JPH06266399A - Encoding device and speech encoding and decoding device - Google Patents

Abstract

PURPOSE:To improve the quality of a decoded speech by the device which separates an input speech into a spectrum parameter and a sound source signal and performs encoding at every frame having fixed time length. CONSTITUTION:The devices are equipped with a 1st code search means which searches for a combination of a sound source model code word with a spectrum code word minimizing the distortion of a synthesized speech and the input speech and outputs a 1st code search result 28, an adaptive sound source code book 6 which contains a 1st or 2nd quantized sound source signal 9 or 10 as an adaptive sound source signal, a driving sound source generating means 4 which generates a driving sound source signal from driving sound source code words in the driving sound source code book 2, a 2nd code search means which searches for a driving sound source code word minimizing the distortion of both the input speech and the synthesized speech generated by using a spectrum code word minimizing the spectrum distortion of a 2nd quantized sound source signal 10 generated by using an adaptive sound source signal and the driving sound source signal and outputs it as a 2nd code search result 13, and an encoding means selecting means 16 which outputs one of the 1st and 2nd code search results.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice coding / decoding device used for digital transmission or storage of voice.

[0002]

2. Description of the Related Art A conventional speech coder / decoder that separates an input speech into spectrum parameters and excitation signals and encodes each fixed-length frame is described in Reference 1 " Examination of analysis and synthesis method "(Katsushi Seza,
Hirohisa Tasaki, Kunio Nakajima, ASJ Autumn Meeting, 1
-6-10, PP209-210, 1991). In this conventional method, a source model defined on the differential waveform of the glottal source wave is used for encoding the source signal, and an autoregressive coefficient (hereinafter abbreviated as AR) and a moving average coefficient (hereinafter abbreviated as MA) are used as spectral parameters. Is used. In the above-mentioned document, as a method for obtaining a high-quality decoded speech in a low bit rate speech transmission of about 2 to 3 Kbps, an analysis and synthesis method (F
VQ-GARMA). There, it is clarified that a highly natural decoded voice can be obtained by vector-quantizing all vocal cord source wave models, AR parameters, and MA parameters. However, there is a case where the decoded speech is deteriorated due to the extraction error of the sound source peak position used for preliminary selection of the vocal cord source wave model and the transmission of the position (source position) matching the vocal cord source wave model. The conventional FVQ-GARMA method generates a voiced sound by driving an ARMA filter with a vocal cord source wave model. As the vocal cord source wave model, a model defined on the differential waveform of the vocal cord source wave is used. At the time of vector quantization of the vocal cord excitation wave model, AR and MA parameters, the encoding unit selects one combination having the maximum SNRseg for each frame from all combinations of the preselected codes. In the decoding unit, the codes obtained in each frame are interpolated and combined.

14 and 15 are block diagrams of a conventional speech coding / decoding apparatus according to the analysis and synthesis method (FVQ-GARMA). FIG. 14 shows an encoding unit, and FIG. 15 shows a decoding unit. In FIGS. 14 and 15, 1 is an input speech, 11 is an AR codebook, 12 is an AR codeword, 14 is a pitch period extraction means, 15 is a pitch period, 19 is a sound source start position extraction means, 20 is a sound source start position, 21 is a sound source model codebook, 22 is a sound source model codeword, 23 is a sound source model generation means, 26 is a MA codebook, 27 is a MA codeword, 29
Is an excitation model codebook, 30 is an excitation model codeword, 31 is an excitation model generation means, 34 is an MA codebook, 35 is an MA codeword, 37 is decoded speech, 44 is an AR codebook, and 45 is an AR.
A code word, 56 is an encoding result, 57 is a code search means, 58
Is a decoding means, 59 is a quantized excitation signal, and 60 is a quantized excitation signal.

First, the encoding unit shown in FIG. 14 will be described. The AR codebook 11 stores a plurality of typical ARs as AR codewords, and the MA codebook 26 stores typical MAs.
A plurality of A codewords are stored, and the excitation model codebook 21 stores a plurality of typical excitation model parameters representing excitation signals of one pitch period as excitation model codewords. The pitch period extracting means 14 extracts the pitch period 15 from the input voice 1 and outputs it. When the preceding frame is unvoiced and the frame is voiced, the sound source start position extraction means 19 extracts the sound source start position 20 from the input voice 1 and outputs it. The excitation model generation means 23 generates a signal in which the excitation signal of one pitch period generated from the excitation model codeword 22 output from the excitation model codebook 21 is repeated in the pitch period 15 and outputs it as a quantized excitation signal 59. . The code search means 57 generates a synthetic speech using the excitation start position 20, the AR codeword 12, the quantized excitation signal 59 and the MA codeword 27, and an AR codeword that minimizes distortion of the input speech 1 and the synthetic speech. And search for combinations of MA codewords and excitation codewords,
The encoded result 56 is output.

FIG. 16 illustrates the operation of the code search means 57 in the first frame of voiced sound. In the figure, the quantized excitation signal and the synthetic speech indicated by the solid line indicate the signal in the frame, and the quantized excitation signal and the synthetic speech indicated by the dotted line indicate the signal in the next frame. Code searching means 57
Encodes the input voice in units of pitch periods in voiced frames. When the frame is the head of the voiced sound, when the quantized sound source signals 59 are arranged from the sound source start position 20 in pitch cycle units, the input speech 1 in the range exceeding the frame is set as the encoding target of the frame and synthesized. Generate audio.

In FIG. 16, synthetic speech is generated from the sound source start position 20 to the pitch periods P1 and P2 as the current frame to be coded. Since the encoding is performed in units of pitch period, the frame delimiter and the pitch period delimiter do not match, and in this example, the time T3 is the time F.
2 does not match, and encoding is performed beyond the current frame only for the time T3 to time F2. For voiced frames other than the first frame of voiced sound, the quantized sound source signal 59
Are arranged until the quantized sound source signal in the preceding frame is exceeded and the synthesized speech is generated until the frame is exceeded. In the figure, speech synthesis is generated for the next frame from the pitch period of P2. That is, P3, P4, P
The pitch frame of 5 is used to encode the next frame to generate synthesized speech.

