JPS58198095A - Line spectrum type voice analyzer/synthesizer - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a line spectrum type speech analysis and synthesis device, and particularly to a line spectrum type voice analysis and synthesis device, which is characterized in that it reduces quantization distortion of line spectrum frequencies and improves the approximation of the spectrum envelope of the voice. Regarding analysis and synthesis equipment.

FIG. 1 shows a block diagram of an example of a conventional line spectrum analysis and synthesis apparatus. As is clear in the figure, the speech analysis side 51 includes LSP parameter extraction means including a linear prediction analyzer 1 and an LSP parameter calculator 2;
Parameter encoder 3, sound source parameter analyzer 4. The speech synthesis side 52 includes a sound source information encoder 5.
LSP parameter decoder7. Parameter calculator 8° Sound source information decoder 9. Excitation signal generator 10. A speech synthesis filter 11 is provided.

The analysis side 51 analyzes and extracts linear prediction coefficients from the audio input signal and sends them to the L S 'P parameter calculator 2 . The linear prediction coefficient is α parameter α1 (i=1
~n1 n is a predetermined positive integer), and the LAP parameter calculator 2 calculates the line spectrum frequency ω1 (i=1 to n, n is a predetermined positive integer) from this α parameter. This (ωi) is generally called an LSP parameter. This (ωi) is sent to the LSP parameter encoder 3, where it is linearly quantized and then transmitted to the speech synthesis side 52 via the data transmission means 6 as analysis data related to the spectral envelope of the speech. Further, the sound source parameter analyzer 4 extracts sound source information including the pitch frequency, voiced/unvoiced discrimination signal, and short-time average power from the input signal, and outputs the extracted sound source information to the sound source information encoder 5. The sound source information encoder 5 quantizes the sound source information and transmits it to the speech synthesis side 52 via the data transmission means 6.

On the speech synthesis side 52, the analysis data related to the spectral envelope of the speech is analyzed using the LSP encoded by the LSP parameter encoder 3 on the speech analysis side 51.
The SP parameter is decoded in the LSP parameter decoder 7, and the linear prediction coefficient α parameter (ωi) of the same order is calculated from the LSP parameter (ωi) in the K parameter calculator 8. From this α parameter, the linear prediction coefficient is further calculated. The coefficients of the speech synthesis filter 11 are controlled by extracting a parameter (Ki) in the form of a speech synthesis filter. On the other hand, the sound source information quantized by the sound source information encoder 5 is input to the sound source information decoder 9 and decoded, and the excitation signal generator 1
Output to 0. The excitation signal generator 10 inputs the sound source information, generates a filter excitation signal, and supplies it to the speech synthesis filter 11.

The voice synthesis filter 11 synthesizes voice using the parameters inputted from the parameter calculator 8 and the filter excitation signal supplied from the vibration signal generator 10, and outputs the voice to the terminal 92.
Output from

FIG. 5 shows an example of a line spectrum frequency pattern in an example of this conventional line spectrum type speech analysis and synthesis apparatus.

In the figure, the vertical axis represents time and the horizontal axis represents line spectrum frequency. Points 10.1 to 108 show the distribution of line spectrum frequencies at time t1. The number of line spectral frequencies in this example is eight. Similarly, t2 to t
The distribution of line spectrum frequencies at each time s is represented by points 201 to 508.

In the figure, line spectrum frequency 30 at time t3
Focusing on 1,302, it can be seen that the values are extremely close. Line spectrum frequency 40 at time t4
The same applies to 1,402. On the other hand, line spectrum frequencies 103, 104 and 104, 1 at time 11
05 etc. are linearly quantized by the LSP parameter encoder 3 in the ratio, but the line spectrum frequency 301
In frequency regions where adjacent frequency intervals are extremely narrow, such as in the case of ゜302 or 401.402, the amount of information related to the spectral envelope of the voice contained in those line spectral frequencies is distributed with high density. On the other hand, in contrast, the line spectral frequency 103.1
In frequency regions where adjacent frequency intervals are large, such as in the case of 04 or 104°105, the density of the amount of information related to the spectral envelope of the voice contained in those line spectral frequencies is relatively low. Therefore, in data transmission using the conventional linear quantization method in the line spectral frequency domain, quantization is uniformly performed regardless of the information density distribution of the analysis data assigned to the line spectral frequency, and as a result, the quantization This causes a problem in that the quantization distortion increases, important information related to the degree of approximation of the audio spectrum envelope is partially lost, and the quantization efficiency decreases. Therefore, when the voice is finally reproduced and extracted on the voice synthesis side 52, the quality of the voice synthesized by the voice synthesis filter 11 controlled by the parameter (Ki) also deteriorates.

