JPH04312000A - Vector quantization method - Google Patents

Abstract

PURPOSE:To reduce the number of quantized bits and the amount of operation, regarding a vector quantization method used for a coding and decoding device to quantize digital signals such as A/D converted speech signals. CONSTITUTION:A vector quantization method has the constitution wherein an encoder A and a decoder B are provided a plurality of code books 10 and 11 grouped depending upon the state of a speech. The encoder A includes an input vector calculation means 1, a partial speech decoding means 3 for selecting a typical vector from the code book 10, a code book selector means 4, and a prediction and selector means for a typical vector selector means 2. Also, the decoder B includes a typical vector read means 7, a speech decoding means 8 and a code book selector means 9, and is used for reading out a typical vector corresponding to received data from the code book 11 for decoding.

Description

[Detailed description of the invention]

0001

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a vector quantization method for quantizing digital signals such as A/D converted audio signals.

0002

[Prior Art] Conventionally, this type of vector quantization method is shown in FIG.
As shown in FIG.
, a representative vector selecting means 102, a transmitting means 103, and a codebook 107 consisting of N code vectors are provided, and the decoder side includes a receiving means 104, a representative vector reading means 105, a speech synthesis means 106, and an encoder side. The same codebook 108 is provided. The codebook used here consists of code vectors that can be applied to all speech conditions.

A conventional vector quantization method will be explained below using the block diagram shown in FIG. First, on the encoder side, A/D
The converted audio is stored for a certain period of time (hereinafter referred to as a frame)
Input vector X = (X1, X
2, ~ , Xn). Then, the representative vector selection means 102 selects the code vectors Y1, Y2, . . . , Yn of the codebook 107 for the input vector X.
Find the code vector Yi = (y1, y2, ..., yn) with the smallest distortion di given by (Equation 1) from among
The index i of the code vector is transmitted to the means 103.
Send by.

0004

[Math 1]

On the decoder side, based on the index information of the representative vector received by the receiving means 104, the representative vector reading means 105 reads a representative vector corresponding to the index information of the corresponding representative vector using the same code vector as on the encoder side. Codebook 108 configured
The speech synthesis means 106 synthesizes (decodes) a speech signal using the read code vector and outputs the synthesized speech signal.

As described above, in the conventional vector quantization method described above, a representative vector is searched and selected for all code vectors in the codebook for an input vector.

[0007]

[Problems to be Solved by the Invention] However, in the conventional vector quantization method described above, the codebook used for vector quantization is composed of code vectors that can be applied to all speech states. Since code vectors of states must be stored in a codebook, it is necessary to increase the number of code vectors in order to improve quantization accuracy. however,
As the number of code vectors increases, the number of bits required to represent the quantized value (quantization bit number) also increases.
There is a problem in that the amount of calculation increases because the distortion with all representative vectors is calculated.

The present invention solves the above-mentioned conventional problems, and aims to provide a vector quantization method that can improve quantization accuracy with a limited number of quantization bits and a limited amount of calculations. be.

[0009]

[Means for Solving the Problems] In order to achieve the above-mentioned object, the present invention has the following features:
It has multiple codebooks grouped by voiceless consonants, silence, etc.), predicts the audio state of the current frame from the audio state of past frames, and generates the corresponding code from the grouped multiple codebooks. Select the book and
It has a configuration that performs quantization using a selected codebook.

[0010] Also, an input vector calculation means for calculating a vector quantity of an input audio signal to be vector quantized, a plurality of first codebook means grouped according to the state of the audio, and a first codebook selection means described later. representative vector selection means for selecting a representative vector from the code vectors of the first codebook means selected in the first codebook means; audio local decoding means for decoding speech from the selected representative vector; encoding means including a first codebook selection means for predicting and selecting a codebook to be used in the next frame from the codebook of the second frame; a plurality of second codebook means grouped according to the state of speech; representative vector reading means for reading out a representative vector corresponding to data from a codebook selected by a second codebook selection means described later; speech decoding means for decoding speech using the read representative vector; The decoding means includes a second codebook selection means for predicting and selecting a codebook to be used in the next frame from the second codebook by referring to the speech.

[0011]

[Operation] With the above-described configuration, the present invention selects which group of codebooks to use for the current frame audio based on the state of past frame audio, and quantizes using the codebook. Therefore, the quantization accuracy can be improved with limited quantization bits and a limited amount of calculation.

0012

DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 is a functional block diagram showing a vector quantization method according to an embodiment of the present invention. As shown in the figure, this embodiment includes a plurality of codebooks 10 and 11 grouped according to the state of speech on the encoder side and the decoder side, and the encoder side includes an input vector calculation means 1 and a representative vector selection means. 2. Local decoding means that synthesizes the same voice as the decoder side using the selected representative vector 3. Predicts the next state using the locally decoded voice and quantizes the voice of the next frame using which group of codebooks The decoder side includes a codebook selection means 4 for determining the index of the code vector, a transmission means 5 for transmitting the index of the code vector to the decoder side, a reception means 6 for receiving the transmitted index, and a representative corresponding to the received index. representative vector reading means 7 for reading vectors from the codebook 11 selected by the codebook selection means 9; speech synthesis (decoding) means 8 for synthesizing (decoding) speech using the read representative vectors; and an encoder from the decoded speech. A codebook selection means 9 is provided for predicting the next state using the same method as the side and determining which group of codebooks should be used to quantize the audio of the next frame.