Next, the decoding unit shown in FIG. 15 will be described.
In the figure, an AR codebook 44, an excitation model codebook 31, M
The A codebook 34 is the same as the AR codebook 11, the excitation model codebook 21, and the MA codebook 26 in the encoding unit, respectively. The sound source model generating means 31 has a pitch period of 15
And the excitation model codeword 30 in the excitation model codebook 29 corresponding to the encoding result 56 is used to generate the quantized excitation signal 60. The decoding means 58 uses the AR codeword 4 in the AR codebook 44 corresponding to the quantized excitation signal 60 and the coding result 56.
5 and the MA codeword 35 in the MA codebook 34 are used to generate the decoded speech 37.

FIG. 17 illustrates the operation of the decoding means 58. The decoding means 58 arranges the quantized excitation signal 60 from the head of the frame until it exceeds the frame to generate the decoded speech 37. In the figure, decoded speech is generated from the beginning of the frame using frame periods 91, 92, and 93. Since the decoded speech is also generated in pitch period units, the decoded speech may be generated over a frame. In FIG. 17, the end time S3 of the pitch cycle 93 exceeds the time F2 of the frame, and even if the time F2 of the frame ends, the decoded speech for the frame is generated until the time S3.

In the next frame, the decoded speech 37 is generated by arranging the quantized excitation signal 60 shown by the dotted line next to the decoded speech decoded so far. In FIG. 17, pitch periods 94, 95 and 96 are used for generation of decoded speech of the next frame. Since the decoded voice generation of the next frame is also performed in pitch cycle units, when the decoded voice generation of the previous frame is performed over the previous frame,
7 As shown below, decoded speech is also generated in the next frame with a shift.

[0010]

In the conventional speech coding / decoding apparatus, the vocal cord excitation code is preselected based on the excitation peak position and the vocal cord excitation code selected in the past frame. There were many errors in automatic extraction of the sound source peak position at the beginning of the word and the transient part, and the preliminary selection sometimes did not work well. FIG. 18A shows the residual waveform, and FIG.
(B) and (c) show differential waveforms of the vocal cord source wave model. When the sound source peak position is correctly extracted (b), when the sound source peak position is erroneously extracted (c), the vocal cord source wave code is selected incorrectly, and the SNRseg rapidly deteriorates. As described above, since only the quantized excitation signal using the excitation model is used for encoding the excitation signal, the quality of the decoded speech may be deteriorated in the case of a speaker whose excitation model is poorly adapted. In addition, the quality of the decoded speech may be deteriorated because the fixed codebooks are quantized for the MA and the excitation model, which have different aspects depending on the pitch period. In addition, the encoding unit determines the head sound source position of the voiced sound and its pitch length so that SNRseg becomes maximum when the sound source is returned in error while finely adjusting the pitch length by interpolation. There is an example in which the decoded voice quality is deteriorated due to a large difference in the results of interpolating each code between the encoding unit and the decoding unit because the configuration is not transmitted to the decoding unit. That is, since the sound source start position in the head frame of voiced sound is not transmitted to the decoding unit, as shown in FIGS. 16 and 17, the number of sound source models included in the same frame in the encoding unit and the decoding unit is May be different. In such a frame, when the fluctuation of the power or the pitch period is large, the first quantized excitation signal of the decoding unit causes a large difference from the first quantized excitation signal of the encoding unit, and the decoded speech There were cases where the quality deteriorated.

The present invention has been made to solve the above problems, and an object thereof is to improve the quality of decoded speech.

[0012]

The speech coding / decoding apparatus according to the invention of claim 1 of the present invention, in order to reduce the deterioration of the sound quality, for example, multimodes GARMA and CELP, and uses the CELP system. A SNRseg is secured by using a sound source that uses the adaptive code and the driving sound source code that are used (see FIG. 18D), and when FVQ-GARMA is used and when the adaptive codebook and the driving sound source codebook are used. The one with the better SNRseg is selected.

In the speech coder / decoder according to the second aspect of the present invention, the synthesized speech generated from the first quantized excitation signal and the spectrum code word in the spectrum codebook is input to the coder. The combination of the excitation model codeword and the spectrum codeword that minimizes the distortion of the voice is searched, and the search result is output as the first code search result to the coding means selecting means together with the distortion at that time, and the first quantization is performed. A first code search means for outputting an excitation signal to an adaptive excitation codebook; an adaptive excitation codebook storing the first quantized excitation signal or the second quantized excitation signal obtained in the preceding frame as an adaptive excitation signal , A driving excitation codebook in which a plurality of prepared excitation signals are stored as driving excitation codewords, and a driving excitation signal in which the driving excitation codewords in the driving excitation codebook are repeated at the pitch cycle are generated. Driving excitation generating means, an adaptive excitation signal in the adaptive excitation codebook, a second quantized excitation signal generated from the driving excitation signal, and a synthetic speech generated using a spectrum codeword in the spectrum codebook. And a drive excitation codeword that minimizes the distortion of the input speech, and outputs the search result as the second code search result together with the distortion at that time to the encoding means selecting means to adapt the second quantized excitation signal. Second code search means for outputting to the excitation codebook, and one of the first code search result and the second code search result, whichever has a smaller distortion is selected as the coding result, and the code of the frame is selected. The coding unit selection unit outputs the coding result and outputs, as a coding unit selection signal, which coding search result is selected, and the decoding unit selects the coding unit input from the coding unit. Decoding means selecting means for selecting the first decoding means and the second decoding means according to the code, the same spectrum codebook as the coding section, the same excitation model codebook as the coding section, and the first quantum. First decoding means for generating decoded speech using a coded excitation signal and a spectrum codeword in the spectrum codebook corresponding to the coding result input from the coding section, and the same adaptive excitation code as the coding section Book, the same driving excitation codebook as the encoding unit, driving excitation code generation means for generating a driving excitation signal from the driving excitation codeword in the driving excitation codebook corresponding to the encoding result and the pitch period, The second decoding means is provided for generating decoded speech from the spectrum codeword, the adaptive excitation signal in the adaptive excitation codebook, and the driving excitation signal output from the driving excitation generating means.

