JPH0236000B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a line spectral vocoder, and particularly to improvements in quantization characteristics and interpolation characteristics of line spectral coefficients.

A vocoder is used for the purpose of converting an audio signal into a low-speed digital code string, and analyzes the audio signal at regular intervals to extract spectral parameters that express vocal tract transmission characteristics and sound source information parameters that express vocal cord vibration, etc. The method extracts voice feature parameters consisting of
Information compression is achieved by taking advantage of the fact that the number of bits required to practically quantize voice feature parameters without distortion is relatively small.

Since it is desirable for a vocoder to transmit voice feature parameters at as long a time interval as possible, interpolation of voice feature parameters is generally performed on the synthesis side.
FIGS. 1a and 1b are waveform diagrams for explaining the effect of interpolation. a indicates distortion of the spectrum in the time axis direction when no interpolation is performed, and b indicates distortion of the spectrum in the time axis direction when interpolation is performed.

At a, a curve 101 shows the change in the actual audio spectrum. Points 102 to 107 indicate spectra corresponding to voice feature parameters analyzed at regular intervals. The stepped line segment group 108 is a reproduced speech spectrum when the speech feature parameters are held for the analysis cycle time length. The area of the portion (hatched portion) encompassed by the curve 101 and the group of stepped line segments 108 expresses the magnitude of distortion in the time axis direction when no interpolation is performed.

In b, curve 101 and points 102 to 107 are a
This is the same as the explanation. A broken line 109 is a reproduced audio spectrum when ideal interpolation is performed on the audio characteristic parameters in the sense described later. curve 1
The part surrounded by 01 and analysis line 109 (shaded part)
The area of represents the magnitude of distortion in the time axis direction when interpolation is performed. From FIGS. 1a and 1b, if ideal interpolation is performed, the distortion in the time axis direction can be reduced to a large extent compared to the case where no interpolation is performed.

Next, the above-mentioned ideal interpolation will be explained. FIG. 2 is a spectrum diagram for explaining ideal interpolation. In the figure, a solid line 201 is the power spectrum envelope at time _t1 , and a solid line 202 is the power spectrum envelope at time _t2 . Dotted line 203 is time
This is the power spectrum envelope obtained by ideal interpolation of t ₁ +t ₂ /2. The dotted line 203 is calculated by the following procedure. The power density at time t ₁ at frequency f ₁ is P ₁ (f ₁ ),
When the power density at time t ₂ is P ₂ (f ₁ ),
Let the logarithmic power density of t ₁ +t ₂ /2 be Log(P ₁ (f ₁ ))+Log(P ₂ (f ₁ )). The power spectrum envelope obtained by interpolation relatively well approximates the power spectrum envelope actually analyzed at time t ₁ +t ₂ /2 if (t ₂ −t ₁ ) is short to some extent.

Usually, the spectral envelope at each observation time is not directly expressed, but is expressed, for example, as a coefficient of a filter having the same frequency response as the spectral envelope. For example, α-parameters, K-parameters, line spectrum frequencies, and the like are known as methods for expressing the spectrum envelope using filter coefficients based on P-order (P is an integer) linear predictive analysis.
Normally, in a vocoder, parameters such as a line spectrum frequency that indirectly expresses the spectrum envelope are determined on the analysis side, and these parameters are interpolated on the synthesis side. In general, the spectral envelope corresponding to the parameter-level interpolation result does not match the ideal interpolation.

One object of the present invention is to easily obtain results close to ideal interpolation regarding interpolation of line spectral coefficients.

Next, quantization of feature parameters will be briefly described. For the purpose of a vocoder, it is necessary to quantize feature parameters with as few bits as possible, but if the number of bits in half-plane quantization is reduced, distortion of the spectral envelope that occurs as a result of quantization increases. This distortion can be alleviated to some extent by implementing so-called nonlinear quantization in which the quantization step size is varied in consideration of the distribution characteristics of the parameters, the fact that sensitivity to quantization varies depending on the position of the parameters, etc.
However, the selection of the parameter interpolation region is also an important factor.

Another object of the invention is to improve the method of quantizing line spectral frequencies.