That is, in the conventional line spectrum type speech analysis and synthesis device, the LSP extracted from the linear prediction coefficients is
Since the parameter (ωi) is encoded by linear quantization and the data is transmitted to the synthesis side, quantization distortion in the narrow frequency interval region of adjacent line spectral frequencies is severe, and the voice This has the drawback of deteriorating the degree of approximation of the spectral envelope and reducing the quantization efficiency.

It is an object of the present invention to eliminate the above-mentioned drawbacks and instead of linearly quantizing line spectral frequencies, the frequency differences between one particular line spectral frequency and all adjacent line spectral frequencies are quantized using a non-linear quantization technique. It is an object of the present invention to provide a line spectrum type speech analysis and synthesis device that performs encoding, has a small amount of quantization, and has high quantization efficiency.

The line spectrum type speech analysis and synthesis apparatus of the present invention analyzes and synthesizes speech by calculating a line spectrum frequency from an input signal and linear prediction coefficients obtained by multilinear prediction analysis. Line spectrum parameter extraction means for analyzing and extracting a line spectrum parameter including a specific line spectrum frequency (ω to n-1) among the line spectrum frequencies from the signal;
The signal related to this linear spectrum parameter is
・Equipped with an encoding means for shape quantization on the speech analysis side,
a decoding means for decoding the coded signal non-linearly quantized by the coding means; and a decoding means for extracting multilinear prediction coefficients from the decoded signal related to the line spectrum parameters to control a speech synthesis filter and generate a synthesized waveform. The speech synthesis device is equipped with a speech synthesis means for outputting the speech.

The present invention will be described in detail below with reference to the drawings.

FIG. 2 is a block diagram showing a first embodiment of the line spectrum type speech analysis and synthesis apparatus of the present invention.

In this embodiment, the speech analysis side 61 uses the linear prediction analyzer 12 and the line spectrum parameters (ω, Δω
i) Pure spectral parameter extraction means including a calculator 13, line spectral parameters (ωk.

Δωi) encoder 14. a sound source parameter analyzer 15 and a sound source information encoder 16, and a line spectrum parameter (ω, Δωi) decoder 18 on the speech synthesis side 62.
．． a parameter calculator 19, a sound source information encoder 20. Excitation signal generator 21. A voice synthesis filter 22 is provided.

On the analysis side 61, an audio input signal is input from a terminal 93 or 9-. The linear prediction analyzer 12 analyzes and extracts coefficients related to waveform reproducibility such as autocorrelation coefficients and covariance coefficients from the input signal, and calculates the spectral parameters (ω, ΔωI
) is output to the calculator 13. Line spectrum parameter (ω
, Δωi) The calculator 13 calculates the multilinear prediction coefficient (αi) by solving a normal equation from the coefficients related to the waveform reproducibility.
That is, the α parameter is determined, and from this α parameter, the predetermined line spectrum frequency ωk (k is a predetermined positive integer) and the frequency difference Δω between all adjacent line spectrum frequencies are calculated.
1 (i-1 to n-1, n is a predetermined positive integer). The calculated ω and Δωi are sent to the line spectrum parameter (ω, Δωi) encoder 14, where they are encoded using a nonlinear quantization technique. As explained using Fig. 5, in the line spectral frequency region where the frequency difference Δω1 (i==1 to n-1) between adjacent line spectral frequencies is small, the analysis data of the speech spectral envelope is distributed with high density. In contrast, the 10-distribution density of the analytical data is low in the line spectral frequency region where Δωi is large. Therefore, in order to maintain the degree of approximation of the analysis data of the voice spectral envelope, it is necessary to divide the line spectrum frequency region with small Δωi and the line spectrum frequency region with large Δωi to correspond to the distribution density of the analysis data. It is necessary to faithfully transmit the lever Δωi to the speech synthesis side 62. For the analytical data of the audio spectrum envelope, we experimentally calculated the frequency difference Δω between adjacent line spectrum frequencies.
Since the statistical distribution data in the line spectrum frequency band of i is known, the quantization precision for the frequency difference ΔωIK is controlled with reference to this distribution data. In other words, Δωi in the line spectrum frequency domain where the frequency difference is small is quantized with high precision, while Δωi in the line spectrum frequency domain where the circular wave number difference is large is quantized with relatively lower precision. . The feature of this line spectrum parameter (ω, Δωi) encoder is that it nonlinearly quantizes a specific line spectrum frequency ω and the frequency difference Δωi, which is the main requirement of the present invention. Line spectrum parameters (ω, Δωi) encoder 1
The nonlinearly quantized ω and Δωi according to 4 are transmitted to the speech synthesis side 62 via the data transmission line 17 as analysis data related to the spectral envelope of the speech. Further, the sound source parameter analyzer 15 extracts sound source information including the pitch frequency, voiced/unvoiced discrimination signal, and short-time average power from the input signal, and outputs it to the sound source information encoder 16 . The sound source information encoder 16 quantizes the sound source information and transmits it to the speech synthesis side 62 via the data transmission means 17.