According to a third aspect of the present invention, there is provided a speech coding / decoding apparatus in which an adaptive excitation codebook of a coding section and a decoding section holds a first quantized excitation signal. Storage means and second sound source storage means for holding a second quantized sound source signal; and code word switching means for switching between the first sound source storage means and the second sound source storage means. A second code search means for selecting from the adaptive excitation codebook an adaptive excitation signal that minimizes the distortion of, and outputting which is included in the second encoding result.

The spectrum codebook, the excitation model codebook, and the driving excitation codebook of the coding unit and the decoding unit of the speech coding / decoding apparatus according to the fourth and fifth aspects of the present invention are respectively pitched. A plurality of sub-codebooks corresponding to the cycle and sub-codebook switching means for switching the sub-codebook according to the input pitch cycle are provided.

According to the sixth and seventh aspects of the present invention, the voice encoding / decoding apparatus is configured to transmit only the sound source position of the head frame of voiced sound in the unvoiced frame immediately before it. In the subsequent frame, the same process of repeating the sound source while interpolating the pitch length in the encoding unit and the decoding unit is performed, so it is not necessary to transmit the sound source position for each frame. That is, in the case of the head frame of voiced sound in the encoding unit, the excitation start position is extracted from the input voice and is output to the decoding unit in the preceding unvoiced frame. A first decoding means for generating decoded speech by synchronizing the first quantized excitation signal with the excitation start position input by the first decoding means is provided.

[0017]

In the invention described in claim 1, for example, G
Since encoding is performed by using different schemes, ARMA and CELP, which gives better results, the encoding quality is improved as compared with either scheme.

According to the second aspect of the present invention, the coding means selecting means selects one of the first code searching means and the second code searching means that reduces distortion of the input voice and the synthesized voice.

According to the third aspect of the present invention, the second code searching means selects the input speech of the first quantized excitation signal and the second quantized excitation signal held in the adaptive excitation codebook. A signal that reduces the distortion of the synthesized speech is selected as the adaptive sound source signal.

Further, in the present invention as defined in claims 4 and 5, the spectral codebook, the excitation model codebook, and the driving excitation codebook switch their own subcodebooks in accordance with the pitch period.

According to the sixth and seventh aspects of the invention, when transmitting the sound source start position in the leading frame of the voiced sound to the decoding unit, the unvoiced frame preceding the voiced frame is transmitted.

[0022]

【Example】

Example 1. 1 and 2 are block diagrams of an embodiment of a speech encoding / decoding apparatus according to the present invention, and the operation of the present invention will be described below with reference to this figure. 1 shows a coding unit and FIG. 2 shows a decoding unit. In FIGS. 1 and 2, the same parts as those in FIGS. 8 and 9 are designated by the same reference numerals, and the description thereof will be omitted. In the figure, 2 is a driving excitation codebook, 3 is a driving excitation codeword, 4 is a driving excitation generating means, 5 is a driving excitation signal, 6 is an adaptive excitation codebook, 7 is an adaptive excitation signal, and 8 is a first code search. Means, 9 is the first quantized excitation signal, 10 is the second quantized excitation signal, 13 is the second code search result, 1
6 is a decoding means selection means, 17 is an encoding means selection signal,
18 is the encoding result, 25 is the second code searching means, 28 is the first code searching result, 33 is the first decoding means, 38 is the encoding means selecting means, 39 is the encoding result, and 40 is adaptive. Excitation codebook, 41 is an adaptive excitation signal, 42 is a second quantized excitation signal, 43 is a second decoding means, 46 is a driving excitation codebook, 47 is a driving excitation codeword, 48 is a driving excitation generating means,
Reference numeral 49 is a driving sound source signal.

First, the encoding unit will be described. As shown in FIG. 10, the first code searching means 25 synchronizes the first quantized sound source signal 24 with the sound source start position 20 in the case of the head frame of the voiced sound, and the first quantized sound source signal and the AR.
Synthetic speech is generated using the codeword 12 and the MA codeword 27, and in the other voiced frames, the first quantized sound source signal 24 is arranged subsequent to the synthetic speech obtained in the preceding frame to generate synthetic speech. , A combination of the AR codeword 12, the MA codeword 27, and the excitation model codeword 22 that minimizes the distortion of the synthesized speech and the input speech 1 is searched, and the search result is set as a coding result 28, which is coded together with the distortion. The first quantized excitation signal 24 in the combination is output to the adaptive excitation codebook 6 as the first quantized excitation signal 9.

The driving excitation codebook 2 stores a plurality of typical one-pitch period signals and Gaussian noise signals among the LPC residual signals as driving excitation codewords. The drive excitation generator 4 generates a drive excitation signal 5 by repeating the drive excitation codeword 3 in the drive excitation codebook 2 at a pitch cycle 15. Hereinafter, the operation method performed by the first code searching means 25 will be referred to as the GARMA method.

The second code searching means 8 synthesizes the driving excitation signal 5 and the quantized excitation signal generated from the adaptive excitation signal 7 in the adaptive excitation codebook 6 and the spectrum codeword 12 in the spectrum codebook 11. A combination of the driving excitation codeword 3 and the spectrum codeword 12 that generates the voice and minimizes the distortion of the synthesized voice and the input voice 1 is searched, and the search result is set as the second code search result 13, and the code is added together with the distortion. The second quantized excitation signal 10 in this combination is output to the adaptive excitation codebook 6. Hereinafter, the method used by the second code search means 8 for encoding will be CELP.
I will call it the method.

The coding means selecting means 16 selects one of the first code search result 28 and the second code search result 13 with the smaller distortion as the coding result 18 in the frame,
An encoding means selection signal 17 indicating which is selected and the encoding result 18 are output.