Conventionally, the relative superiority of α parameters, K parameters, and line spectrum frequencies obtained as a result of linear predictive analysis has been discussed in terms of their quantization sensitivities and interpolation characteristics. Regarding quantization sensitivity, it is said that the α parameter is the highest (in other words, it is unsuitable for quantization), the K parameter is the next highest, and the line spectrum frequency is the lowest. or,
Regarding the interpolation characteristics, the K parameter has the largest distortion in the time axis direction (that is, it is unsuitable for interpolation),
Next, the α parameter is the largest, and the line spectrum frequency is the smallest. Recent vocoders often use line spectrum frequencies that are more advantageous than other parameters in both quantization and interpolation characteristics. Although the line spectrum frequency has excellent quantization characteristics and interpolation characteristics as described above, it is not necessarily perfect and also has some problems. The typical problem is P
Regarding the next (P is an integer) line spectral frequencies ω ₁ to ω _P , ω ₁ to ω _P are not independent from each other.
For example, the stability of a synthetic filter with K parameters K ₁ to K _P is as follows: |Ki|<1 (i=
1, 2, ... P), whereas the stability of the synthesis filter with line spectral frequencies ω ₁ to ω _P is 0 < ω ₁ < ω ₂ ... < ω _P < π (however, the unit is radians) must be satisfied. In addition, in the synthesis filter using line spectrum frequencies ω ₁ to ω _P , adjacent coefficients ω _i and ω _i+1 are ω _i ≒ ω _i+1 (i=1, 2,...P
-1), it has extremely high selectivity.

Figure 3 shows the 8th order line spectrum frequency obtained by analyzing approximately 1.9 seconds of audio. It can be seen from FIG. 3 that, for example, there are many places where the interval between ω ₁ and ω ₂ is extremely narrow. From the figure, if we independently quantize ω ₁ and ω ₂ , we get
Each small amount of quantization error causes the interval between ω ₁ and ω ₂ to change relatively significantly, causing a large change in the selectivity of the synthesis filter. Furthermore, as a single frequency of line spectrum frequency does not correspond to clear parameters of physical properties such as formants, its temporal change characteristics are unclear, and it is not always possible to obtain sufficient interpolation characteristics by linear interpolation etc. It's not something you can do.

On the other hand, from the relationship between ω ₁ and ω ₂ in Figure 3, ω ₁
It is clear that a pole with high selectivity exists near the middle between and ω ₂ and is thought to correspond to the first formant.

Since conventional line spectrum vocoders independently quantize line spectrum frequencies, the interval between ω ₁ and ω ₂ changes relatively significantly due to the influence of quantization errors, which greatly changes the selectivity of the synthesis filter. The first disadvantage is that the line spectrum frequency is interpolated independently, and the second disadvantage is that it is not always possible to correspond to formants, etc., and it is not possible to obtain sufficiently satisfactory interpolation characteristics. had.

SUMMARY OF THE INVENTION An object of the present invention is to provide a high-quality vocoder by improving the quantization characteristics and interpolation characteristics of line spectral frequencies in a line spectral type vocoder.

In the line spectrum type vocoder of the present invention, on the analysis side, adjacent odd-order line spectrum frequencies (ω ₁ , ω ₃ ,
..., ω _P-1 ) and even-order line spectral frequencies (ω ₂ , ω ₄ , ..., ω _P ) are treated as a pair, and the intermediate value and interval of the line spectral frequencies constituting each pair are calculated. The main means is a means for output quantization, and the main means on the synthesis side is a means for independently interpolating the intermediate value and the interval.

The present invention relates to a line spectrum type vocoder, in which adjacent odd-order line spectrum frequencies and even-order line spectrum frequencies are collectively quantized and interpolated, thereby changing the interval between the line spectrum frequencies due to quantization. The first effect is that the stability of the synthesis filter is improved, and the second effect is that good interpolation characteristics are obtained because the interpolation characteristics correspond relatively well to the formants.

Next, embodiments of the present invention will be described with reference to the drawings. FIG. 4 is a block diagram for explaining an embodiment of the present invention. The first embodiment includes an analysis side 410,
It is composed of a transmission line 430 and a combining side 440.
Furthermore, the analysis side 401 includes low-pass filters 1 and 41.
1, A/D converter 412, window processor 413, autocorrelation coefficient measuring device 414, linear prediction analyzer 415, residual power calculator 416, line spectrum frequency analyzer 417, line spectrum frequency quantizer 418, Residual power quantization 419, pitch, v
-uv analyzer 420, pitch, v-uv quantizer 4
21 and a multiplexer 422. Synthesis side 4
40 is a demultiplexer 441, a pitch, a v-uv decoder 442, a pulse generator 443, a noise generator 4
44, v-uv switch 445, variable gain amplifier 446, residual power decoder 447, line spectrum frequency decoding signal 448, intermediate frequency interpolator 4
49, frequency interval interpolator 450, line spectrum frequency restorer 451, LSP synthesis filter 452,
D/A converter 453, low pass filters 2, 45
Consists of 4.