The encoded signal of the predetermined line spectrum frequency ω and the frequency difference Δωi transmitted through the stage 17 is decoded by the line spectrum parameter (ω, Δωi) decoder 18 and output to the K parameter calculator 19. do. In the K parameter calculator 19, the linear prediction coefficient α parameter is calculated from the above-mentioned ω and Δωi, and the α parameter of ! 7, and the coefficients of the speech synthesis filter 22 are controlled by the parameters. On the other hand, regarding the sound source information, the coded signal of the sound source information transmitted via the transmission line 17 is decoded by the sound source information decoder 20 and output to the excitation signal generator 21 .

The excitation signal generator 21 inputs the sound source information, generates a filter excitation signal, and supplies it to the speech synthesis filter 22 . The voice synthesis filter 22 synthesizes voice using the parameters inputted to the parameter calculator 19 and the filter excitation signal supplied from the excitation signal generator 21, and outputs the voice to the terminal 94.
Output from

Next, a second embodiment of the present invention will be described with reference to FIG. In the figure, the analysis side 71 is the same as in the first embodiment. On the synthesis side, the analysis data related to the spectral envelope of the audio of ω and Δωi encoded by the nonlinear quantization method and the encoded signal of the sound source information are sent to the synthesis side 7 via the data transmission means 28.
2, among which ω and Δωi are decoded by a line spectrum parameter (ω, Δωi) decoder 29 and output to an α parameter calculator 30. α par 1
3- In the parameter calculator 30, for said ω and Δωi
A linear prediction coefficient α parameter is calculated from the α parameter, and the number of speech synthesis filters 33tvi4 is controlled by this α parameter. On the other hand, regarding the sound source information, the encoded signal of the sound source information transmitted from the analysis side 71 is transmitted to the sound information decoder 3.
1 and output to the excitation signal generator 32. The excitation signal generator 32 inputs the sound source information, generates an inverter excitation signal, and supplies it to the speech synthesis filter 33. The voice synthesis filter 33 synthesizes voice using the α parameters inputted from the α parameter calculator 30 and the filter excitation signal supplied from the excitation signal generator 32, and outputs the synthesized voice from the terminal 96.

Next, a third embodiment of the present invention will be described with reference to FIG. In the figure, the analysis side 81 is the same as in the first and second embodiments. On the synthesis side, the analysis data related to the spectral envelope of the voice of ω and Δωl encoded by the non-linear quantization method and the encoded signal of the sound source information are sent via the data transmission means 39 to the synthesis side 14- 82, among which ω and Δωi are decoded by the line spectrum parameter (ω, Δωl) decoder 40 and output to the LSP parameter calculator 41. L.S.
In the P parameter calculator 41, the above ω and Δ
A linear prediction coefficient α parameter is calculated from ωi, the order of the calculated α parameter is converted as necessary, and an LSP parameter is calculated from the order-converted α parameter. Controls the order of 44. On the other hand, regarding the sound source information, the encoded signal of the sound source information transmitted from the analysis side 81 is decoded by the sound source information decoder 42 and output to the excitation signal generator 43. The excitation signal generator 43 inputs the sound source information, generates a filter excitation signal, and supplies it to the speech synthesis filter 44 .