FIG. 3 is a diagram for explaining the operation of the adaptive excitation codebook 6. The adaptive excitation codebook 6 stores the first quantized excitation signal 9 in the excitation storage means 50 when the first code searching means is selected in the frame, and when the second code searching means is selected. The second quantized excitation signal 10 is stored in the excitation storage means 50 and output as the adaptive excitation signal 7.

In the conventional adaptive excitation codebook 6, the quantized excitation signal is stored by the code search means, whereas in the present embodiment, the quantized excitation signal stored in the adaptive excitation codebook 6 is stored as the first. The one code searching means 25 selects the latest quantized sound source signal from the first quantized sound source signal 9 outputted by the second code searching means 8 and the second quantized sound source signal 10 outputted by the second code searching means 8. Then, it is stored in the sound source storage means 50.
Which is the latest quantized excitation signal can be determined by the coding means selection signal 17 output from the coding means selection means. Therefore, the adaptive excitation codebook 6 inputs the encoding means selection signal 17 and switches the switch to switch the first quantized excitation signal 9 and the second quantized excitation signal 10 to the excitation storage means 50. input.
In this way, the adaptive excitation codebook 6 can supply the most recently used quantized excitation signal to the second code searching means.

Next, the decoding unit 1b shown in FIG. 2 will be described. In the figure, driving excitation codebook 46 and adaptive excitation codebook 4
0 is the same as the driving excitation codebook 2 and the adaptive excitation codebook 6 in the encoding unit 1a, respectively. The decoding means selecting means 38 outputs the coding result 18 as it is as the coding result 39 to either the first decoding means 33 or the second decoding means 43 according to the coding means selection signal 17.

The first decoding means 33 uses the first quantized excitation signal 32 and the MA codeword 35 in the MA codebook 34 and the AR codeword 45 in the spectrum codebook 44 corresponding to the coding result 39. Then, the decoded speech 37 is generated, and the first quantized excitation signal 32 is directly output to the adaptive excitation codebook 40 as the first quantized excitation signal 36.

The driving sound source generating means 48 has a pitch period of 15
And a drive excitation signal 49 is generated from the drive excitation codeword 47 in the drive excitation codebook 46 corresponding to the encoding result 39. The second decoding means 43 is a spectrum codebook 4 corresponding to the quantized excitation signal generated from the adaptive excitation signal 41 and the driving excitation signal 49 in the adaptive excitation codebook 40 and the encoding result 39.
Generate a decoded speech 37 using the AR codeword 45 in 4;
The quantized excitation signal is output as a second quantized excitation signal 42 to the adaptive excitation codebook.

As described above, in this embodiment, the first coding search means performs coding using only the quantized excitation signal using the excitation model stored in the excitation model codebook 21. On the other hand, the second code search means uses the adaptive excitation codebook to perform encoding by using the difference from the quantized excitation signal encoded immediately before. In this way, by operating the two code search means with different encoding methods respectively and comparing the operation results,
A feature of this embodiment is that one with a smaller coding distortion is selected. That is, this embodiment uses the GARMA method and CE.
It is characterized by using two LP systems and selecting and outputting the better one of the coding results obtained from the two systems.

Example 2. In the first embodiment, the GA
Although the case where two methods of RMA method and CELP method are used for comparison and selection is shown, the two encoding methods are not limited to these methods, and other methods may be used. Alternatively, the same system may be used, or one of them may be improved or modified. Further, it is not limited to the combination of two methods, and may be a combination of three or more methods.

Example 3. FIG. 4 is a configuration diagram of the adaptive excitation codebook 6 in one embodiment of the speech encoding / decoding device according to the present invention. Hereinafter, the operation of the adaptive excitation codebook 6 will be described with reference to this figure. The same parts as those in FIG. 3 are denoted by the same reference numerals. The adaptive excitation codebook 6 stores the first quantized excitation signal 9 in the first excitation storage means 51 when the first code searching means is selected, and when the second code searching means is selected,
The second quantized excitation signal 10 is stored in the second excitation storage means 52. The switching means 53 is the first sound source storage means 51.
And the signals stored in the second sound source storage means 52 are output as the adaptive sound source signals 7, respectively. The second code search means 8 selects the adaptive sound source signal 7 that reduces the distortion between the synthetic speech and the input speech, and outputs the selection result included in the code search result 13.

That is, the second code searching means 8 outputs the selection signal 8a to the applied excitation codebook 6, and switches the excitation signals stored in the first excitation storage means 51 and the second excitation storage means. The first sound source storage means 51 stores the latest quantized sound source signal 9 output from the first code search means. The second sound source storage means 52 stores the latest quantized sound source signal output from the second code search means. The second code searching means 8 outputs the selection signal 8a and operates the code word switching means 53, so that the quantized excitation signal stored in the first excitation storage means 51 and the second excitation storage means 52. , Respectively, and try encoding based on both quantized excitation signals. As a result, the one with the smaller distortion is used for encoding and output as the code search result.

The adaptive excitation codebook 40 of the decoding unit is shown in FIG.
Is the same as the adaptive excitation codebook 6 in FIG. The second decoding means 43 of the decoding unit selects the adaptive excitation signal 41 from the adaptive excitation signal codebook 40 according to the encoding result 39.

Example 4. In the third embodiment, the case where there are two code search means is shown, but when there are three or more code search means, the excitation code is stored in the applicable excitation codebook 6 shown in FIG. Means exist corresponding to each code searching means, and the switching means 53 switches the quantized excitation signal stored in these three or more excitation storage means.

Example 5. FIG. 5 is a block diagram of the MA codebook 26 in one embodiment of the speech coding / decoding apparatus according to the present invention. The operation of the MA codebook 26 will be described below with reference to this figure. The same parts as those in FIG. 1 are denoted by the same reference numerals. M
The A codebook 26 has a plurality of subcodebooks 54, and the subcodebook switching means 55 selects one of the subcodebooks according to the input pitch period 15, and the MA code in the selected subcodebooks. Output word 27.

