This speech coding system has a configuration in which the LPC analysis unit 111 and the LPC quantization unit 320 are removed from the speech coding system according to the third embodiment shown in FIG. In the third embodiment, the input audio signal is subjected to linear pre-analysis, and is separated into LPC coefficients and drive signals and encoded. However, in the present embodiment, encoding is performed without separation.

Specifically, in the third embodiment, the pitch period and the pitch waveform of the drive signal are extracted.
In the present embodiment, the pitch cycle of the input voice signal is extracted by the pitch cycle analysis unit 311, and the pitch waveform of the input voice signal is cut out by the pitch waveform analysis unit 312.

According to the present embodiment, the LPC analysis required in the first to third embodiments becomes unnecessary, and the present invention can be applied to an input voice signal having a small effect of the LPC analysis. In the following embodiments, an example in which the LPC analysis is not performed is shown. However, as in the relationship between the present embodiment and the third embodiment, the present invention is also applicable to the case where the LPC analysis is performed. <Regarding the Decoding Side> FIG. 8 is a diagram showing a configuration of a speech decoding system corresponding to the speech encoding system of FIG. The configuration is such that the inverse quantization unit 420 is removed. In this case, the synthesizer 450 generates a decoded speech signal by connecting the inversely quantized pitch waveforms with the inversely quantized pitch period by an appropriate method.

(Fifth Embodiment) FIG. 9 shows a fifth embodiment of the present invention.
1 shows a speech encoding system according to an embodiment. This speech coding system includes the first and second prediction units 341 and 3 in the speech coding system according to the fifth embodiment shown in FIG.
This shows the configuration of the evaluation unit and the evaluation unit 350 in detail, and other components and operations thereof are the same as those of the fourth embodiment. The prediction unit 341 includes a comparison unit 3411 and a pitch pattern codebook 3412, and the prediction unit 342 includes a comparison unit 3421 and a waveform pattern codebook 3422. Further, the evaluation unit 350 includes two error integration units 3511 and 352.
Consists of two.

First, the first predictor 341 will be described. The pitch pattern codebook 3412 of the prediction unit 341 stores a plurality of pairs of pitch patterns obtained by learning from many voices and subsequent pitch periods appearing next to the patterns. Here, the pitch pattern represents a change in the pitch cycle of the voice, and is assumed to be the pitch cycle of the past N frames. At the time of encoding, the delay unit 3
A comparison unit 3411 compares the past N quantized pitch periods stored in 31 with the pitch pattern (N pitch periods) stored in the pitch pattern codebook 3412, and the change form of the pitch period is the closest. A candidate is searched for, and a subsequent pitch period paired with the candidate is output as a predicted pitch period.

Next, the second predictor 342 will be described. The operation of the prediction unit 342 is the same as that of the prediction unit 342, except that the parameter to be handled is changed from the pitch period of the prediction unit 341 to a pitch waveform. That is, the past M stored in the delay unit 332
The number of pitch waveforms and the M number of pitch waveforms stored in the waveform pattern codebook 3422 are compared to find the closest one, and the subsequent pitch waveform paired with this is output as a predicted pitch waveform.

Next, the evaluation section 350 will be described. In the error integration section 3511 of the evaluation section 350, the pitch analysis section 31
The analysis value of the pitch period obtained in step 1 and the predicted value of the pitch period output from the prediction unit 341 are input, and the difference is evaluated. At this time, not only the difference in the pitch cycle of the current frame but also how much the change in the pitch cycle of the entire frame is shifted back to the past several frames is evaluated. This shift in change is referred to as an integral value here.

Even if the pitch cycle is temporarily shifted, there is no problem from the viewpoint of auditory naturalness. However, if the pitch cycle is continuously shifted for a long period of time, an unnatural intonation is lost and the naturalness is impaired. Therefore, the error accumulating unit 3511 determines that the larger the difference between the analysis value and the predicted value of the pitch period over a long period of time, the larger the difference, and sends the information to the bit distribution unit 360 as error information.

The error accumulating unit 3512 is different from the error accumulating unit 3512 except that the information to be handled is not a pitch period but a pitch waveform.
Same as 11. That is, the temporal change of the predicted value of the pitch waveform and the analysis value are compared, and the error information of a larger value is output to the bit distribution unit 360 as the deviation continues for a long time.

In the bit distribution section 360, the error integration section 35
The pitch period and the number of quantization bits of the pitch waveform are determined on the basis of the error information from
The number of bits is output to 22.

The quantization unit 321 calculates the pitch period of the current frame based on the analysis value of the pitch period obtained from the pitch period analysis unit 311 and the predicted value of the pitch obtained from the comparison unit 3411 using the allocated number of bits. decide. Normally, the pitch period of the current frame is determined so as to be closer to the analysis value of the pitch period, an index indicating this is output, and at the same time, the quantization value is supplied to the delay unit 331 and stored for use in the next frame. Keep it.

The quantization unit 322 also performs quantization of the pitch waveform in the same manner. Codebooks are often used for quantization,
This codebook is different from the pitch pattern codebook 3412 and the waveform pattern codebook 3422, and is included in the quantization unit 322 in FIG. 9 and is not explicitly shown.