The audio signal is input to the low pass filter 1, 411 via the waveform input terminal 401. The low-pass filter 1, 411 limits the band of the audio signal to, for example, 3.4 kHz and outputs it to the A/D converter 412. A/D
The converter 412 samples the audio signal band-limited to 3.4 kHz at, for example, 8 kHz, further quantizes it, and outputs it to the window processor 413. The window processor 413 takes the sampled and quantized audio signal as one block, for example, 240 samples, that is, 30 mSEC, and uses the autocorrelation coefficient measuring device 41 in block units.
4, pitch, and output to the v-uv analyzer 420.
The output period in each block is, for example, 20 m.
It is SEC. The autocorrelation coefficient measuring device 414 measures the delay of the blocked audio signal from delay "0" to delay "P".
(P is an integer). The autocorrelation coefficient of delay "0" is the power of the blocked audio signal. The autocorrelation coefficient measuring device 414 sends the P+1 autocorrelation coefficients from delay "0" to delay "P" to the linear prediction analyzer 415, and calculates the residual power using the autocorrelation coefficient of delay "0". output to the device 416. The linear prediction analyzer 415 performs linear prediction analysis using the autocorrelation method from the P+1 autocorrelation coefficients to calculate the P-order α parameter and the normalized prediction residual power. Of the calculation results, the linear prediction analyzer 415 further outputs the normalized prediction residual power to the residual power calculator 416 and the P-order α parameter to the line spectrum frequency analyzer 417. The residual power calculator 416 calculates the residual from the delayed “0” autocorrelation coefficient, that is, power supplied from the autocorrelation coefficient measuring device 114 and the normalized prediction residual power supplied from the linear prediction analyzer 415. The power is calculated and the result is output to the residual power quantizer 419. The residual power quantizer 419 performs log-linear quantization on the residual power, for example, and outputs the result to the multiplexer 422 . The line spectrum frequency analyzer 417 is disclosed in, for example, patent application No. 57-
114195 “Line spectrum type speech analysis and synthesis device” From the P order α parameter inputted by the method described in “Line spectrum type speech analysis and synthesis device”, the P order line spectrum frequencies ω ₁ , ω ₂ , …,
ω _P , (0<ω ₁ <ω ₂ ...<ω _P <π) is obtained, and the frequency is further output to the line spectrum frequency quantizer 418. The line spectral frequency quantizer 418 calculates the intermediate value of P/2 sets of line spectral frequencies, ω ₁ +ω ₂ /2.
, ω ₃ +ω ₄ /2,…, ω _P-1 +ω _P /2, and the interval of the P/2 set of line spectral frequencies, (ω ₂ −ω ₁ ), (ω ₄ −ω ₃ ),…,
(ω _P −ω _P-1 ) is calculated, and the calculation results are each independently linearly quantized. Line spectrum frequency quantizer 4
18 outputs the quantization result to multiplexer 422.
The pitch and v-uv analyzer 42 analyzes the pitch period and the voiced/unvoiced discrimination signal from the blocked audio signal supplied from the window processor 413, and sends the analysis results to the pitch and v-uv quantizer 421.
Output to. A pitch/v-uv quantizer 421 linearly quantizes the pitch period, for example, and outputs the result to a multiplexer 422 . The pitch and v-uv quantizer 421 processes the voiced/unvoiced discrimination signal by setting the pitch period to "0" in the case of unvoiced signal. The multiplexer 422 is the residual electron quantizer 4
The quantized residual power information supplied from the line spectral frequency quantizer 418 and the intermediate value of the quantized P/2 set of line spectral frequencies and the quantized line spectral frequency of P/2 supplied from the line spectral frequency quantizer 418 The interval and quantized pitch period information supplied from the pitch and v-uv quantizer 421 are multiplexed, and the multiplexed result is sent to the demultiplexer 44 via the transmission line 430.
Output to 1. The demultiplexer 441 separates the multiplexed quantized data and sends the quantized residual power information to the residual power decoder 447 as the intermediate value of the quantized P/2 set of line spectrum frequencies and The line spectrum frequency interval of P/2 is output to the line spectrum frequency decoder 448, and the quantized pitch period information is output to the pitch and v-uv decoder 442, respectively. pitch, v-uv decoder 442
decodes the quantized pitch period information, and further generates a voiced/unvoiced discrimination signal by determining that it is unvoiced if the decoding result is "0", and determining that it is voiced if it is not "0". Pituchi, v-uv decoder 4
42 outputs the decoded pitch period information to the pulse generator 443 and the voiced/unvoiced discrimination signal to the v-uv changeover switch 445, respectively. The pulse generator 443 generates a pulse train having a period corresponding to the supplied pitch period information, and further outputs the generated pulse train to the v-uv changeover switch 445.
Noise generator 444 generates white noise and further outputs the generated white noise to v-uv changeover switch 445. The v-uv changeover switch 445 is pitch,
If the voiced/unvoiced discrimination signal supplied from the v-uv decoder 442 is voiced, a pulse train is selected; if it is unvoiced, white noise is selected, and the selection result is output to the variable gain amplifier 446. Residual power decoder 447 decodes the quantized residual power and outputs the decoding result to variable gain amplifier 446. The variable gain amplifier 446 amplifies the amplitude of the pulse train or white noise supplied via the v-uv switch 445 with a gain proportional to the square root of the residual power supplied from the residual power decoder 447, and further amplifies it. The resulting pulse train or white noise is output to the LSP synthesis filter 452. The line spectral frequency decoder 448 decodes the intermediate value of the quantized P/2 set of line spectral frequencies supplied from the demultiplexer 441 and outputs the intermediate value to the intermediate frequency interpolator 449. Line spectral frequency decoder 448 further decodes the interval of the quantized P/2 sets of line spectral frequencies supplied from demultiplexer 441 and outputs the interval to frequency interval interpolator 450 . The intermediate frequency interpolator 449 independently linearly interpolates the intermediate values of the P/2 sets of line spectral frequencies supplied from the line spectral frequency decoder 448. The interpolation interval is, for example, one sampling period, 125 μSEC. The intermediate frequency interpolator 449 further outputs the interpolated intermediate value of the P/2 set of line spectral frequencies to the line spectral frequency restorer 451. Frequency interval interpolator 4
50 independently linearly interpolates the intervals of P/2 sets of line spectrum frequencies supplied from the line spectrum frequency decoder 448. The interpolation interval is, for example, one sampling period, 125 μSEC. Frequency interval interpolator 450
further outputs the interpolated P/2 set of line spectral frequency intervals to the line spectral frequency restorer 451. The line spectral frequency restorer 451 receives intermediate values of P/2 sets of interpolated line spectral frequencies supplied from the intermediate frequency interpolator 449 and P/2 sets of interpolated line spectral frequencies supplied from the frequency interval interpolator 450. The P-order line spectrum frequency is restored from the spectral frequency interval, and the LSP synthesis filter 4 uses the restored line spectrum frequency as a filter coefficient.
Output to 52. The LSP synthesis filter 452 receives the signal supplied from the variable gain amplifier 446 as input, resynthesizes the audio waveform using the line spectrum frequency as a filter coefficient, and further outputs the resynthesized audio waveform to the D/A converter 453. . The D/A converter 453 converts the resynthesized audio waveform, which is a sampling series, into a continuous audio waveform, and outputs it to the low-pass filter 2, 454. Low-pass filter 2, 454
removes aliasing components included in the continuous speech waveform and outputs resynthesized speech to the waveform output terminal 460. Note that in this embodiment, instead of linearly quantizing the line spectral frequency intervals, it is clear that nonlinear quantization can be performed in consideration of the quantization sensitivity regarding the line spectral frequency intervals and the distribution characteristics of the intervals. It is.