For example, in the figure, the sub codebook 1 is 15 ms.
The MA codebook corresponding to the pitch period of is stored. The sub-codebook 2 stores a codebook corresponding to a pitch period of 16 ms. The sub-codebook 3 stores a codebook corresponding to a pitch period of 17 ms. 15ms in this way
To the codebook of, for example, 25 ms, is stored, and the subcodebook switching means 55 can select the subcodebook based on the input pitch period 15. For example, pitch period 15
Is 16 ms, the sub codebook switching means 55
Selects sub-codebook 2 and outputs it as MA codeword 27. Note that the MA codebook 34 of the decoding unit is also the MA codebook 26.
It has the same configuration as. Further, the AR codebook, the excitation model codebook, or the driving excitation codebook can be configured in the same manner as in FIG.

As described above, the feature of this embodiment is that a plurality of codebooks are prepared in each codebook according to the pitch period. The voice has characteristics such as the difference between men and women, the speed of conversation, or the pitch of the voice, but whether the input voice is a man or a woman, or fast or slow, or high. Features such as low are often reflected in the pitch period. That is, there is a fact that the correspondence between the MA and the sound source model differs depending on the pitch period. The feature of this embodiment is that a plurality of MA codebooks and excitation model codebooks are prepared in accordance with the pitch cycle and quantization is performed using the codebook corresponding to the pitch cycle.

Example 6. 6 and 7 are block diagrams of an embodiment of the speech coding / decoding apparatus according to the present invention, and the operation of the present invention will be described below with reference to these figures. FIG. 6 shows the encoding unit, and FIG. 7 shows the decoding unit. The same parts as those in FIGS. 1 and 2 are designated by the same reference numerals and the description thereof will be omitted. When the frame is the head frame of voiced sound, the encoding unit transmits the sound source start position 20 to the decoding unit in the unvoiced frame preceding the frame. That is,
The sound source start position 20 is coded and transmitted before the head frame of voiced speech is coded and transmitted. Therefore, the transmission of the voiced frame is delayed by one frame.
Even if there is a delay in the transmission of frames, the decoding time in the decoding unit only shifts, and the decoding quality is not affected.
The sound source start position 20 may be encoded together with the encoding of the head frame of voiced sound. However, in this case, the coded information amount of the head frame of the voiced sound is reduced by the coding of the sound source start position 20.

In the decoding section, it is predetermined that the first decoding means 33 decodes the leading frame of the dominant sound, and when the frame is the leading frame of the voiced sound, the first decoding means 43. Generates a decoded speech 37 by arranging the first quantized excitation signal 32 from the input excitation start position 20.

This example will be described with reference to FIG. 17 showing a conventional example. In the conventional example, the decoding operation is performed from the beginning of the frame, but according to this embodiment, since the decoding starts from the sound source start position 20, it is decoded from time T1 shown in FIG. . Conventionally, since the decoding was performed from the time F1, the decoding between the time F1 and the time T1 was originally not necessary, but the synthesized voice was generated.
According to this embodiment, the sound source start position 20, that is, the time T
Since the decoding is started from 1, there is no synthesized voice in a portion that is not originally necessary. In addition, since the start decoding start positions coincide with each other, the pitch cycle deviation is eliminated, and the quality of decoded speech is improved. As shown in FIG. 17, since the pitch periods P1, P2, P3 of the input speech 1 are different from the pitch periods Q1, Q2, Q3 of the decoded speech 37, this deviation causes deterioration of the quality of the decoded speech. ing. In contrast,
According to this embodiment, since the decoded speech also starts from the sound source start position 20, the pitch period becomes the same as that of the input speech, which is useful for improving the quality of the decoded speech.

Example 7. Although the sound source start position 20 is input to the first decoding unit 33 in the sixth embodiment, the sound source start position 20 is input to the second decoding unit 4.
You may enter it in 3. In the case of the sixth embodiment, as described above, it is premised that the first encoding unit combines the head frame of the voiced sound, so the sound source start position 20 is input only to the first decoding unit 33. However, if it is unclear which of the first and second encoding means encodes the head frame of the voiced sound, the sound source start position 20 is set to the first and second decoding. By inputting to both of the decoding means, either of the first and second decoding means can start decoding using the sound source start position 20.

Example 8. In Examples 1 to 7,
When the first code searching means is selected in the encoding means selecting means, the first quantized excitation signal is stored in the adaptive excitation codebook, but the signal obtained by driving the MA filter with the first quantized excitation signal is stored. It is also possible to store.

Example 9. Although AR and MA are used as the spectral parameters in the first to eighth embodiments,
It is also possible to use other spectral parameters such as AR only, cepstrum, etc.

Example 10. In the fifth embodiment, a case has been shown in which a plurality of codebooks are prepared for each codebook used for the speech coding / decoding apparatus shown in FIG. 1 or FIG. The method of switching in cycles can also be used for the conventional speech coding / decoding apparatus shown in FIGS. 8 and 9. That is, the method of selecting the codebook according to the pitch period 15 is the first embodiment.
The method can be applied to other methods regardless of the method shown in FIG.

Example 11. In the above embodiment, the case where the voice is encoded and decoded as the voice encoding / decoding device has been described. However, the encoding device performing only the encoding or the decoding device performing only the decoding is described. Also in, it is possible to apply the encoding part and the decoding part of each embodiment.

Example 12 In the above embodiment,
Although the case of encoding and decoding a voice has been described, the voice in the present invention is not limited to a voice uttered by a human from the vocal tract and can be applied to a voice uttered by a non-human organism such as an animal or a beast. Is. Similarly, as long as it is input as sounds, not limited to voices produced by living things, it can be applied to the case of encoding and decoding these sounds. For example, a sound of a musical instrument, a fricative sound, or the like may be input and encoded and decoded. Further, the sound is not limited to being perceptible by a human being, and may be a sound such as an ultrasonic wave or a low sound wave that cannot be perceived by a human ear.

Example of evaluation experiment. FIG. 8 shows a configuration diagram of this evaluation experiment when the number of subframes is 2. MODE in the figure
0 means FVQ-GARMA system, MODE1 is C
It means ELP method. First, several sets of LSPs (AR parameters) are preselected for a frame. Next, for each LSP, a Mode with good SNRseg is selected for each subframe. Finally the SNR of the whole frame
LSP that maximizes seg and MOD in each subframe
A combination of E is selected.