The speech synthesis filter 44 uses the LSP parameters input from the LAP parameter calculator 41 and the excitation signal generator 4.
The synthesized voice is synthesized with the filter excitation signal supplied from 3, and outputted from a terminal 98.

As explained in detail above, the line spectrum type speech analysis and synthesis device of the present invention uses lines encoded by the conventional linear quantization method as data related to linear prediction coefficients to be transmitted from the analysis side to the synthesis side. The spectral frequency (ωi), i.e. the frequency difference Δω between one given line spectral frequency ω and all adjacent line spectral frequencies encoded by a non-linear quantization technique in place of the LSP parameters.
The use of i has the effect of reducing quantization distortion and improving quantization efficiency.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of a conventional line spectrum type speech analysis and synthesis device, and FIGS. Second. FIG. 5, a block diagram of the third embodiment, is an explanatory diagram of the time distribution of line spectrum frequencies. In the figure, 1.12, 23.34...Linear prediction analyzer, 2
・・・・・・LSP parameter calculator, 3・・・・・・
LAP parameter encoder, 4, 15, 26. 37...
...Sound source parameter analyzer, 5, 16, 27.38
...Sound source information encoder, 6.17, 28.39
...... Data transmission means, 7... LAP parameter decoder, 8.19... K parameter calculator, 9, 20, 31° 42... Sound source information decoder, 10, 21, 32° 43...Excitation signal generator, 11, 22, 33.44...Speech synthesis filter, 13, 24.35... ...Line spectrum parameter (ω, Δωi) calculator, 14°25.36
...Line spectrum parameter (ω, ΔωD encoder, 18, 29.40...Line spectrum parameter (ω, Δωi) decoder, 30...
α parameter calculator, 41... LAP parameter calculator, 51, 61, 71.81... Analysis side, 52°62.72.82... Synthesis side, 91 ~
98... Edge 17- Rumor! Clause 6 ゝ← N Odor 0 9 Procedural Amendment - Issued by Mr. Commissioner of the Patent Office ■, Indication of the case 1982 Patent Application No. 08096
No. 3 No. 2, Title of the invention: Line spectrum type speech analysis and synthesis device 3, Relationship to the amended case Applicant: 5-33-1 Shiba, Minato-ku, Tokyo (423) B Hon Electric Co., Ltd. Representative: Tadahiro Sekimoto 4. Agent 5, insert the following sentence between lines 19 and 20 on page 15 of the contents of the amendment in column 6 of the "Detailed Description of the Invention" of the specification to be amended. [Note that the accumulation of quantization distortion caused by sequentially quantizing and decoding frequency differences between line spectral frequencies can be easily removed by the following method. That is, Δω is first quantized. Let δ be the distortion generated as a result of quantizing Δω1. Next, instead of quantizing Δω, (Δω2−δ1) is quantized. Let δ be the distortion generated as a result of quantizing (Δω, −δ,). Thereafter, quantization is similarly performed up to (Δωn-1−δn-2). ”

Claims (1)

[Scope of Claims] (1) In a line spectrum type analysis and synthesis device for analyzing and synthesizing speech by extracting line spectrum frequencies from linear prediction coefficients obtained from an input signal by linear prediction analysis, Line spectrum parameter extraction means for analyzing and extracting line spectrum parameters including a specific line spectrum frequency and frequency differences between all adjacent line spectrum frequencies; and nonlinear quantization of signals related to the line spectrum parameters. a decoding means for decoding the encoded signal that has been nonlinearly quantized by the encoding means;
A line spectrum type voice characterized in that the voice synthesis side is equipped with a voice synthesis means for extracting a linear prediction coefficient from the decoded signal related to the line spectrum parameter, controlling a voice synthesis filter, and outputting a synthesized voice waveform. Analysis and synthesis equipment. (2. In the device described in claim (1),
A line spectrum type speech analysis supplementary device characterized in that a parameter is used as the linear prediction coefficient for controlling the speech synthesis filter included in the speech synthesis means. (3) Range of % allowable rust In the equipment described in paragraph (1),
A line spectrum type speech analysis and synthesis device, characterized in that an α parameter is used as the linear prediction coefficient for controlling the speech synthesis filter included in the speech synthesis means. (4) In the device according to claim (1),
A line spectrum type speech analysis and synthesis device, characterized in that an L8P parameter is used as the linear prediction coefficient for controlling the speech synthesis filter included in the speech synthesis means.