JP2797348B2 - Audio encoding / decoding device - Google Patents

Description

Description: BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech encoding / decoding apparatus used for digital communication, voice mail, and the like.

2. Description of the Related Art FIG. 6 (a) shows a conventional speech coding apparatus, and FIG.
FIG. 1B shows a conventional speech decoding apparatus.

In FIG. 6 (a), 14 is a predictor for calculating a prediction error and a short-term / long-term prediction filter coefficient from the audio signal A / D converted by the linear PCM, and 15 is a predictor, as shown in FIG.
A codebook 15a in which a plurality of signal sequences (representative vectors) of a predetermined length are stored in advance, and a vector quantizer 15b for vector-quantizing the prediction error from the predictor 14 and outputting the number of the representative vector , A quantization value of the short-term / long-term prediction filter coefficient from the predictor 14 and a quantizer 15
Is a multiplexer for multiplexing the numbers of representative vectors from.

In FIG. 6 (b), 17 is a separator for separating the short-term prediction filter coefficient and the long-term prediction filter coefficient from the encoder and the number of the representative vector, and 18 is the same code as the codebook 15a of the encoder. A dequantizer 19 having a book (not shown) and outputting a representative vector corresponding to the number from the separator 17 as a prediction error; 19 is a short-term prediction filter coefficient and a long-term prediction filter coefficient; This is a synthesizer that synthesizes an audio signal using a representative vector.

 Next, the operation of the above conventional example will be described.

In FIG. 6 (a), when an audio signal that has been A / D-converted by the linear PCM is input, the predictor 14 calculates a short-term prediction filter coefficient in order to remove a correlation between adjacent sample values of the audio signal. A short-term prediction error is obtained using a short-term prediction filter coefficient.

Further, in order to remove the periodic correlation of the sound source pitch of the audio signal, a long-term prediction filter coefficient is obtained from the short-term prediction error, and a prediction error is obtained from the long-term prediction filter coefficient.

The quantizer 15 calculates the square distance between the signal sequence of the prediction error and each representative vector of the codebook 15a, and outputs the number of the representative vector having the smallest value as a quantization value.

Therefore, the audio signal is transmitted from the multiplexer 16 to the decoder as data compressed to the short-term prediction filter coefficient and the long-term prediction filter coefficient and the representative vector number.

In FIG. 6B, the synthesizer 19 is a predictor of the encoder.
A filter having inverse characteristics to the fourteen filters is provided. Therefore, the representative vector can be decoded into a speech signal by a filter corresponding to the short-term / long-term prediction filter coefficient from the encoder.

However, the above-described conventional speech coding apparatus and speech decoding apparatus perform the same processing irrespective of the characteristics of voice such as voiced voice, unvoiced voice, and silent voice. Is not encoded according to the characteristics of the audio,
There is a problem that the quality of the decoded speech is not good and the coding efficiency cannot be improved.

The present invention solves such a problem, and an object of the present invention is to provide a speech encoding / decoding device capable of improving encoding efficiency.

Means for Solving the Problems In order to achieve the above object, a speech coding apparatus of the present invention classifies a speech signal into a voiced sound stationary part, a voiced sound transient part, an unvoiced sound, and the like. Outputs the prediction error of the voiced transient part and the compressed voiced stationary part which are compressed on the axis, and quantizes the unvoiced sound and the prediction error of the voiced stationary part and the voiced transient part. It is.

Further, in order to achieve the above object, the speech decoding apparatus of the present invention inversely quantizes the unvoiced sound and the like and the quantized values of the voiced stationary part and the voiced transient part. The voiced sound transient part is synthesized with the voice signal, and the synthesized voiced sound stationary part is expanded on the time axis.

Operation The present invention can quantize according to the characteristics of voices such as voiced voices, unvoiced voices, and voiceless voices according to the above configuration, so that the quality of decoded voices is good and the coding efficiency can be improved. it can.