As described above, in the present embodiment, by evaluating the shift of the pitch cycle and the change of the pitch waveform over a long period of time, it becomes possible to perform low-rate encoding that can obtain decoded speech that sounds more audible and natural. In addition, it is effective to normalize the pitch period in order to reduce the amount of memory when creating the pitch pattern codebook 3411. <Regarding the Decoding Side> FIG. 10 is a diagram showing a configuration of a speech decoding system corresponding to the speech encoding system of FIG. 9, and the first and the first speech decoding systems of the fifth embodiment shown in FIG. 4 shows the configuration of the second prediction units 441 and 442 in detail. That is, the prediction unit 441 includes a comparison unit 4411 and a pitch pattern codebook 3412,
The prediction unit 442 includes the comparison unit 4421 and the waveform pattern codebook 4
422.

Next, the operation of the speech decoding system will be described. In bit allocation section 410, the number of quantization bits allocated to quantization sections 321 and 322 in FIG. 9 is obtained from the bit allocation information input to input terminal 213, and these are supplied to inverse quantization sections 421 and 422. Each given.

A pitch period index is input to the input terminal 301, and the index and the first prediction unit 441
Inverse quantization unit 4 based on the predicted pitch period output from
At 21, the pitch period is decoded and input to the synthesis unit 450, and at the same time, the delay unit 431 is prepared for the processing of the next frame.
Is input to Since the prediction unit 441 has exactly the same configuration including the codebook as the prediction unit 331 of the speech coding system in FIG. 9, the same output as the prediction unit 331 can be obtained without side information.

The pitch waveform index is input to the input terminal 215, and this index and the second prediction unit 442
Inverse quantization unit 4 based on the predicted value of the pitch waveform output from
At 22, the pitch period is decoded and input to the synthesizing unit 450, and at the same time, the delay unit 432 prepares for the processing of the next frame.
Is input to

The synthesizing section 450 generates a decoded speech by connecting the pitch waveforms at a pitch period by using an appropriate means. Sixth Embodiment <On the Encoding Side> FIG. 11 shows a speech encoding system according to a fifth embodiment of the present invention. This speech coding system removes the bit allocation unit 360 in the speech coding system according to the fourth embodiment shown in FIG. 7, and according to the evaluation result of the evaluation unit 350, the first and second quantization units 32.
1 and 322, a coding parameter selection control unit 370 that determines which of the pitch period and the quantization value of the pitch waveform is to be selected as a coding parameter, and is controlled by the coding parameter selection control unit 370. ,
One of the pitch cycle and the quantization index of the pitch waveform (pitch cycle / pitch waveform quantization index) from the second quantization units 321 and 322 (pitch cycle / pitch waveform quantization index) is replaced with a changeover switch 380 for extracting as an encoding parameter 1015. I have. The encoding parameter selection control unit 370 outputs encoding parameter selection information 1014 indicating which of the pitch period and the quantization index of the pitch waveform has been selected as the encoding parameter.

Then, either the pitch period or the pitch waveform is selected for each frame, and all the coded bits that can be allocated to the selected one are assigned and coding is performed. If the coded bits are not allocated, the predicted value is used as it is. <Regarding the Decoding Side> FIG. 12 is a diagram showing a speech decoding system corresponding to the speech encoding system of FIG.
In this speech decoding system, the coding parameter (pitch cycle / pitch waveform quantization index) input to the input terminal 215 is changed according to the coding parameter selection information 1014 input to the input terminal 214.
1, 422, and input. That is, the pitch period quantization index is calculated by the inverse quantization unit 421.
The pitch waveform quantization index is calculated by the inverse quantization unit 422.
Respectively. The subsequent operation is the same as in FIG.

As described above, according to the present embodiment, in a frame in which encoded data of a pitch period is transmitted, a predicted value is used for generation of decoded speech in a pitch waveform, and a frame in which encoded data of a pitch waveform is transmitted is used. Then, by using the predicted value for the generation of the decoded speech in the pitch period, the consistency of the decoded speech with the original speech is lost as compared with the fourth embodiment, and a lower-rate encoding becomes possible. .

(Eighth Embodiment) FIG. 13 shows a speech decoding system according to an eighth embodiment of the present invention. This embodiment combines the first embodiment and the fourth embodiment, and further includes a pitch shape analyzer 51 in an analyzer 510.
1 and a gain analysis unit 512 to analyze the pitch waveform by decomposing it into a shape and a gain, and to extract all four types of parameters from the input speech signal, that is, pitch period, pitch shape, gain, and LPC coefficient. ing. Subsequent processing is the same as in the previous embodiments, and encoding is performed by allocating more bits to a parameter whose predicted value may sound unnaturally audible.

As described above, by classifying the parameters of the audio signal more finely, the bits can be concentrated on the necessary parameters, and the encoding efficiency is further improved. Although a speech decoding system corresponding to the speech encoding system in FIG. 13 is not particularly shown, it is apparent from the embodiments described above, and thus detailed description will be omitted.

