JPS6272000A - Encoding method for voice - 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 [Industrial Application Field] The present invention relates to a speech encoding method for extracting a vector parameter from input speech and encoding it at a low bit rate.

[Conventional technology]

Conventionally, the method of encoding audio at a low bit rate of 1000 bps or less has been vector quantization (for example, A, Buzo+etc. 5peech Coding).
based 1ponVector Quantiz
ation,' IEEE, ASSP-28,198
0) and variable frame rate encoding (e.g. tubing, sheet metal:
- There are two methods: speech information compression by linear approximation of solameter (Speech Research Group Material S-78-13, 1978). The former vector quantization method maintains the frame unit (speech analysis unit) constant and calculates the spectrum of each frame.
It quantizes P parameter information to about 8 bits, and is characterized by treating the parameters as one vector.

However, this method removes only spatial redundancy, that is, frequency redundancy, and when the speed is lower than 500 bps, the quality deteriorates rapidly because the frame unit is constant.

On the other hand, the latter variable frame rate encoding method changes the frame unit (frame length) in response to temporal changes in the spectrum, removes redundancy in time, and reduces the average transmission rate to 1/2. Even if the number is reduced to about 3, there is little deterioration in quality. However, this method is essentially a grammeter (
Because it depends on the interpolation characteristics (linear), the transmission speed is 2 per second.
If the number of frames is less than 5 frames (total 600 bps), the quality will deteriorate rapidly.

There is also a method that encodes the time series of the latest spectral parameters in segment units (Japanese Patent Application No. 59-8085
No. 5, Shiraki, Honda; Very low bit speech coding using spatio-temporal spectrum, Acoustical Society of Japan Proceedings 1-2-3゜1984 3
Month). This method cannot sufficiently reduce encoding distortion because matching with the standard pattern is performed in a fixed dimension and the determination of the segment position and the selection of the standard button are not integrated.

An object of the present invention is to provide a speech encoding method that can reproduce speech with good text intelligibility even at a low bit rate of 600 bps or less.

[Failure to solve the problem]

According to this invention, a spectral nomurameter time series of input speech is divided into segments, and the segment division position is corrected so that the matching distance between the segment series and a standard pattern with a fixed time length is minimized. Determine the standard/g tongue and segment division positions. In this invention, encoding distortion is minimized by integrating the determination of segment division positions and the selection of standard patterns. Furthermore, the creation of the standard tsukutan described above is performed by clustering the segment series of the spectral parameters of the learning speech, obtaining a standard pattern for each class, and using this standard pattern, two procedures are performed: modifying the segment division position of the learning speech and updating the standard pattern. By repeating this process, we aim to reliably reduce encoding distortion.

In the conventional segment-based encoding method, the time-series segmentation (dividing into segments) of four-dimensional audio and the selection of standard and tangents are processed separately, resulting in good segmentation. cannot be obtained, so the encoding distortion has not been minimized. In addition, good segmentation cannot be obtained even if the standard 4-tank is created, and the standard 4-tank is constructed based on that segmentation. It was difficult to make it smaller. Furthermore, in this invention, the matching measure with the standard pattern is not a fixed time length, but is a measure that matches the standard pattern with the time length of the input segment so that the spectral distortion upon decoding is minimized.

〔Example〕

FIG. 1 shows an embodiment of the speech encoding method of the present invention. Audio input from input terminal 1 is band-limited by low-pass filter 2 and input to AD converter 3, where it is periodically sampled (
in this example, 8000 times per second) and converted into a digital signal. The output of this AD converter 3 is the LPC analyzer 4
The spectrum/grammeter of the input voice is extracted. L
The input speech spectrum/grammeter time series calculated through PG analysis is divided into segments by the segmentation unit 5, for example, at boundary points determined in advance by inspection.

In this embodiment, the phoneme boundaries determined by the reading of Funadalam are used as segment points. The divided segment series is modified by the modification unit 6 using dynamic programming so that the distance between the segment sequence and the standard button stored in advance in the standard button memory 7 within a continuous short speech section is minimized. After the segment division position is corrected, the corrected segment division position or segment length is encoded, and the number of the standard pattern most similar to this is output. The matching distance between the segment series and the standard nogtan is defined by the weighted Euclidean distance including power after performing linear transformation on the standard yatan prepared in advance to make its length equal to the input segment length. .

In this example, the 12th order LSP (L
The standard baton is a 13 x 10-order matrix in which 10 parameters of logarithmic speech (L...L) and logarithmic speech/Yawa P1 are arranged horizontally. If the input segment length is l,
Assuming that the matrix of the input segment is Xj (13) linear matrix) and the projection matrix to be made one-dimensional from SH10 by 5m transformation is HJ, the matching distance between Xj and Xo is calculated by the following equation.

d (X', Xj)2=llXj-X'Hz If2
・・・・・・・・・・・・・・・・・・・・・(1)
Is this invention standard? The matching distance between the input segment and the standard tongue is determined by making the tongue equal to the length of the input segment.

Modifying the input segment division position using dynamic programming is as follows:
Do as follows. Let σ(Ts) be the cumulative distance (sum of matching distances) up to time Ts within one short sound interval Im, and M be the number of segments within one short sound interval Im. Set the division position correction width Δ appropriately, and set the time Ts-1 according to an arbitrary recurrence formula.
decide.

σ(Ts)=min (σ(Ts-1)+d(Ts-1
, Ts)) ""...""・(2) Ts-Madashi IT, -T, -, l≦(Δ-1)/2: s=
1.2...Mσ(To)=0: d is time Ts
Determine the time TM that minimizes the cumulative distortion σ(TM) at the end point of the value obtained by quantizing the input segment from −1 to Ts using equation (1), and correct the position of each segment obtained using equation (2). Determine points sequentially.