Embodiments Hereinafter, embodiments of the present invention will be described with reference to the drawings. FIG. 1 (a) is a block diagram showing an embodiment of a speech encoding device according to the present invention, FIG. 1 (b) is a block diagram showing an embodiment of a speech decoding device according to the present invention, FIG. 2 is a detailed block diagram showing the predictor of FIG. 1 (a), FIG. 3 is a detailed block diagram showing the quantizer of FIG. 1 (a),
FIG. 4 is a waveform diagram of a general audio signal, and FIG.
5 is a flowchart for explaining the operation of the classifier of FIG.

In FIG. 1 (a), reference numeral 1 denotes a classifier for analyzing characteristics of an input audio signal and classifying the characteristics into a voiced sound stationary portion, a voiced sound transient portion, unvoiced sound, and no sound, as described later. 2
This is a time axis compressor that thins out a similar waveform of the voiced stationary part of the audio signal based on the determination data from the, and compresses it on the time axis.

Reference numeral 3 denotes a predictor for obtaining a prediction error and a short-term / long-term prediction filter coefficient based on the voiced sound of the audio signal. As shown in FIG. The short-term prediction filter 21 and a short-term prediction filter 21 for removing a correlation between adjacent sample values of voiced sounds.
A short-term prediction analyzer 22 for determining a short-term prediction filter coefficient of
A long-term prediction filter 23 for obtaining a prediction error based on a short-term prediction error from the short-term prediction filter 21 and a long-term prediction filter of the long-term prediction filter 23 using a short-term prediction error to remove a periodic correlation between voiced sound source pitches. It is composed of a long-term prediction analyzer (pitch predictor) 24 for obtaining coefficients, and a quantizer 25 for quantizing the short-term prediction filter coefficients and the long-term prediction filter coefficients, respectively.

Numeral 4 is a quantizer for vector-quantizing the prediction error from the predictor 3, for example, and as shown in FIG. 3, the quantizer 4 is for a voiced stationary part, a voiced transient part, and an unvoiced sound. , A codebook containing the representative vectors for silence
One of the codebooks 31 to 34 is selected based on the determination data of the audio signals classified by the classifier 1 and the codebooks 31 to 34, and the signal sequence of the prediction error from the predictor 3 and the two representative vectors of the codebook are selected. It is composed of a vector quantizer 35 that calculates the squared distance and outputs the number of the representative vector having the smallest value as a quantization value.

Reference numeral 5 denotes multiplexing for multiplexing the determination data of the audio signal classified by the classifier 1, the quantized values of the short-term / long-term predicted filter coefficients from the predictor 3, and the number of the representative vector from the quantizer 4. Switches 10 and 11 are switches that are switched according to the characteristics of the sound classified by the classifier 1.

In FIG. 1 (b), a separator that separates transmission data from the encoding device into speech signal determination data, quantized values of short-term / long-term prediction filter coefficients, and representative vector numbers.
7 has the same codebook (not shown) as the codebooks 31 to 34 of the encoder, selects the codebook according to the determination data of the audio signal from the separator 6, and corresponds to the number from the separator 6. An inverse quantizer 8 for outputting a representative vector to be output as a prediction error; a synthesizer 8 for synthesizing a voiced sound of an audio signal by using the short-term prediction filter coefficient and the long-term prediction filter coefficient and the representative vector from the inverse quantizer 7; Is a time axis expander that expands the voiced stationary part of the audio signal on the time axis,
Reference numerals 12 and 13 denote switches that are switched according to the determination data of the audio signal from the separator 6.

 Next, the operation of the above embodiment will be described.

In FIG. 1A, an audio signal that has been A / D converted by the linear PCM is input to an encoding device at fixed time intervals (frames).

The audio signal (analog signal) is, as shown in FIG.
It is composed of a silence, a relatively long unvoiced sound which is the beginning of a voice, a relatively short voiced sound transition part continuing between the unvoiced sound and the voiced sound stationary part, and a voiced sound stationary part.

In this voice signal, (1) the voiced sound stationary part has a characteristic that similar waveforms continuously show periodicity and the correlation between samples and pitch periods is strong, and (2) the voiced sound transient part has , Similar waveforms are not continuous, but have a characteristic that the correlation between samples and pitch periods is strong. (3) Unvoiced sound has a characteristic that the correlation between samples and pitch period is weak and the signal changes drastically. (4) Silence has a characteristic that the signal amplitude is small.

When such an audio signal is input, the classifier 1
As shown in the figure, the power Pw of the waveform in one frame is calculated (step 51), and it is determined whether the power Pw is equal to or more than a threshold (step 52). If the power Pw is less than the threshold, the frame is determined to be silent (step 53). If the power Pw is equal to or greater than the threshold, the frame is determined to be sound and the process proceeds to step 54 and below.

In step 55, the autocorrelation coefficient γ (i) of the waveform in one frame is calculated, and the maximum value γmax at the designated i is obtained from the calculated autocorrelation coefficient γ (i) (step 56). It is determined whether or not the maximum value γmax is equal to or larger than a threshold (step 57). If the maximum value γmax is less than the threshold,
The frame is determined to be an unvoiced sound part (step 58).

In step 60, it is determined whether or not the previous frame is a voiced sound. In the case of NO, the frame is determined to be a voiced sound transition section (step 61), and in the case of YES, the process proceeds to step 62 and subsequent steps.

In step 62, the pitch period Pn is calculated from the autocorrelation coefficient γ (i), and the rate of change ρ from the pitch period Pn-1 of the previous frame is calculated (step 63). Is determined (step 64). If the rate of change ρ is less than the threshold value, the frame is determined to be a voiced sound stationary part, and if the change rate ρ is equal to or greater than the threshold value, the frame is determined to be a voiced sound transient part.

The classifier 1 calculates four types of characteristics of the audio signal,
For example, it outputs 2-bit decision data to the switches, 10, 11, the quantizer 4, and the multiplexer 5.

In a frame determined to be a voiced sound stationary part, the switch 10 is switched to the time axis compressor 2 side, and in a frame determined to be a voiced sound transient part, the switch 11 is switched to the predictor 3
Switch to the side.

Accordingly, the silence and unvoiced sound of the audio signal are directly input to the quantizer 4 and quantized to the representative vector numbers of the codebook for silence 34 and the codebook 33 for unvoiced sound, respectively. The quantized signal is input to the quantizer 4 and quantized to the representative vector number of the voiced transient codebook 32. The voiced stationary part is compressed by the time axis compressor 2 and then quantized via the predictor 3. The voice signal is quantized into a representative vector number of the codebook 31 for voiced stationary part.

Here, in the case of encoding silence and unvoiced speech, since the predictor 3 is not used, the bits allocated to the prediction filter coefficients and the like are allocated to the bits of the quantizer 4 to keep the number of transmitted bits constant. .

The decision data of the audio signal is transmitted to the decoder together with the prediction filter coefficient and the number of the representative vector, so that the decoding device can decode the original audio signal, and also has characteristics such as voiced sound, unvoiced sound, and silence. Since the quantized value is decoded in accordance with, the quality of the decoded speech is good.

As described above, the audio encoding / decoding device of the present invention receives an audio signal, calculates the power in one frame, and if the power is less than a threshold, converts the frame to silence. If the value is less than the threshold, the frame is determined to be voiced, and the autocorrelation coefficient within one frame is calculated.If the value is less than the threshold, the frame is determined to be unvoiced. Is determined to be a voiced sound, it is determined whether or not the previous frame is a voiced sound, and if it is determined that the frame is not a voiced sound, the frame is determined to be a voiced sound transient portion, and it is determined that the frame is a voiced sound. Calculate the pitch cycle by the autocorrelation coefficient, calculate the rate of change from the pitch cycle of the previous frame, if the rate of change is less than the threshold, determine the frame as a voiced sound stationary part, In this case, a classification means is provided to determine the voiced sound transient part. Therefore, the speech signal can be accurately classified into a voiced sound stationary part, a voiced sound transition part, an unvoiced sound, and a silent sound.
Further, since vector quantization using a codebook is performed, encoding can be performed efficiently with less information transmission.

[Brief description of the drawings]

FIG. 1 (a) is a block diagram showing an embodiment of a speech encoding device according to the present invention, FIG. 1 (b) is a block diagram showing an embodiment of a speech decoding device according to the present invention, Fig. 2
FIG. 1 (a) is a detailed block diagram showing the predictor, FIG. 3 is a detailed block diagram showing the quantizer of FIG. 1 (a),
FIG. 4 is a waveform diagram of a general audio signal, and FIG.
FIG. 6 (a) is a flowchart for explaining the operation of the classifier, FIG. 6 (a) is a block diagram showing a conventional speech encoding device, and FIG. 6 (b) is a block diagram showing a conventional speech decoding device. FIG. 7 is a detailed block diagram showing the quantizer of FIG. 6 (a). 1 ... Classifier, 2 ... Time axis compressor, 3 ... Predictor, 4
..... quantizer, 5 ... multiplexer, 6 ... separator, 7 ...
Inverse Quantizer, 8: Synthesizer, 9: Time Axis Decompressor, 31 ...
... Codebook for voiced sound steady part, 32 ... Codebook for voiced transient part, 33 ... Codebook for unvoiced sound, 34 ... Codebook for silence.

Continuation of the front page (56) References JP-A-55-11616 (JP, A) JP-A-63-59127 (JP, A) JP-A-62-194299 (JP, A) JP-A-59-12499 (JP, A) JP-A-59-82608 (JP, A) JP-A-63-204300 (JP, A) (58) Fields investigated (Int. Cl. ⁶ , DB name) G10L 3/00-9/18 H03M 7/30 H04B 14/04 JICST

Claims (2)

(57) [Claims] An audio signal is input, the power in one frame is calculated, and if the power is less than a threshold, the frame is determined to be silent, and if the power is greater than the threshold, it is determined to be sound. The autocorrelation coefficient in one frame is calculated, and if the value is less than the threshold value, the frame is determined to be unvoiced. If the value is equal to or greater than the threshold value, the frame is determined to be voiced. If it is determined that it is not a voiced sound,
The frame is determined to be a voiced sound transition part, and when it is determined that the frame is a voiced sound, a pitch cycle is calculated by an autocorrelation coefficient, a change rate with respect to a pitch cycle of a previous frame is calculated, and the change rate is less than a threshold value. In the case of, the frame is determined to be a voiced stationary part, and if it is greater than or equal to a threshold, the classification means is determined to be a voiced transient part, and the voiced stationary part of the audio signal classified by the classification means is classified into a time axis. Time axis compression means for compressing upward, a voiced sound transient part of the audio signal classified by the classification means, and prediction means for outputting a prediction error of the voiced stationary part compressed by the compression means, and the classification means A speech encoding device having quantization means for quantizing an unvoiced sound of the classified speech signal, silence, and an output signal of a prediction error output by the prediction means; a voiced sound stationary unit from the speech encoding device; Vocal transients Dequantizing means for dequantizing the quantized value of the unvoiced sound, the unvoiced sound, and a voiced sound stationary part dequantized by the dequantizing means, a synthesizing means for synthesizing the voiced sound transient part into a voice signal, A speech encoding device comprising: a time axis extension unit for extending a voiced sound steady part synthesized by the synthesis unit on a time axis. 2. A quantization means comprising a codebook in which a voiced sound stationary part, a voiced sound transient part, an unvoiced sound, and a representative vector for unvoiced sound are stored, and a codebook based on speech signal determination data classified by the classification means. 2. The speech code according to claim 1, wherein a selected value is calculated, a square distance between a signal sequence of the prediction error from the prediction means and each representative vector of the codebook is calculated, and the number of the representative vector having the smallest value is quantized and output.・
Decryption device.