[0072]

As described above, according to the present invention, it is possible to realize low-rate speech encoding / decoding in which the consistency with the original speech is somewhat lost, but an auditory natural decoded speech can be obtained. it can.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of a speech coding system according to a first embodiment of the present invention.

FIG. 2 is a block diagram showing a configuration of a speech decoding system according to the first embodiment of the present invention.

FIG. 3 is a block diagram showing a configuration of a speech coding system according to a second embodiment of the present invention.

FIG. 4 is a block diagram showing a configuration of a speech decoding system according to a second embodiment of the present invention.

FIG. 5 is a block diagram showing a configuration of a speech coding system according to a third embodiment of the present invention.

FIG. 6 is a block diagram showing a configuration of a speech decoding system according to a third embodiment of the present invention.

FIG. 7 is a block diagram showing a configuration of a speech coding system according to a fourth embodiment of the present invention.

FIG. 8 is a block diagram showing a configuration of a speech decoding system according to a fourth embodiment of the present invention.

FIG. 9 is a block diagram showing a configuration of a speech coding system according to a fifth embodiment of the present invention.

FIG. 10 is a block diagram showing a configuration of a speech decoding system according to a fifth embodiment of the present invention.

FIG. 11 is a block diagram showing a configuration of a speech coding system according to a sixth embodiment of the present invention.

FIG. 12 is a block diagram showing a configuration of a speech decoding system according to a sixth embodiment of the present invention.

FIG. 13 is a block diagram showing a configuration of a speech coding system according to a seventh embodiment of the present invention.

[Explanation of symbols]

100 audio signal input terminals 110, 310, 510 analysis unit 111 LPC analysis unit 112 drive signal analysis units 121, 122, 321, 322, 521, 522 quantization units 131, 132, 331, 332, 531, 532 delay unit 141, 142, 341, 242, 541, 542 prediction unit 150, 350, 550 evaluation unit 160, 360, 560 bit allocation unit 170, 370 encoding parameter selection control unit 180, 380 switching Switch 190 Analysis section 191 Pitch period analysis section 192 Pitch waveform analysis section 210, 410 Bit allocation section 221, 222, 421, 422 Inverse quantization section 231, 232, 431, 432 Delay section 241, 242, 441, 442: prediction unit 250, 450 ... synthesis unit 260, 460: switch Switch 311 Pitch period analysis unit 312 Pitch waveform analysis unit 313 Gain analysis unit 320 LPC quantization unit 511 Pitch shape analysis unit 512 Gain analysis unit 1001 LPC coefficient quantization index 1002 Drive signal quantization index 1003 ... bit allocation information 1004 ... coding parameter selection information 1005 ... LPC coefficient / drive signal quantization index 1011 ... pitch period quantization index 1012 ... pitch waveform quantization index 1013, 1018 ... bit allocation information 1014 ... coding parameter selection information 1015 ... Pitch period / pitch waveform quantization index 1016 ... Pitch shape quantization index 1017 ... Gain quantization index 1020 ... Decoded speech signal 3411, 3421 ... Pitch pattern codebook 3 12,3422 ... waveform pattern codebook 3511,3512 ... error integration unit 4411,4421 ... pitch pattern codebook 4412,4422 ... waveform pattern codebook

Claims (4)

[Claims] 1. A speech encoding method for encoding a plurality of parameters obtained by analyzing input speech, wherein a predicted value of each parameter is obtained from past encoded data, and an analysis value and a predicted value of each parameter are obtained. By performing the comparison, the degree of the effect on the auditory naturalness when the predicted value is used to generate the decoded speech is evaluated. Based on the evaluation result, more codes are assigned to a parameter having a higher degree of the effect. A coded bit, and transmitting the coded data of each parameter and bit allocation information indicating allocation of coded bits to each parameter. 2. A plurality of parameters obtained by analyzing an input voice, the effect of which on the perceived naturalness when a decoded voice is generated using encoded data of the parameters and a predicted value of the parameters. Based on the evaluation result of the degree, input the bit allocation information indicating the allocation of the coded bits determined so that the degree of the influence is more allocated to the larger parameter, and decode the coded data according to the bit allocation information. A speech decoding method characterized by generating a decoded speech by performing the decoding. 3. A speech encoding method for encoding a plurality of parameters obtained by analyzing an input speech, wherein a predicted value of each parameter is obtained from past encoded data, and an analysis value and a prediction value of each parameter are obtained. By comparing the values, the degree of influence on the auditory naturalness when the predicted value is used to generate the decoded speech is evaluated. Based on the evaluation result, only the parameter having a larger degree of the influence is evaluated. A speech encoding method comprising: selecting and transmitting encoded data of an analysis value of the parameter; and transmitting selection information indicating a type of the selected parameter. 4. A method in which a plurality of parameters obtained by analyzing an input voice are selected based on an evaluation result of a degree of influence on auditory naturalness when a decoded voice is generated using a predicted value of the parameter. The coded data of the analysis value of the parameter having a greater degree of influence and selection information indicating the type of the selected parameter are input. Based on the selection information, the code of the analysis value is selected for the selected parameter. A speech decoding method comprising: decoding decoded data to generate decoded speech; and generating a decoded speech by using the predicted value as it is for a parameter not selected.