FIG. 9 shows the internal structure of MODE0. M
ODE0 is similar to FVQ-GARMA, and selects one that maximizes SNRseg from all combinations of preselected vocal cord source wave model code and MA code.

FIG. 10 shows an internal configuration diagram of the MODE1.
In CELP of MODE1, pitch synchronization processing is performed in order to synchronize with MODE0. First, the adaptive codebook 1
Let the pitch length be a repetition vector P. If the previous subframe is MODE0, the adaptive codebook is a vocal cord source wave model driven by an MA filter. Next, one pitch length of the driving sound source codebook is set as a repeating vector C. Then, the ratio of the gains of the vectors P and C that maximizes SNRseg is determined. The short-term prediction residual signal was used for learning the driving sound source codebook.
Since the subframe length is long in the unvoiced frame, the subframe is divided into a plurality of parts and driven by white noise. No adaptive codebook is used. Twenty Japanese short sentences spoken by five men and women were used for learning the codebook. However, 10 sentences were used for learning the driving sound source codebook.
For the evaluation data, 10 sentences of Japanese short sentences other than the learning were used (5 sentences for each male and female, who were not used for learning, uttered one different sentence).

In order to investigate the performance of this system, a synthetic sound was created under the conditions shown in FIG. However, the peak position of the sound source was automatically extracted. In Fig. 12, the present system 2.4 Kbps (MGARM
A) SNRseg and DC (Cepstrum Dis)
tortion) by gender. The results show that males are inferior in both SNRseg and CD. This is mainly due to the fact that MODE1 is often selected in men with large pitch fluctuations, and because the sound source is repeated with the same pitch length without interpolating the pitch length in MODE1, the deviation of the sound source peak position becomes large. Considered to be

In order to examine the subjective quality of the synthetic speech of this method, synthetic speech was prepared under the conditions shown in FIG. 11 and a simple pair-comparison test was conducted by 6 test subjects. This system 2.4Kbps (MGARMA) that automatically extracts the sound source peak position and the conventional 2.4 that manually corrects the sound source peak position
Kbps FVQ-GARMA (MFVQ), the conventional 2.4 Kbps FVQ-GA that automatically extracts the sound source peak position
A comparison was made with RMA (AFVQ) and 4.8 Kbps CELP Basic Method (CELP). The test results are shown in FIG.

This method is better than AFVQ, and it was confirmed that the robustness against the extraction error of the sound source peak position was improved. In the comparison between this method and MFVQ, a reverberation was partially felt in sentences where this method was not selected. It has been confirmed that the present method is more stable than the MFVQ in terms of variations in sound quality. 4.8 if it reaches the same level as MFVQ by improving the sound source peak position extraction and interpolation
It can be expected that the same quality as KCELP can be obtained.

[0056]

As described above, according to the first aspect of the present invention, a better method is selected from the two kinds of encoding methods, and therefore, a better code than the case of one kind of encoding method is selected. Can be converted.

According to the second aspect of the present invention, the encoding means selecting means selects the one of the second code searching means and the first code searching means that reduces the distortion of the synthesized speech and the input speech. The quality of the decoded speech is improved for speakers with poor model fit.

The adaptive excitation codebook according to the third aspect of the present invention stores the first quantized excitation signal and the excitation signal as the adaptive excitation signal, and the second code searching means reduces distortion of the synthesized speech and the input speech. Since one of them is selected and used, the quality of decoded speech is improved.

In the speech coding / decoding apparatus according to the present invention, the spectrum codebook, the excitation model codebook, and the driving excitation codebook of the coding unit and the decoding unit correspond to the pitch period. Since a plurality of created sub-codebooks are provided and these sub-codebooks are used by switching according to the pitch period, the quality of decoded speech is improved.

Further, in the speech coding / decoding apparatus according to the sixth and seventh aspects of the invention, the sound source start position in the leading frame of the voiced sound is output to the decoding unit as an unvoiced frame preceding the voiced frame, and decoding is performed. Since the first decoding means of the encoding unit generates the decoded voice by synchronizing the first quantized excitation signal with the excitation start position, it is the same as the synthesized voice of the encoding unit without increasing the transmission amount of voiced frames. To generate the decoded sound of and the quality of the decoded sound is improved.

[Brief description of drawings]

FIG. 1 is a configuration diagram showing a speech coding / decoding device according to a first embodiment of the present invention.

FIG. 2 is a configuration diagram showing a speech encoding / decoding device according to the first embodiment of the present invention.

FIG. 3 is a configuration diagram showing an adaptive excitation codebook according to the first embodiment of the present invention.

FIG. 4 is a configuration diagram showing an adaptive excitation codebook according to a third embodiment of the present invention.

FIG. 5 is a configuration diagram showing an MA codebook according to a fifth embodiment of the present invention.

[Fig. 6] Fig. 6 is a configuration diagram showing a speech encoding / decoding device of embodiment 6 of the present invention.

FIG. 7 is a configuration diagram showing a speech encoding / decoding device according to embodiment 6 of the present invention.

FIG. 8 is a block diagram showing a speech encoding / decoding device of an evaluation experiment based on the present invention.

FIG. 9 is a configuration diagram showing a speech encoding / decoding device of an evaluation experiment based on the present invention.

FIG. 10 is a configuration diagram showing a speech encoding / decoding device of an evaluation experiment based on the present invention.

FIG. 11 is a diagram showing conditions of an evaluation experiment based on the present invention.

FIG. 12 is a diagram showing SNRseg and CD in an evaluation experiment based on the present invention.

FIG. 13 is a diagram showing a result of an evaluation experiment based on the present invention.

FIG. 14 is a configuration diagram showing a conventional speech encoding / decoding device.

FIG. 15 is a configuration diagram showing a conventional speech encoding / decoding device.

[Fig. 16] Fig. 16 is a diagram for explaining the operation of the code search means of the conventional speech encoding / decoding device.

FIG. 17 is a diagram for explaining the operation of the decoding means of the conventional speech encoding / decoding device.

[Fig. 18] Fig. 18 is a diagram explaining a problem of a conventional speech encoding / decoding device.

[Explanation of symbols]

 1 Input Speech 2 Drive Excitation Codebook 3 Drive Excitation Codeword 4 Drive Excitation Generator 5 Drive Excitation Signal 6 Adaptive Excitation Codebook 7 Adaptive Excitation Signal 8 Second Code Search Means 9 First Quantized Excitation Signal 10 Second Quantized excitation signal 11 AR codebook 12 AR codeword 13 Second code search result 14 Pitch cycle extraction means 15 Pitch cycle 16 Coding means selection means 17 Coding means selection signal 18 Coding result 19 Excitation start position extraction means 20 Excitation start position 21 Excitation model codebook 22 Excitation model codeword 23 Excitation model generation means 24 First quantized excitation signal 25 First code search means 26 MA codebook 27 MA codeword 28 First code search result 29 Excitation Model codebook 30 Excitation model codeword 31 Excitation model generation means 32 First quantized excitation signal 33 First decoding means 34 MA codebook 35 MA codeword 36 First quantized excitation signal 37 Decoded speech 38 Decoding means selection means 39 Coding result 40 Adaptive excitation codebook 41 Adaptive excitation signal 42 Second quantized excitation signal 43 Second decoding Conversion means 44 AR codebook 45 AR codeword 46 driving excitation codebook 47 driving excitation codeword 48 driving excitation generation means 49 driving excitation signal 50 excitation source storage means 51 first excitation storage means 52 second excitation storage means 53 switching means 54 Sub-codebook 55 Sub-codebook switching means 56 Encoding result 57 Code search means 58 Decoding means 59 Quantized excitation signal 60 Quantized excitation signal

Claims (7)