FIG. 2 is a block diagram showing the hardware configuration of an embodiment of the vector quantization method of the present invention. 21
is an A/D converter that converts audio signals into digital signals,
22 is a central processing unit (CPU) that converts the input audio signal into an input vector suitable for vector quantization and performs operations as shown in FIG. A read-only memory (ROM) stores a codebook, and a random access memory (RAM) 24 stores data used in vector quantization.

Next, the operation of one embodiment of the vector quantization method described above will be explained with reference to the flowchart of FIG. 3, which explains the operation of the CPU of FIG. 2, and the functional block diagram of FIG. 1. First, when audio is input (step 1), the input vector calculating means 1 calculates the input vector X=(X1, X2, . . . , Xn) of the input signal to be vector quantized (step 2). Then, the representative vector selection means 2 selects the code vector Ymi with the smallest distortion dmi given by (Equation 2) from among the code vectors Y1, Y2, ..., Yn of the codebook m selected by the codebook selection means 4. =(
y1, y2,..., yn) However, (m=1, 2,...
, M) and use it as a representative vector. (Step 3).

[0016]

[Math 2]

The speech local decoding means 3 decodes the speech using the selected representative vector Ymi in the same manner as on the decoder side (step 4), and using the speech decoded by the codebook selection means 4, for example, , a codebook to be used in the next frame is selected from M codebooks in the following manner (step 5).

1. If the decoded audio is silent, there is a high possibility that the next frame's audio will be silent, so a silent codebook is selected. The codebook for silence includes not only code vectors for silence, but also code vectors for the onset of voice, since there is a possibility that the next frame's voice will be at the onset of voice.

2. If the decoded speech is a consonant, there is a high possibility that the speech in the next frame will be a vowel, so a codebook for vowels is selected. The vowel codebook includes not only vowel code vectors but also consonant code vectors since there is a possibility that the next frame of speech is a consonant.

3. Also, if a consonant continues for a certain amount of time,
Since there is a high possibility that the next frame's audio will be silent, a silent codebook is selected.

4. If the decoded speech is a vowel, there is a high possibility that the speech in the next frame will be a vowel, so a codebook for vowels is selected. The vowel codebook includes not only vowel code vectors but also consonant code vectors since there is a possibility that the next frame of speech is a consonant.

[0022] In this way, from the previous audio state,
By predicting the next speech state, the codebook m to be used for the next quantization is selected.

The index of the selected representative vector Ymi is sent to the decoder side by the transmitting means 5 (step 6).

On the decoder side, the receiving means 6 receives the index of the representative vector (step 11), and the codebook selecting means 9 selects the representative vector corresponding to the index data received by the representative vector reading means 7. Read from codebook m (step 12)
, speech synthesis (decoding) using the read representative vector
The means 8 decodes and outputs the audio (step 13). Further, the codebook selection means 9 selects a codebook to be used in the next frame from among the M groups in the same manner as on the encoder side (step 14).

As described above, according to the vector quantization method of the embodiment of the present invention, by providing a plurality of codebooks grouped according to the nature of the speech, it is possible to decide which codebook to use depending on the nature of the input speech. Since the arrangement is such that it is possible to select whether vector quantization is to be performed, quantization accuracy can be improved with a limited number of quantization bits and a limited amount of calculation.

[0026]

According to the present invention, as is clear from the description of the above embodiments, in the vector quantization method for quantizing digital signals such as A/D converted audio signals, grouping is performed according to the state of audio. The system uses multiple codebooks to predict the current frame audio state from the past frame audio states, determines which group of codebooks to use, and performs quantization using the codebook. By doing so, it is possible to provide a vector quantization method that can improve quantization accuracy even with a limited number of quantization bits and a limited amount of calculations.

[Brief explanation of drawings]

FIG. 1 is a functional block diagram showing a vector quantization method according to an embodiment of the present invention.

[Figure 2] Block diagram showing the hardware configuration of the method

FIG. 3: (A) is a flowchart representing encoder processing of the same method; (B) is a flowchart representing decoder processing of the same method;

[Figure 4] Functional block diagram showing a conventional vector quantization method

[Explanation of symbols]

Claims (2)

[Claims] Claim 1: A method comprising a plurality of codebooks grouped by speech state, predicting the speech state of the current frame from the speech state of past frames, and predicting the speech state of the current frame from the plurality of grouped codebooks. A vector quantization method that selects a corresponding codebook and performs quantization using the selected codebook. 2. Input vector calculation means for calculating a vector quantity of an input audio signal to be vector quantized, a plurality of first codebook means grouped according to the state of the audio, and a first codebook selection means described below. representative vector selection means for selecting a representative vector from the code vectors of said first codebook means selected in said first codebook means; audio local decoding means for decoding speech from said selected representative vector; encoding means including a first codebook selection means for predicting and selecting a codebook to be used in the next frame from the plurality of first codebooks; and a plurality of second codebooks grouped according to the state of speech. codebook means; representative vector reading means for reading out a representative vector corresponding to received data from a codebook selected by a second codebook selection means described later; speech decoding for decoding speech using the read representative vector; A vector quantization method comprising decoding means including second codebook selection means for predicting and selecting a codebook to be used in the next frame from the second codebook by referring to the decoded speech.