[Brief explanation of drawings]

Figure 1 a and b are waveform diagrams to explain the effect of interpolation, Figure 2 is a spectrum diagram to explain ideal interpolation, and Figure 3 is a line spectral frequency diagram found by analyzing real speech. FIG. 4 is a waveform diagram for explaining temporal changes, and a block diagram for explaining one embodiment of the present invention. 401... Waveform input terminal, 410... Analysis side,
411...Low pass filter 1, 412...A/
D converter, 413...Window processor, 4
14...Autocorrelation coefficient measuring device, 415...Linear prediction analyzer, 416...Residual power calculator, 417...
... line spectrum frequency analyzer, 418 ... line spectrum frequency quantizer, 419 ... residual power quantizer, 420 ... pitch, v-uv analyzer, 42
1... Pitch, v-uv quantizer, 422... Multiplexer, 430... Transmission line, 440... Combining side,
441... demultiplexer, 442... pitch, v-
uv decoder, 443...pulse generator, 444
... Noise generator, 445 ... V-UV changeover switch, 446 ... Variable gain amplifier, 447 ... Residual power decoder, 448 ... Line spectrum frequency decoder, 449 ... Intermediate frequency interpolator, 450...
...Frequency interval interpolator, 451...Line spectrum frequency restorer, 452...LSP synthesis filter, 45
3...D/A converter, 454...Low pass filter 2, 460...Waveform output terminal.

Claims (1)

[Claims] 1. In a line spectrum type vocoder that analyzes and synthesizes speech by extracting line spectrum frequencies from linear prediction coefficients obtained by linear prediction analysis from an input audio signal, adjacent odd-order line spectrum frequencies and even-order lines are used. The analysis side includes means for handling the spectrum frequencies as a pair and calculating and quantizing the intermediate value and interval of the line spectral frequencies constituting each pair, and means for independently interpolating the intermediate value and the interval. A line spectrum type vocoder characterized in that it has on the synthesis side.