[Claims] 1. A pitch cycle for extracting a pitch cycle from an input signal in a coding apparatus for searching a model that minimizes distortion of an input signal and a synthesized signal and coding the input signal using the search result. Extracting means; first encoding means for encoding one input signal by selecting one model suitable for the input signal from a plurality of models for the pitch period extracted by the pitch period extracting means; A second encoding unit that encodes an input signal by a method different from that of the first encoding unit is compared with the encoding results of the first encoding unit and the second encoding unit, and the comparison is made. An encoding apparatus comprising: a selection unit that selects and outputs one of the encoding results of the first encoding unit and the second encoding unit based on the result. 2. A speech coding / decoding apparatus which separates an input speech into spectral parameters of a vocal tract and a sound source signal and performs coding / decoding for each frame of a fixed time length, wherein a pitch is provided from the input speech to a coding unit. Pitch cycle extraction means for extracting a cycle and outputting it to the sound source model generation means and the driving sound source generation means, a spectrum codebook in which a plurality of typical spectrum parameters are stored as spectrum codewords, and a sound source representing a sound source signal of one pitch cycle Excitation model codebook that stores a plurality of typical model parameters as excitation model codewords, and excitation signal of one pitch cycle generated from excitation model codewords in the excitation model codebook is repeated in the pitch cycle. And a first quantized excitation signal as a first quantized excitation signal to the first code search means. And a search for a combination of a sound source model code word and a spectrum code word that minimizes distortion of the synthesized speech and the input speech generated from the spectrum code word in the spectrum codebook, and the search result as the first code search result at that time. The first code search means that outputs the first quantized excitation signal in this combination to the adaptive excitation codebook and the first code search means in the preceding frame. The adaptive excitation codebook in which the first quantized excitation signal or the second quantized excitation signal output by the second code search means is stored as an adaptive excitation signal, and a plurality of signals prepared in advance as driving excitation codewords Driving excitation codebooks each of which is stored, and driving excitation generation means for generating a driving excitation signal in which the driving excitation codewords in the driving excitation codebook are repeated at the pitch cycle. A second quantized excitation signal is generated from the adaptive excitation signal and the driving excitation signal, and a spectrum codeword that minimizes spectral distortion in the frame is selected from the spectrum codebook. Generates synthesized speech using the second quantized excitation signal, searches for a driving excitation codeword that minimizes distortion between the synthesized speech and the input speech, and codes the search result as the second code search result along with the distortion at that time. Output to the conversion means selection means, the second code search means for outputting the second quantized excitation signal to the adaptive excitation codebook, among the first code search result and the second code search result, Encoding means selecting means for selecting and outputting one having a smaller distortion as an encoding result of the frame and outputting which code search result is selected as an encoding means selecting signal. In the decoding unit, the same spectrum codebook as the coding unit, the same excitation model codebook as the coding unit, and the excitation model in the excitation model codebook corresponding to the coding result input from the coding unit. An excitation model generation unit that is the same as the encoding unit that generates and outputs the first quantized excitation signal from the codeword and the pitch period input from the encoding unit, the same adaptive excitation codebook as the encoding unit, and the encoding The same driving excitation codebook as the unit, generating decoded speech using the first quantized excitation signal and the spectrum codeword in the spectrum codebook corresponding to the coding result input from the coding unit, First decoding means for outputting the first quantized excitation signal to an adaptive excitation codebook, a driving excitation codeword in the driving excitation codebook corresponding to the encoding result, and a driving excitation signal from the pitch period Drive sound source generation A second quantized excitation signal generated from the adaptive excitation signal in the adaptive excitation codebook, the driving excitation signal output from the driving excitation generation means, and the decoded speech from the spectrum codeword; The second decoding means for outputting the quantized excitation signal to the adaptive excitation codebook and the decoded speech of the first decoding means and the second decoding means according to the coding means selection signal input from the coding section. A speech coding / decoding apparatus comprising: a decoding means selecting means for selecting decoded speech. 3. In the above speech coding / decoding apparatus, a first quantized excitation signal when the first code searching means is selected is stored in the adaptive excitation codebooks of the encoding unit and the decoding unit. First excitation storage means, second excitation storage means for storing the second quantized excitation signal when the second code search means is selected, and first quantized excitation signal as an adaptive excitation signal And a switching means for switching and outputting the second quantized excitation signal, and the encoding section generates a second quantized excitation signal from the adaptive excitation signal and the driving excitation signal, and the second quantized excitation signal. A synthetic speech is generated using the signal and the AR code word, an adaptive sound source signal that reduces distortion of the synthetic speech and the input speech is selected, and the selection result is included in the second code search result and output. It is equipped with a second code search means, and enters the decoding unit from the coding unit. A second decoding means adapted to select and use the first quantized excitation signal or the second quantized excitation signal in the adaptive excitation codebook as the adaptive excitation signal according to the applied encoding result. The speech coding / decoding apparatus according to claim 2, characterized in that: 4. In a speech coding / decoding apparatus which separates an input speech into spectral parameters of a vocal tract and a sound source signal and performs coding / decoding for each frame of a fixed time length, at least a typical coding unit is used. A spectrum codebook in which a plurality of spectrum parameters are stored as spectrum codewords, and a excitation model codebook in which a plurality of typical parameters of excitation models representing excitation signals of one pitch period are stored as excitation model codewords, Any one of a driving excitation codebook in which a plurality of prepared signals are stored as driving excitation codewords, and at least one of a spectrum codebook, an excitation model codebook, or a driving excitation codebook provided in the encoding unit. , A plurality of sub-codebooks corresponding to the pitch cycle and an encoding / decoding process according to the pitch cycle input from the pitch cycle extraction means. A sub-codebook switching unit for switching the sub-codebook to be used is provided, and the decoding unit has the same spectrum codebook or excitation model code as the coding unit for switching the sub-codebook according to the pitch period input from the decoding unit. A voice encoding / decoding device comprising a book or a driving excitation codebook. 5. A speech coder / decoder that separates an input speech into spectrum parameters and excitation signals and encodes each frame with a fixed time length, wherein an encoding unit extracts a pitch period from the input speech and the excitation source. Model generation means, driving excitation generation means, spectrum codebook, pitch period extraction means for outputting to the excitation model codebook, spectrum codebook storing a plurality of typical spectrum parameters as spectrum codewords, and excitation source with one pitch period An excitation model codebook storing a plurality of typical excitation model parameters representing signals as excitation model codewords, and a plurality of subcodes corresponding to pitch periods in the spectrum codebook or the excitation model codebook And the sub-codebook used for encoding / decoding processing according to the pitch cycle input from the pitch cycle extraction means. And a sub-codebook switching means for outputting the excitation signal of one pitch cycle generated from the excitation model codeword in the excitation model codebook repeated in the pitch cycle to the code search means as a quantized excitation signal. Model generation means, searching for a combination of the excitation model codeword and the spectrum codeword that minimizes distortion of the synthesized speech and the input speech generated from the quantized excitation signal and the spectrum codeword in the spectrum codebook,
Code search means for outputting the search result as a code search result, a decoding unit, the same spectrum codebook as the coding unit, the same excitation model codebook as the coding unit, and the coding input from the coding unit. An excitation model generation unit that is the same as the encoding unit that generates a quantized excitation signal from the excitation model codeword in the excitation model codebook corresponding to the result and the pitch period input from the encoding unit, and outputs the quantization excitation signal to the decoding unit; , A decoded speech is generated using a spectrum codeword in the spectrum codebook corresponding to the quantized excitation signal and a coding result input from a coding unit, and the quantized excitation signal is stored in the adaptive excitation codebook. A voice encoding / decoding device comprising a decoding means for outputting. 6. In the speech coding / decoding apparatus, when the frame is a frame in which the unvoiced frame is changed to a voiced frame for the first time, the coding unit extracts from the input speech the position at which the voiced sound starts as the sound source start position. Then, in the unvoiced frame preceding the frame, the excitation start position extraction means for outputting to the decoding unit, and the decoding unit synchronizes the first quantized excitation signal with the excitation start position input from the encoding unit. The speech coding / decoding apparatus according to claim 2, further comprising a first decoding unit configured to generate decoded speech. 7. A speech coder / decoder that separates an input speech into spectrum parameters and excitation signals and encodes each frame of a fixed time length, wherein an encoding unit extracts a pitch period from the input speech and generates an excitation. Pitch cycle extraction means for outputting to the model generation means and the driving excitation generation means, a spectrum codebook in which a plurality of typical spectrum parameters are stored as spectrum codewords, and a typical excitation model parameter representing a one pitch cycle excitation signal Excitation model codebook in which a plurality of typical excitation model codewords are stored, and one pitch period excitation signal generated from the excitation model codeword in the excitation model codebook is quantized by repeating the excitation signal in the pitch period. Excitation model generating means for outputting to the code search means as an excitation signal, the quantized excitation signal and the spectrum in the spectrum codebook. A code search means for searching a combination of a sound source model code word and a spectrum code word that minimizes distortion of synthesized speech and input speech generated from a Koutor code word, and outputting the search result as a code search result, and the frame is unvoiced. In the case of a frame that has changed from a frame to a voiced frame for the first time, a sound source start position extraction unit that extracts the position where the voiced sound starts as a sound source start position from the input voice, and outputs it to the decoding unit in the unvoiced frame preceding the frame, In the decoding unit, the same spectrum codebook as the encoding unit, the same excitation model codebook as the encoding unit, and the excitation model codeword in the excitation model codebook corresponding to the encoding result input from the encoding unit. And the same excitation model generation unit as the encoding unit that generates a quantized excitation signal from the pitch period input from the encoding unit and outputs it to the decoding unit If the frame is a frame that has changed from a voiceless frame to a voiced frame for the first time, the quantized excitation signal synchronized with the excitation start position input from the encoding unit and the encoding result input from the encoding unit A speech coding / decoding apparatus comprising: decoding means for generating decoded speech using a spectrum codeword in a spectrum codebook and outputting the quantized excitation signal to the adaptive excitation codebook.