JPS5912499A - Voice encoder - 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 speech encoding device, and more particularly to a speech encoding device that separates and encodes speech into short-time spectrum envelope information and sound source information.

BACKGROUND OF THE INVENTION With the development of linear prediction techniques, a number of speech encoding devices have been realized for the purpose of highly efficient transmission of speech. Among these devices, there is an encoding device that has a table for vector quantizing the short-time spectral color fringe information vector of speech. For example, as shown in FIG. 1, this table consists of short-time spectrum envelope information vectors and indices indicating them. In FIG. 1, the short-time spectral envelope information vector is shown as an example of a linear prediction coefficient vector. The linear prediction coefficient vectors in this table are vectors representing several model spectra designed in advance so that the average distortion during vector quantization is minimized for the speech to be encoded. .

A conventional encoding device having this vector quantization table will be explained using FIG. 2. Note that in FIG. 2, linear prediction coefficients are shown as an example of short-time spectrum envelope information. In the same figure, an audio signal passes through a low-pass filter 1 and is A-D converted by an A-D converter 2.
It is divided into frames of approximately ms&, and each frame is processed.

Then, the vector quantizer 3 performs an operation of fitting the linear prediction coefficient vector of each frame to a model with the minimum distortion distance from among the model spectra in the quantization table 4.

As an example of this vector quantizer 3, an apparatus using an inverse filter bank is shown in FIG. Vector quantization table 4
Let the zero linear prediction coefficient vector be a2 to a. Here n is the total number of vectors in the table. Further, each vector 81 to an is a vector as shown below.

“i” (ail, ai2+ ・・・・・・+aim)
(1<:i<, n) Here, m indicates the order of the model spectrum.

The inverse filter banks A1(Z) to An(Z) in FIG. 3 have coefficients al to Δan, and each audio frame is input to these filters in parallel. Since each filter is an inverse filter, 2 of these outputs per audio frame interval
By comparing the sums of the powers, the one with the smallest distortion distance can be selected from Δ a1 to this model spectrum. That is, the value of the sum of squares is a
1 is closest to the spectrum of the input audio frame. Therefore, the third
The output Δ of the vector quantizer shown in the figure outputs the index of the vector quantization table 4 of FIG. 2 corresponding to a i.

Through the above operations, the vector quantizer 3 converts the spectrum of each frame into a model spectrum K t , and the corresponding index sequence becomes the output of the vector quantizer 3.

Next, the pinch/gain calculator 6 calculates the pitch and gain of each frame and quantizes them. The quantized pitch and gain values are encoded together with the output of the vector quantizer 3 by an encoder 6.

FIG. 4 shows a decoding device for the above encoding. The sent code is decoded by the decoder A, and the index and gain of the quantization table 8 are decoded.

Separated into pitches. The index is input to the quantization table 8, and the linear prediction coefficient vector of the corresponding model spectrum is output to the synthesis filter 1o.

The pitch and gain are input to the sound source signal generator 9. The output of the synthesis filter 1o is converted into an audio output by a DA converter 11 and a low-pass filter 12.

The above is the configuration of a conventional audio encoding device having a vector quantization table. The quality of audio in this device is determined by the number of model ΔΔ−rus vectors in the vector quantization table, that is, the total number of a1 to an. Therefore, if this encoding device is to encode any voice with small distortion, the number of 81 to an must be made very large.
It has several problems as follows. First, designing a large quantization table takes a very long time. Secondly, as the number of model spectra increases, the information compression rate during encoding deteriorates. Third, the number of inverse filters in the vector quantizer increases, making the device configuration complicated.

In view of the above drawbacks, the present invention provides a speech encoding device capable of highly efficient encoding by preparing quantization tables with different characteristics in advance and switching these according to the quality of input speech. It is. An embodiment of the present invention will be described below with reference to the drawings.

FIG. 5 is a block diagram of a speech encoding device in one embodiment of the present invention.