Next, how to make standard Yatan will be explained with reference to Figure 2. First, a voice prepared in advance for standard pattern learning is input, a spectral parameter time series of the learning voice is created, and the time series is further divided into segments. Is this segment boundary known? Cluster the parameter time series and create an initial standard/Fmethane for each class. This method is performed, for example, by Gray's method using the distance measure of equation (1).

Regarding Gray's method, see, for example, A-Buzo et al.
peachCoding based upon Ve
ctor Quantli;ati:On+ IEEE
.

See ASSP-28pp562-pp574 (1980). Using this initial standard pattern, you can use it for learning purposes. The segment division position of the parameter time series is corrected. This modification method uses the dynamic programming method described above. This modification does not increase the total quantization distortion. That is, if the total quantization distortion at the initial segment division position is D(0) and the total quantization distortion after correction is D(1), then D (0)≧D(1) ...・・・・・・
・・・・・・・・・・・・・・・・・・(3) holds true. Next, the standard pattern is updated according to the procedure shown below from the segment series of the time series of stiffness/tornocrameter of the learning speech whose division positions have been corrected.

In other words, the standard before any update? Let Ni be the number of segments whose division positions have been corrected. When the division position is corrected, the set of segments created from the segments quantized according to this standard method is clustered again, and the segment series whose division position has been corrected is clustered again, and one class is defined as (Xy;). ＞=1.2. ..., Ni),
The standard pattern X is updated to a projection matrix Ht corresponding to the center of gravity of this set, H,: Xy: transposed matrix B, an inverse matrix. In general, the standard pattern &i before updating does not match the standard pattern X1 after updating G', so updating the standard pattern does not increase the total quantization distortion. In other words, if the quantization distortion after updating the standard pattern is D(2), then D(1)≧D(2)...
・・・・・・・・・・・・・・・・・・(5) holds true. In the same way, the total quantization distortion is calculated by modifying the segment division position and updating the standard pattern. 3)≧・・・・・・
・≧D(K)≧D≧(K+1) (6). This standard i4 tan is updated until the reduction rate of the total quantization distortion becomes equal to or less than a predetermined value. Regarding the -membrane inverse matrix, for example, Rao, Mitra/Shibuya.

Please refer to "Other Translations 1 General Inverse Matrices and Their Applications" Tokyo Tosho (1973).

FIG. 3 shows an example in which the total quantization distortion is a non-increasing column.

In this example, the number of segments is 2000. The dimension of the standard Q tan in the time direction is 10. The standard number of clicks is 64, and the segment division position is corrected when the longest segment length is 32. The correction width Δ=33. (LPG analysis has an analysis window length of 30 m5.
ec + shift length 1010 m5e speaker is male - first name].

From this figure, it is verified that the quantization distortion is in a non-increasing sequence, and it can be seen that the distortion is reduced to about 80 factors compared to the initial value, and is significantly reduced after one update.

Returning to the explanation of FIG. As described above, the segment division position of the input waste vector time series is corrected by the segment position correction unit 6, and the division position (segment length) is encoded. Further, the number of the optimal standard pattern, the input audio pinch information, and the duration information of each segment are combined by a multiplexer 8 and output as encoded output.

This audio encoded output is transmitted or stored, and for decoding, a standard pattern is obtained by referring to the standard pattern memory 9 using the standard pattern number, and then linear conversion is performed using the duration information, and the spectrum is The parameter time series is restored, and the LPC synthesis unit 10 performs LP from this and the pitch information.
A signal corresponding to the C analysis input is synthesized, the synthesized output is converted into analog by a DA converter 11, and an analog voice is outputted to an output terminal 13 through a low-pass filter 12.

〔Effect of the invention〕

The number of segments is 20,000 standard number 4 tan'i 10
24, the segment division position was corrected once (correction width Δ=9), and an intelligibility test of 100 syllables was conducted using the updated standard pattern. When the correction width Δ=13, the phoneme intelligibility was 78. Good audio quality was obtained. In this case, the average number of segments is about 8 per second, so the spectral information of this encoded speech is 8 x (10
+5') = 120 bps. If the phoneme intelligibility is 75% or higher, 50 out of 100 people will have a sentence intelligibility of 100.

Therefore, the phoneme clarity of 78 points is a good result.

As explained above, according to the present invention, it is possible to obtain spectral information, such as coded speech that is sufficiently clear even at extremely low speeds such as about 120 bps. It has the advantage that it can be used for configuration, etc.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing an example of the present invention, FIG. 2 is a diagram showing the procedure in the standard button creation section in FIG. 1, and FIG.
The figure is a diagram showing the relationship between segment correction, the number of repetitions of A'methane new, and reduction in coding distortion.

Claims (2)

[Claims] (1) Extract the spectral parameters of the input audio frame by frame, divide the time series of the extracted spectral parameters into segments, and while modifying the division position, divide the segment series into a standard with a fixed time length prepared in advance. A speech encoding method that determines a division position and the most similar standard pattern so that the matching distance with a pattern is minimized, and outputs a code indicating the determined division position and standard pattern. (2) Input the training speech, extract its spectral parameters frame by frame, divide the time series of the extracted spectral parameters into segments, cluster the segments, determine the standard pattern for each class, and Using the determined standard pattern, select the most similar standard pattern while correcting the segmentation position of the segment series of the training speech using the determined standard pattern, recluster the segment series at the modified segmentation position, and create the standard of each class. Redetermine the pattern, repeat the above-mentioned division correction, re-clustering, and redetermining the standard pattern at least once, and finally make the re-determined standard pattern the standard pattern using the encoding of the input speech. A speech encoding method according to claim 1, characterized in that: