JP2001195094A - Voice coding method and voice decoding method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a voice coding method and a voice decoding method in which no band dividing scheme is used, the constitution/process is made simpler and the cost is reduced. SOLUTION: When the coding of voice signals is conducted, a plurality of predictors 12-1, 12-2,...12-n having respectively different characteristics with respect to the voice signals obtain the linear prediction values and predictive residuals of present signals from the past signals in the time region, select a predictor in which predictive residual is minimum among the output signals of the predictors and the output signals of the selected predictor are outputted with the signals that specify the predictor. During decoding, an appropriate predictor is selected using the signals that specify the predictor.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to audio signal encoding, and more particularly, to an audio encoding method and an audio decoding method for compressing and encoding the data amount of a digital audio signal.

[0002]

2. Description of the Related Art Various types of so-called high-efficiency encoding techniques for compressing, recording, and transmitting a data amount of an audio signal (audio signal) have been developed. Generally, sampled and uniformly quantized audio signals have redundancy, and if highly efficient encoding is performed, efficient transmission can be achieved with a smaller amount of information. There are roughly two types of high efficiency coding. One is irreversible coding (lossy coding), and "auditory coding" that controls so that data deterioration is not detected audibly is often used. The other is lossless encoding (lossless encoding), and the encoded / decoded data completely matches the original data. Lossless coding is commonly used in computer data compression, but is not often used in audio coding because a large compression ratio cannot be obtained. However, high-sampling audio, etc., which have recently attracted attention, are signals having greater redundancy than before, and lossless coding is considered to be effective.

In the prior art described in Japanese Patent Application Laid-Open No. 9-46234, filed by the applicant company, after dividing a digital audio signal into bands, the energy value obtained for each band is determined according to the amount of information. The sample value quantized by bit allocation is encoded as frequency domain information, the quantized sample value is inversely quantized, and the original signal is subjected to band synthesis in order to reconvert it into a time domain signal. The residual signal from the signal is encoded as information in the time domain, and the two are multiplexed to perform encoding for lossless compression.

[0004]

However, in the above-mentioned prior art, as described above, the configuration and processing for dividing the input signal into bands are complicated, and the amount of calculation is extremely large, causing an increase in cost. .

Accordingly, it is an object of the present invention to provide a speech encoding method and a speech decoding method which are simpler in configuration and processing and can reduce costs without using a band division technique.

[0006]

In order to achieve the above object, the present invention comprises the following means (1) to (3). That is,

1) A reference sample value is obtained in response to an input digital audio signal, and a linear prediction of a current signal is performed from a past signal in a time domain by a predetermined linear prediction method in response to the digital audio signal. Obtaining a prediction residual from the digital audio signal and the predicted value, selecting a linear prediction method in which the prediction residual is a minimum value in response to the digital audio signal, And an audio decoding method for decoding an original digital signal from a signal generated by multiplexing, in a predetermined format, the selected prediction residual and information specifying a linear prediction method of the selected prediction value. Separating the multiplexed signal, extracting the reference sample value, the selected prediction residual, and information identifying a linear prediction method of the selected prediction value; Using a single predictor corresponding to the selected prediction residual and calculating with a predictor capable of setting different prediction coefficients, and using information for specifying a linear prediction method of the selected prediction value. Setting a prediction coefficient of the single predictor at the time of decoding. 2) responding to the input digital audio signal and obtaining a reference sample value in frame units of a predetermined time; and responding to the digital audio signal, among the predicted values of the current signal predicted from the past in the time domain. Selecting a linear prediction method that minimizes the prediction residual in subframe units obtained by further dividing the frame, specifying the reference sample value, and a linear prediction method of the selected prediction value. Information, a prediction residual from the selected prediction value,
Multiplexing in a predetermined format in which prediction residuals from the selected prediction value are packed in subframe units, and decoding the original digital signal from the data encoded by the audio encoding method. Separating the multiplexed signal and extracting information identifying the reference sample value, the selected prediction residual, and a linear prediction method for the selected prediction value. Calculating a predicted value from the extracted reference sample value, the prediction residual, and a linear prediction method;
Restoring the original audio signal from the calculated predicted value. 3) responding to the two input digital audio signals and obtaining a reference sample value in frame units of a predetermined time; predicting a current signal predicted from the past in the time domain in response to the digital audio signals; Selecting a linear prediction method that minimizes the prediction residual among the values in subframe units obtained by further dividing the frame; the reference sample value; and a linear prediction method of the selected prediction value. And a prediction residual from the selected prediction value are multiplexed in a predetermined format including one header and bit number information according to the prediction residual, and the prediction residual is And packing the prediction residual based on the bit number information according to.

[0008]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Preferred embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram of a preferred embodiment of a speech encoding device (encoder) to which the present invention is applied. PCM as input audio signal
The data is sent, and the one-frame buffer 10 first operates as a reference sample acquiring unit, and sends the first sample value of each frame to the multiplexer 22. One frame buffer 10 further includes a plurality of predictors 12- having different characteristics.
1, 12-2,..., 12-n are sent in units of one frame. One frame is determined by the sampling frequency fs as shown in FIG.
Obtain 00 sample data. Here, the different characteristics are different prediction coefficients. Each predictor 12-1,
The output signals of 12-2,..., 12-n are output to the corresponding subtracters 14-1, 14-2,
.. 14-n, where a residual (difference) is detected. That is, the subtracters 14-1, 14-2,...
14-n operates as prediction residual detection means. This residual data is provided to the buffer / selector 16. In the present embodiment, the buffer / selector 16 has a function of obtaining a prediction residual for each subframe obtained by subdividing one frame. The time length of the subframe is one-tenth of one frame as shown in FIG. 2, and specifically, 0.1 to 0.5
ms.

A selection signal generator 18 connected to the buffer / selector 16 sub-generates a selection signal for selecting an output signal of a predictor having the minimum residual among these residual data. A selection signal for selecting an output signal of the predictor having the smallest residual is generated for each subframe, and is sent to the buffer / selector 16. That is, the selection signal generator 18
Sequentially scans the residual signal from each subtractor temporarily stored in the buffer / selector 16, finds the signal with the smallest residual for each subframe, and responds to the corresponding predictor next. The output signal of the subtractor is selected. The buffer / selector 16 sends only the selected signal to the packing circuit 20. Further, the selection signal generator 18 generates a bit number flag of each subframe and sends it to the packing circuit 20. The packing circuit 20 packs the selected residual signal with the number of bits specified by the bit number flag. I do.

[0010] The output signal of the packing circuit 20 is sent to a multiplexer 22, where the output signal of one frame buffer 10
, The number of bits flag (5 bits per subframe) and the predictor selection flag (3 bits per subframe) obtained by the selection signal generator 18, A bit stream of a predetermined format is generated together with the generated header (40 bits). The predictor selection flag is a signal for specifying the predictor having the minimum residual,
It is used when decoding an encoded audio signal. An audio decoding device for performing decoding will be described later with reference to FIG. FIG. 14 is a format diagram showing a data array for one frame after packing. This example shows a case where two channels (ch) of left and right (L, R) audio signals are compressed and recorded, and the number of samples in one frame is 1 per channel.
600, the number of subframes in one frame is 80 per channel. “Qb” in FIGS. 2 and 14 is the initial number of quantization bits.

Here, the principle of lossless speech (audio signal) encoding in the present invention will be described. If the audio signal to be encoded is a signal such as white noise, there is no redundancy, and lossless encoding does not work. However, real audio signals have several properties that lead to redundancy. The following are two typical examples.

1) As shown in FIG. 9, an audio signal often has energy deviated to a lower frequency side with respect to a reproduction frequency band. Further, as shown in FIG. 10, when viewed in the time domain, the waveform is smooth with respect to the interval between sampling points, and there is a large correlation between adjacent samples, so that prediction is possible. 2) As shown in FIG. 11, the signal amplitude of each section of the audio signal has a bias. Also, as shown in FIG.
There is also a bias in the amplitude appearance probability in the section, and generally, the smaller the amplitude, the higher the appearance probability. Naturally, there is a correlation between the contents of FIG. 11 and FIG.

As a method for improving the efficiency by using the property (first property) described with reference to FIGS. 9 and 10, there are a method of converting a signal into a frequency domain and normalizing the signal for each band, and a method of converting the signal into a time domain. A method is conceivable in which a linear prediction value of a current signal viewed from a past signal is calculated, and a residual between the prediction value and the actual measurement value is encoded (if the prediction is appropriate, the prediction residual amplitude is larger than the actual measurement value). Greatly reduced). When comparing the former and the latter, there is no great difference in the obtained effects, and the amount of calculation is several times larger in the former, so that it is more advantageous to use the time domain method. In audio coding, a frequency domain technique is often used to control a signal for each frequency band.

In order to use the property (second property) shown in FIG. 11, it is conceivable to divide the signal into sections of a fixed length and normalize the section with the maximum value for each section. In order to utilize the property of FIG. 12, entropy coding represented by Huffman coding is generally performed to optimize the code length allocated to data. Here, if the section length in FIG. 11 is shortened, the bias in the amplitude in the section and the bias in the appearance establishment in FIG. 12 are reduced, and the effect of entropy coding is also reduced.

According to the present invention, based on linear prediction coding in the time domain, a plurality of predictors having different characteristics are selectively used in accordance with an input signal (using the first property), and signals are compared. A linear predictive lossless audio encoding system characterized by dividing the data into short sections and performing the above-described processing for each section (using the second property) is established.

The operation of the embodiment shown in FIG. 1 for realizing the linear prediction lossless audio encoding according to the present invention will be further described together with the main specifications shown in FIG. In an encoding circuit, that is, an encoder, a plurality of predictors calculate linear prediction values / prediction residuals, select a predictor having the minimum residual for each subframe, and pack the result with the number of bits. Further, the main information and the auxiliary information are combined and multiplexed in a predetermined format for each frame to form a bit stream. Here, by performing the processing for each relatively short subframe, FIG.
The redundancy described in item 2 is effectively reduced. Further, since the total number of bits of each frame after compression generally changes depending on the redundancy of the signal, the operation is essentially performed at a variable transmission rate (variable bit rate). The signal encoded in the present embodiment, that is, the output signal of the multiplexer 22 can be recorded on an optical recording medium such as a CD-ROM or a DVD (digital versatile disk). Although not shown, a decoder for decoding a bit stream generated by the encoding device of the present invention includes:
Decoding can be performed for each independent frame. In this decoding, a prediction equivalent to that at the time of encoding can be performed using the predictor selection flag in the auxiliary information.

Here, the predictors 12-1, 1 in FIG.
Consider the case where a linear predictor is used for 2-2 and 12-n. That is, consider the case of reducing the redundancy in FIGS. 9 and 10 using linear prediction. An input sequence x
The n-order linear prediction value Pi for i is calculated by the following equation.

[0018]

(Equation 1)

Here, ak (k = 1, 2,..., N) is a prediction coefficient. The prediction residual di, which is the difference between the predicted value and the measured value, is

[0020]

(Equation 2)

It can be seen that di is a linear filter output for the xi series. In encoding, di
It is necessary to determine the order n and coefficient ak of the prediction filter so that is minimized. The coefficient ak may be an integer or a non-integer. FIG. 3 shows the frequency characteristics of a typical predictor using integer coefficients. If the coefficients are non-integer, more flexible control of the filter characteristics is possible. However, if the coefficient bit length becomes longer due to non-integerization, the operation word length of the predictor also increases, and care must be taken in managing the operation accuracy of the encoder / decoder. In the present embodiment, a total of six types of primary, secondary, and tertiary predictors based on integer coefficients and three types of fourth-order predictors of non-integer coefficients (in increments of 0.5) obtained by extending the integer coefficients by one bit are prepared. Selected used. An example of the non-integer coefficient used this time is as follows.

Fourth order: a1 = 3.5 a2 = −5.0 a3 = 3.5 a4 = −1.0

Next, evaluation of the compression ratio in the present embodiment will be described. The compression ratio of a plurality of audio sources was evaluated using the above-described encoding method. FIG. 4 shows an outline of the evaluation source. There are 20 songs of the conventional sampling frequency (fs = 44.1 kHz) and 11 songs of high bit / high sampling sources (each source has a length of about 3 to 5 minutes). In FIG. 4, source numbers are set according to source specifications and genres. First, FIG. 5 (source number NP6) and FIG. 6 (source number UC1) show an example of a temporal transition of the average number of bits per sample (bit rate) in each frame. As shown in the figure, the bit rate changes every moment depending on the tune. In addition, the conventional fs
In many sources, the compression effect at the peak is often small.

Next, the compression ratio of each source = (data size after compression) / (data size of source) × 100 [%]
Is shown. Figure 7 shows fs = 44.1kHz / Qb = 16 bit source, Figure 8 shows
fs = 88.2kHz, 96kHz / Qb = Compression rate of 20-bit source.
In general, it is the upper bits that contribute to the reduction of redundancy in lossless coding, and the additional bits (4 bits) are hardly compressed by 20-bit quantization with bits extended to the LSB side.
Therefore, it is necessary to consider the assumption that the compression ratio is deteriorated (increased) in the case of 20 bits.

Referring to the figure, in the case of the conventional fs = 44.1 kHz / Qb = 16 bit source, the compression ratio is about 44% to 83%, and the compression effect particularly in pop rock (NP *) is small. Assuming that these sources extend the number of bits to 20 bits, the compression ratio is estimated to be about 55% to 86%. Fs = 88.2kH
In the high sampling source of z, 96kHz, the energy in the high frequency range above fs / 4 is small, and the redundancy is expected to be large accordingly. In this example, the compression ratio is relatively good, about 44% to 62%. If a compression ratio of about 50% (in the case of 20 bits) can be expected, practical application is possible. Source UJ1 to UJ3
Has a poor S / N ratio due to the old analog master, and the compression ratio is degraded. Regarding the high-sampling source, although the possibility in the pop / rock system that could not be considered this time is unknown, lossless coding is considered to be effective.

Next, a preferred embodiment of a speech decoding apparatus for decoding in response to a variable rate bit stream generated by encoding according to the present invention will be described with reference to FIG. The audio decoding apparatus of FIG. 13 has a demultiplexer 24 for decomposing a variable rate bit stream which is an output signal of the multiplexer 22 of FIG. 1 (or a signal obtained by recording this output signal on a recording medium and reproducing) into signal components. ing. In the output signal of the demultiplexer 24, the frame head sample value is output as a part of the PCM data. In the output signal of the demultiplexer 24, the bit number flag indicates an unpacking circuit 26 for unpacking the audio signal.
Given to.

In the output signal of the demultiplexer 24, the predictor selection flag indicates a plurality of predictors 30-1, 30-2,.
Used to select one of 30-n. The output signal of the selected predictor is supplied to the adder 28, added to the output signal of the unpacking circuit 26, and output as PCM data. Since the predictor selection flag changes for each sub-frame as described in the description of the encoding, the predictors 30-1, 30-2,.
One of 30-n will be selected. Therefore, decoding can be performed in a procedure reverse to the encoding, and the original PCM
Data can be obtained without loss.

In the configuration shown in FIG. 13, a plurality of predictors are used, as in the speech coding apparatus of FIG. 1. However, at the time of decoding, the speed of encoding is not required. It is also possible to use a variable one and set a desired prediction coefficient by a predictor selection flag.

[0029]

As described above, according to the present invention, the compression efficiency is improved especially in the case of predictive coding of an audio signal, and the band division method is not used, so that the configuration processing is simpler and the cost is reduced. Can be provided.

[Brief description of the drawings]

FIG. 1 is a block diagram of a preferred embodiment of a speech encoding device to which the present invention has been applied.

FIG. 2 is a table showing coding specifications in the embodiment of FIG. 1;

FIG. 3 is a graph showing frequency characteristics of a typical predictor.

FIG. 4 is a list of test sources used for evaluating the compression efficiency of encoding according to the present invention.

FIG. 5 is an example of an evaluation result of compression efficiency of encoding according to the present invention.

FIG. 6 is another example of the evaluation result of the compression efficiency of encoding according to the present invention.

FIG. 7 is a graph showing a compression rate of encoding according to the present invention for each source having a sampling frequency of 44.1 kHz.

FIG. 8 is a graph showing a compression rate of encoding according to the present invention for each of sources having a sampling frequency of 88.2 kHz and 96 kHz.

FIG. 9 is a graph showing a frequency distribution of an audio signal.

FIG. 10 is a waveform diagram showing a change in amplitude of an audio signal with respect to time.

FIG. 11 is a diagram showing a change in amplitude of an audio signal for each section (segment);

FIG. 12 is a graph showing a relationship between an amplitude of an audio signal and an appearance probability.

FIG. 13 is a block diagram showing a preferred embodiment of a speech decoding device to which the speech decoding method of the present invention is applied.

FIG. 14 is a format diagram showing a data array for one frame after packing.

[Explanation of symbols]

10 1 frame buffer (reference sample value acquisition means) 12-1, 12-2,..., 12-n predictor 14-1, 14-2,. Reference Signs List 16 buffer / selector 18 selection signal generator (means for selecting a predictor / means for generating information specifying a predictor) 20 packing circuit (packing means) 22 multiplexer (means for multiplexing) 24 demultiplexer (means for extracting 26 unpacking circuit 28 adder 30-1, 30-2,... 30-n predictor

────────────────────────────────────────────────── ───

[Procedure amendment]

[Submission date] February 7, 2001 (2001.2.7)

[Procedure amendment 1]

[Document name to be amended] Statement

[Correction target item name] Claims

[Correction method] Change

[Correction contents]

[Claims]

[Procedure amendment 2]

[Document name to be amended] Statement

[Correction target item name] 0007

[Correction method] Change

[Correction contents]

1) A reference sample value is obtained in response to an input digital audio signal, and a current signal is converted from a past signal in the time domain by a plurality of linear prediction methods having different characteristics in response to the digital audio signal. its linear prediction value of
Respectively, obtaining a prediction residual from the digital audio signal and the linear prediction value, selecting a linear prediction method in which the prediction residual is a minimum value in response to the digital audio signal, A speech decoding method for decoding an original digital signal from a signal generated by multiplexing a sample value, the selected prediction residual, and information specifying a linear prediction method of the selected prediction value in a predetermined format. Separating the multiplexed signal, extracting the reference sample value, the selected prediction residual, and information identifying a linear prediction method of the selected prediction value; Calculating by a single predictor corresponding to the selected prediction residual and a predictor capable of setting different prediction coefficients, and decoding by using information specifying a linear prediction method of the selected prediction value Sometimes said simple Step a speech decoding method comprising setting a prediction coefficient of the predictor. 2) obtaining a reference sample value in units of frames of a predetermined time in response to the input digital audio signal; and a plurality of linear processors having different characteristics in response to the digital audio signal.
Linear prediction of current signal from past in time domain by measurement method
Each value is predicted, its predicted linear prediction value and the previous
Selecting a linear prediction method such that a prediction residual obtained from the speech signal is minimized in units of sub-frames obtained by further dividing the frame; and a step of selecting the linearity of the reference sample value and the selected prediction value. Multiplexing information specifying a prediction method and a prediction residual from the selected prediction value in a predetermined format in which prediction residuals from the selected prediction value are packed in subframe units; A speech decoding method for decoding an original digital signal from data encoded by a speech encoding method comprising: separating the multiplexed signal, the reference sample value, and the selected prediction residue. Extracting information that specifies a difference, a method of linearly predicting the selected predicted value, and predicting from the extracted reference sample value, the prediction residual, and the linear prediction method. Steps and, and restoring the original audio signal from the calculated prediction value, voice decoding method consists of calculating the value. 3) in response to the digital audio signals of two channels inputted, obtaining a reference sample value in frames of a predetermined time, in response to said digital audio signal, a plurality of different properties
Of the current signal from the past in the time domain by the linear prediction method
Each linear prediction is predicted and its predicted linear prediction
Selecting a linear prediction method such that a prediction residual obtained from the measured value and the audio signal is minimized in subframe units obtained by further dividing the frame; and the reference sample value and the selected prediction. Information identifying a linear prediction method of the value, a prediction residual from the selected prediction value,
In a predetermined format including one header and bit number information according to the prediction residual, and packing the prediction residual based on the bit number information according to the prediction residual; An audio coding method comprising:

Claims (3)

[Claims] 1. In response to an input digital audio signal,
Obtaining a reference sample value, and obtaining a linear prediction value of a current signal from a past signal in a time domain by a predetermined linear prediction method in response to the digital audio signal, and predicting the linear prediction value from the digital audio signal and the prediction value. Obtaining a residual, and selecting a linear prediction method in which the prediction residual is a minimum value in response to the digital audio signal, the reference sample value, the selected prediction residual, and the selected prediction A speech decoding method for decoding an original digital signal from a signal generated by multiplexing information specifying a linear prediction method of a value in a predetermined format, comprising: separating the multiplexed signal; Extracting the selected prediction residual and information specifying a linear prediction method of the selected prediction value; and a single predictor corresponding to the selected prediction residual. Calculating with a predictor capable of setting a prediction coefficient, and setting a prediction coefficient of the single predictor at the time of decoding using information specifying a linear prediction method of the selected prediction value, Voice decoding method. 2. A step of obtaining a reference sample value in units of a predetermined time frame in response to an input digital audio signal, and a step of predicting a current signal predicted from a past in a time domain in response to the digital audio signal. Selecting a linear prediction method that minimizes the prediction residual among the values in subframe units obtained by further dividing the frame; the reference sample value; and a linear prediction method of the selected prediction value Multiplexing the information identifying and the prediction residual from the selected prediction value in a predetermined format in which the prediction residual from the selected prediction value is packed in subframe units. A speech decoding method for decoding an original digital signal from data encoded by a speech encoding method, comprising: separating the multiplexed signal; Extracting the value, the selected prediction residual, and information specifying a linear prediction method of the selected prediction value; and the extracted reference sample value, the prediction residual, and the linear prediction method. A speech decoding method, comprising: calculating a predicted value from, and restoring an original audio signal from the calculated predicted value. 3. A step of responding to two input digital audio signals and obtaining a reference sample value in a frame unit of a predetermined time, and a step of responding to the digital audio signals and predicting a current time predicted from a past in a time domain. Selecting a linear prediction method in which the prediction residual is minimized among the prediction values of the signal in subframe units obtained by further dividing the frame; and the reference sample value and the selected prediction value. While multiplexing information specifying a linear prediction method and a prediction residual from the selected prediction value in a predetermined format including one header and bit number information according to the prediction residual, Packing the prediction residual based on bit number information corresponding to the prediction residual.