JPH08202396A - Voice prediction coding method - Google Patents

Abstract

PURPOSE: To unnecessitate a memory by which plural noise excitation sources are stored by selecting a noise excitation source with less quantity of arithmetic processing. CONSTITUTION: In a voice prediction coding method by which a voice is coded utilizing reproducing a voice by driving a synthetic filter 4 of which a quantized filter coefficient calculated based on sampling of plural input voices is set per frame unit by a time series vector component outputted from two excitation sources of a pitch excitation source 11 expressing a pitch component and a noise excitation source 31 expressing a noise component, a frame is further divided into plural sub-frames, the noise excitation source 31 is expressed by making a pulse for each sub-frame, and a position and amplitude of this pulse is quantized and coded for each sub-frame.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech predictive coding method and, more particularly, to a synthesis filter using two excitation sources, a pitch excitation source expressing a pitch component and a noise excitation source expressing a noise component. The present invention relates to a voice predictive coding method for encoding voice by utilizing driving and reproducing voice.

[0002]

2. Description of the Related Art VSELP, CS-CELP, and other methods are known as speech predictive coding methods for driving a synthesis filter using two excitation sources. These are IAGerson and
MAJasiuk: “Vector Sum Excited Linear Prediction
(VSELP) Speech Coding at 8 kbps ”, proc.ICASSP '90
, Pp.461-464, 1994, or A. Kataoka, T. moriya and d.
S.Hayashi: “An 8-kbit / s Speech Coder Based on Con
jugate Structure CEIP ”, proc.ICASSP '93, pp.592-
595, 1993, and others.

Now, a conventional example of the voice predictive coding method will be described with reference to FIG. First, the sampling value series of the input speech waveform input from the input terminal 1 is supplied to the filter coefficient determination unit 2, and the filter coefficient is calculated here. The filter coefficient calculated in the filter coefficient determination unit 2 is then supplied to the filter coefficient quantization unit 3, the filter coefficient is quantized here, and the quantized filter coefficient is supplied to the synthesis filter 4, where The synthesis filter is set.

The excitation source for exciting the synthesis filter 4 is 2
It consists of individual excitation sources. One of the excitation sources is a pitch excitation source 11 and the other is a noise excitation source 21. The pitch excitation source 11 is composed of a plurality of pitch period components, and the selected pitch period component candidate is multiplied by the pitch gain in the pitch gain multiplication unit 12 and output. The noise excitation source 21 is composed of a plurality of noise waveform components, and the selected noise excitation source candidate is multiplied by the noise gain in the noise gain multiplication unit 22 and output. The synthesis filter 4 is excited and driven by the sum of the output of the pitch gain multiplication unit 12 and the output of the noise gain multiplication unit 22.

The distortion calculating section 5 is arranged so that the distortion, which is the difference between the input speech input through the input terminal 1 and the synthetic speech output from the synthesis filter 4, is minimized. The excitation component candidate is selected, and at the same time, the optimum gain is set for each excitation component candidate. Code output unit 6
Is to encode and output the quantized filter coefficient supplied from the filter coefficient quantization unit 3 and each excitation component candidate selected in the distortion calculation unit 5, and these codes are output via the output terminal 7. Is transmitted.

A method of selecting an optimum noise excitation source candidate in the conventional example of the speech predictive coding method shown in FIG. 4 will be described with reference to FIG. FIG. 5 shows a circuit which constitutes a part of the distortion calculation section 5 of the conventional example. In this conventional example, the optimum noise excitation source candidate is selected by convolving the impulse response waveform of the synthesis filter having the time length equal to the frame length and the noise excitation source candidate, and the input having the time length equal to the frame length. It is necessary to perform the distortion calculation between the waveform obtained by removing the pitch component from the speech and the waveform of the convolution operation result as many as the number of noise excitation source candidates, and for this reason, it is necessary to perform a large number of arithmetic processings. is there. Then, a memory having a large storage capacity for storing a noise excitation source having a relatively long time length is required.

By the way, the noise excitation source candidate is stored in a memory, and the noise excitation source is expressed by a plurality of pulses without using the noise excitation source, and a voice that drives a synthesis filter by using two excitation sources. ACELP is a method of encoding
Is disclosed (R. Salami, C. Laflamme and JP.Ado
ul: “ACELP Speech Coder at 8 kbit / s with a 10msfr
ame: A Candidate for CCITT Standardization ”, proc.
See IEEE Workshop on Speech Coding, pp.23-24, 1993). In this method, a noise excitation source having a time length of 5 msec is represented by a pulse train of 4 or 5 constant amplitudes, and the noise excitation source is specified only by the pulse positions of these pulse trains. It eliminates the memory required for storage. However, since this method synthesizes a voice having the same long time as the input voice in order to specify the position of one pulse, it is necessary to perform a large number of arithmetic processes, and therefore, 4 to 5 are required. It is necessary to perform arithmetic processing a large number of times to specify each pulse.

[0008]

SUMMARY OF THE INVENTION The present invention provides a speech predictive coding method for driving a synthesis filter using two excitation sources, a pitch excitation source expressing a pitch component and a noise excitation source expressing a noise component. In the above, there is provided a speech predictive coding method for selectively setting a noise excitation source with a small amount of calculation without storing a noise excitation source candidate as a noise excitation source in a memory.

[0009]

A synthesis filter 4 in which a quantization filter coefficient calculated based on a plurality of samples of input speech is set represents a pitch excitation source 11 representing a pitch component and a noise component. In a voice predictive coding method for coding a voice by utilizing the time-series vector components output from the two excitation sources of the noise excitation source 31 to drive the voice for each frame to reproduce the voice, Further, it is divided into a plurality of subframes, a noise excitation source 31 is expressed by setting a pulse for each subframe, and the position and amplitude of this pulse are quantized and coded for each subframe. Configured method.

Then, a speech predictive coding method is constructed in which a plurality of pulse amplitudes for each subframe are collectively quantized and quantized.
In addition, we constructed a speech prediction coding method that quantizes multiple pulse positions for each subframe. Furthermore, we constructed a speech predictive coding method that quantizes multiple pulse amplitudes and pulse positions for each subframe.

Here, the synthesis filter in which the quantized filter coefficient calculated based on a plurality of samples of the input speech is set as a pitch excitation source expressing a pitch component and a noise excitation source expressing a noise component. In the audio predictive coding method in which the audio is encoded by using the time-series vector component output from each excitation source to drive the audio for each frame to reproduce the audio, the frame is further divided into a plurality of subframes. The noise excitation source is expressed by dividing and arranging one pulse with a constant amplitude for each subframe, and the position of this pulse is quantized for each subframe.
Furthermore, we constructed a speech predictive coding method that determines the gain of the noise excitation source for each frame, quantizes it, and encodes the pulse position and amplitude. Then, a speech predictive coding method for collectively quantizing a plurality of pulse positions for each subframe is configured.

[0012]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENT An embodiment of the present invention will be described with reference to FIG. First, the sampling value series of the input speech waveform input from the input terminal 1 is supplied to the filter coefficient determination unit 2,
Here the filter coefficients are calculated. The filter coefficient calculated in the filter coefficient determination unit 2 is
The filter coefficient is supplied to the filter coefficient quantizer 3, the filter coefficient is quantized here, and the quantized filter coefficient is supplied to the synthesis filter 4, and the synthesis filter is set.

The excitation source for exciting the synthesis filter 4 is 2
It consists of individual excitation sources. One of the excitation sources is the pitch excitation source 11. The pitch excitation source 11 is composed of a plurality of pitch period components, and the selected pitch period component candidate is multiplied by the pitch gain in the pitch gain multiplication unit 12 and output. The other of the excitation sources is a pulse generation source 31, which generates one pulse for each subframe which is a section obtained by dividing the input speech frame into a plurality of sections.

The distortion calculator 5 selects a pitch cycle component candidate from the pitch cycle components of the pitch excitation source 11 so that the distortion, which is the difference between the input speech and the synthesized speech, becomes the smallest, and the pitch cycle component is selected. The optimum gain is set for the candidate, and at the same time, the position and amplitude at which the pulse generated in the pulse generation unit 31 should be generated are set. The specific method of setting the position and amplitude of the pulse to be generated in the distortion calculation section 5 is as follows.

The sampling frequency of the input voice is 8.0 kHz,
Encoded frame length is 10msec and number of subframes is 1
If there are 0 frames, there are 8 positions where pulses should be generated in each subframe. Here, in the state where all parameters other than the pulse position and pulse amplitude of the noise excitation source to be set and the input speech frame are locked, the pulse position and the pulse amplitude are sequentially applied from the first subframe to the tenth subframe. Perform the setting operation of. First, for the first sub-frame, in the above-mentioned state, a pulse is set at a first position among the eight positions where a pulse should be generated, the pulse amplitude is changed, distortion calculation is performed, and the result is stored. Then, subsequently, a pulse is set at the second position out of the eight positions, the pulse amplitude is changed, the distortion is calculated, and the result is stored.
Also, distortion calculation is performed for the eighth position and the calculation result is stored. All the stored 1st to 8th position and pulse amplitude distortion calculation results are compared, and only one position and pulse amplitude showing the minimum value are set as the pulse position and pulse amplitude of the first sub-frame. Similarly, for the second subframe, in the above-described state, the distortion calculation of the position and the pulse amplitude of the eight positions where the pulse should be generated is performed, the results are stored, and these are compared to show the minimum value. The position and pulse amplitude are the pulse position and pulse amplitude of the second subframe. Similarly, for the third to tenth subframes, one pulse position and one pulse amplitude are set. When the pulse position and the pulse amplitude are set one by one for all the first to tenth subframes, the expression of the noise excitation source is completed.

The code output unit 6 includes a filter coefficient quantization unit 3
The quantized filter coefficient supplied from the distortion calculation unit 5
The position and amplitude of each pulse of the pitch excitation source candidate and the noise excitation source selected in 1 are encoded and output, and these symbols are transmitted through the output terminal 7. According to the present invention, by using the pulse for the noise excitation source, it is possible to eliminate the need for performing the convolution calculation of the impulse response waveform of the synthesis filter and the noise excitation source candidate, which is very heavy in the calculation processing in the conventional example. This will be described with reference to FIG. 2, which is a circuit forming a part of the distortion calculation section 5 of the present invention. That is, in FIG.
Instead of performing the above-mentioned convolution operation, the shift register is used to read the impulse response waveform of the synthesis filter input to this with a pulse indicating the position, and this is assumed to correspond to the convolution operation result in the conventional example. The processing load can be reduced. Furthermore, the frame length L
Since the pulse is specified for each sub-frame having a shorter time length L ′ as compared with, it is possible to reduce the calculation processing load for distortion calculation.

Here, by quantizing the position and / or the amplitude of each pulse of the noise excitation source in the code output section 6 collectively, the efficiency of the quantization can be improved. Another embodiment of the present invention will be described with reference to FIG. In FIG. 3, parts corresponding to those in FIG. 1 are given the same reference numerals. The amplitude of the pulse generated by the pulse generation unit 41 constituting the noise excitation source is constant, and the pulse generation unit 41 specifies the noise waveform component only by the position of the pulse. The output of the pulse generation unit 41 is multiplied by the noise gain in the noise gain gain multiplication unit 42 and supplied to the synthesis filter 4 together with the pitch excitation source candidate. The code output unit 6 encodes and outputs the quantized filter coefficient supplied from the filter coefficient quantization unit 3, the selected pitch excitation source candidate, the position of each pulse of the noise excitation source candidate, and the noise gain, respectively.

The code output unit 6 can improve the efficiency of quantization by collectively quantizing the positions of each pulse of the noise excitation source. According to the present invention, in order to select a noise excitation source, it suffices to synthesize a voice for a short time corresponding to a subframe obtained by dividing a voice frame. It is significantly reduced compared to the method. Furthermore, since all noise excitation sources are expressed by pulses and the noise excitation sources are encoded with respect to the position and amplitude of the pulse, a memory for storing noise excitation source candidates is unnecessary.

The sampling frequency of the input voice is 8.0 kHz,
Under the condition of the coded frame length of 10 msec, the size of the code length of the noise excitation source of the conventional coding method is set to 256, and in the present invention, the voice frame is divided into 10 and the subframe length is set to 1 msec. Comparing the amount of calculation required to specify the noise excitation source when the number of pulses is 10 pulses, the amount of calculation of the present invention can be reduced to 10% or less as compared with the conventional encoding method. It is estimated to be. Furthermore, the present invention does not require a memory to store the noise excitation source required in conventional coding methods.

[0020]

As described above, according to the present invention, since the noise excitation source having an extremely short time length is expressed by using a single pulse, a short time length is selected in selecting the noise excitation source. It is only necessary to synthesize the voice, and the noise excitation source can be selected with a small amount of calculation processing. Furthermore, since the noise excitation source is represented by a pulse, the noise excitation source can be specified by the position and amplitude of the pulse, so that a memory for storing a plurality of noise excitation sources can be eliminated.

[Brief description of drawings]

FIG. 1 is a block diagram illustrating an embodiment of the present invention.

FIG. 2 is a block diagram which constitutes a part of a distortion calculation section of the present invention.

FIG. 3 is a block diagram illustrating another embodiment of the present invention.

FIG. 4 is a block diagram illustrating a conventional example.

FIG. 5 is a circuit forming a part of a distortion calculation unit of a conventional example.

[Explanation of symbols]

 1 Input Terminal 2 Filter Coefficient Determining Section 3 Filter Coefficient Quantization Section 4 Synthesis Filter 5 Distortion Calculating Section 6 Code Output Section 7 Output Terminal 11 Pitch Excitation Source 12 Pitch Gain Multiplying Section 31 Pulse Generation Source

Claims (6)

[Claims] 1. A synthesis filter in which a quantization filter coefficient calculated based on a plurality of samples of input speech is set, and two synthesis filters are provided: a pitch excitation source expressing a pitch component and a noise excitation source expressing a noise component. In the audio predictive coding method in which the audio is encoded by using the time-series vector component output from the excitation source to drive the audio for each frame to reproduce the audio, the frame is further divided into a plurality of subframes. Then, a noise excitation source is expressed by raising one pulse for each subframe, and the position and amplitude of this pulse are quantized and encoded for each subframe. 2. The speech predictive coding method according to claim 1, wherein a plurality of pulse amplitudes for each subframe are collectively quantized and quantized. 3. The speech predictive coding method according to claim 1, wherein a plurality of pulse positions for each subframe are collectively quantized and quantized. 4. The speech predictive coding method according to claim 1, wherein a plurality of pulse amplitudes and pulse positions for each subframe are collectively quantized. 5. A synthesis filter in which a quantized filter coefficient calculated based on a plurality of samples of input speech is set, and two synthesis filters are provided: a pitch excitation source expressing a pitch component and a noise excitation source expressing a noise component. In the audio predictive coding method in which the audio is encoded by using the time-series vector component output from the excitation source to drive the audio for each frame to reproduce the audio, the frame is further divided into a plurality of subframes. Then, a noise excitation source is expressed by raising one pulse with a constant amplitude for each subframe, the position of this pulse is quantized for each subframe, and the gain of the noise excitation source is determined for each frame. And a pulse position and amplitude are coded, and a speech predictive coding method is characterized. 6. The speech predictive coding method according to claim 5, wherein a plurality of pulse positions for each subframe are collectively quantized and quantized.