In Fig. 5, 1 is a low-pass filter that passes the low frequency range of the input audio signal, and 2 is a low-pass filter that converts the signal output from the low-pass filter from analog to digital, and converts the analog-to-digital converted signal. 10~20m5e
A-D converter 3 divides each frame into approximately C frames, and 3 quantizes the spectrum of each frame into a model spectrum and outputs the corresponding index series.
A vector quantizer as shown in the figure; 5 is a pitch/gain divider that calculates and quantizes the pitch and gain of each frame; 6 is a pitch/gain calculator 5 that calculates the pitch and gain of each frame; This is an encoder that encodes the determined pitch and gain.

The above-described low-pass filter 1, A-D converter 2, vector quantizer 3, pitch/gain calculation section 5, and encoder 6 have the same configuration as shown in FIG. The difference from the configuration in Figure 2 is the second
Instead of the vector quantization table 4 shown in the figure, tables T1 to Tn each store different features in advance.
, an adaptive quantization table is constructed by a table number detector 13a that selects an optimal table number according to the audio spectrum envelope information vector from the widths of tables T1 to T2, and a table switch 13b that switches tables.

The operation of the speech encoding device configured as described above will be described below.

Amazu audio signal input is converted into a digital signal by a low-pass filter 1 and an A-D converter 2, and is
It is processed every frame of about ec.

Next, the table number detector 13a to which the voice characteristics are input observes the average distortion between the input voice spectrum envelope information vector and the vectors of each table T1 to Tn for a certain period of time, and determines the average distortion according to the quality of the input voice. An optimal table number is selected and this number is sent to the table switch 13b. An optimal table is selected by the table switch 13b, and the information is sent to the vector quantizer 3. Next, encoder 6
Now, the index obtained by the vector quantizer 3 and the pitch and gain information obtained by the pitch/gain calculator 5 are encoded at the same time. When a table switch occurs, this table number is also encoded and the input audio is encoded.

On the other hand, FIG. 6 shows the configuration of a decoding device corresponding to the encoding device of FIG. 5. The sent code is decoded using a decoding booklet and separated into the index, gain, and pitch of the adaptive quantization table 13, respectively. This index is input to the adaptive quantization table 13, separated into the table number and the index of the table, and the linear prediction coefficient vector of the corresponding model spectrum is output to the synthesis filter 10. The subsequent processing is equivalent to the conventional example.

In this embodiment, the linear prediction coefficient Betta 1H is used as an example of the short-time spectral envelope information vector, but it can be used for any other short-time spectral envelope information vector.

As described above, the present invention provides low-distortion encoding for various voices by preparing in advance a plurality of quantization tables with different characteristics and providing table number detection means and table switching means for switching the quantization tables. It is of great industrial value, as it is easy to do this, and the table number information only needs to be encoded when switching, allowing substantially more efficient encoding than in the past.

[Brief explanation of drawings]

FIG. 1 is a diagram showing a vector quantization table, FIG. 2 is a block diagram of a conventional speech encoding device, FIG. 3 is a block diagram showing a specific configuration of a vector quantizer, and FIG. A block diagram of a decoding device corresponding to the code A1 device in FIG.
FIG. 6 is a block diagram of a speech encoding device according to an embodiment of the present invention, and FIG. 6 is a block diagram of a decoding device corresponding to the encoding device of FIG. 6. 1...Low pass filter, 2...A-D converter, 3...Vector quantizer, 13...
...Adaptive quantization table, 13a...Table number detector, 13b...Table switch, T1
~TN...Table. Name of agent: Patent attorney Toshio Nakao and 1 other person No. 1
Figure 2 Figure 3 Figure 4

Claims (1)

[Claims] A feature extracting means for sequentially extracting features from an audio signal, a plurality of vector quantization tables each storing different features in advance, and a method according to the features of the audio sent from the feature extracting means. A speech encoding device comprising table number detection means for detecting an optimal table among the vector quantization tables, and table switching means connected to the vector quantization table and switching tables according to the output of the table number detection means.