JP2560860B2 - Multi-pulse type speech coding and decoding device - Google Patents

Description

TECHNICAL FIELD The present invention relates to a multi-pulse type speech code decoding apparatus.

[Prior Art] The input voice signal is analyzed, the spectrum envelope information and the sound source information that constitute the voice information of the input voice signal are extracted on the analysis side, and these voice information are transmitted to the synthesis side via the transmission path. However, a multi-pulse type voice coder / decoder for synthesizing voice has been well known recently.

The spectral envelope information represents spectral distribution information of the vocal tract system that generates the input speech signal, and is usually a number of LPC coefficients corresponding to the analysis order obtained by linear predictive coding (LPC) analysis, for example, α parameter, K parameter, etc. are used.

On the other hand, the sound source information indicates a fine structure of the spectrum envelope, and is called a residual signal obtained by removing the spectrum envelope information from the input speech signal. Pulse).

The multi-pulse type speech coding / decoding apparatus obtains the spectrum envelope information and the excitation information, quantizes them based on a predetermined quantized bit allocation rule, and transmits the quantized information.

[Problems to be Solved by the Invention] The above-described conventional multi-pulse type speech code decoding apparatus has the following problems. In the conventional multi-pulse type speech codec / decoder, the number of multi-pulses, which is excitation information, is usually a predetermined fixed number. As described above, this sound source information expresses the fine structure of the spectral envelope. When the input speech is voiced, it shows a pitch structure, and when it is unvoiced, it has no clear pitch structure and is random. It is known to be close to a pulse.

If the input voice is voiced and the number of multi-pulses is fixed, the following problems occur. That is, the pitch period changes in the range of approximately 50 Hz to 500 Hz depending on the difference in the speaker. That is, the number of pitch period sections included in the analysis frame having a constant section length changes.

Nevertheless, if the number of multi-pulses in the analysis frame is fixed, the number of multi-pulses assigned to each pitch section in the analysis frame may not be the same.

If the bit rate is high and the number of multipulses allocated in the analysis frame is large enough, it does not matter so much even if the number of multipulses allocated to each pitch section is different. This is because each pitch section is assigned with a multi-pulse that almost represents the speech waveform in that section.

However, when the bit rate is low and the number of multi-pulses allocated in the analysis frame is small, it can be said to be a fatal problem. This is because, in a certain pitch section in the analysis frame, it is not possible to express the outline of the voice waveform in that section, and as a result, the sound quality is greatly deteriorated.

This problem can be avoided by making the number of multi-pulses in the analysis frame variable so that the number of multi-pulses in each pitch section in the analysis frame is the same and expressing the voice waveforms in each pitch section in a substantially balanced manner. it can.

Therefore, an object of the present invention is to improve the sound quality by varying the number of multi-pulses in the analysis frame and expressing the voice waveforms of the respective pitch sections in a balanced manner when the input voice is a voiced sound. It is an object to provide a multi-pulse type speech code decoding apparatus.

[Means for Solving the Problem] In a multi-pulse speech code decoding apparatus, a method based on correlation processing is known as a multi-pulse analysis method for obtaining the amplitude and position of multi-pulse which is excitation information. This method calculates the cross-correlation coefficient of the impulse response waveform of the LPC synthesis filter that represents the spectral envelope and the weighted waveform that is subjected to the perceptual weighting process called noise shaping on the input speech waveform, and the absolute value of the correlation value. The pulse is set at a large point. In this case, each time one pulse is set, the cross-correlation coefficient is corrected using the auto-correlation coefficient of the impulse response waveform. After the correction, as in the previous case, the part where the absolute value of the cross-correlation coefficient is large is searched for and the pulse is set.
When the pulses are sequentially set in this manner, the value of the cross-correlation coefficient gradually decreases. Finally, when the cross-correlation coefficient becomes sufficiently small, the speech waveform of the analysis frame is sufficiently expressed. In other words, it can be seen that the cross-correlation coefficient can be used as an index of how much the speech waveform is represented by the multi-pulses set up to that point.

Here, the cross-correlation coefficient power Pφ (n) is defined as follows, where φ (n, i) is the cross-correlation coefficient of the i-th tap after the setting of the n-th pulse.

In addition, the cross-correlation coefficient power Pφ after setting the nth pulse
The power difference D (n) between (n) and the cross-correlation coefficient power Pφ (n−1) after setting the (n−1) th pulse is defined as follows.

D (n) = P [phi] (n) -P [phi] (n-1) (n = 1,2,3, ...) (2) The power difference D (n) is calculated from (n-1) pulse to n. It can be said that it indirectly expresses the degree of sound quality improvement when the pulse is used. Therefore, it can be said that the pulse number is not sufficient while the power difference D (n) is large. As described above, when the input voice is a voiced sound, the power difference D (n) is large until the outline of each pitch section is expressed. Therefore, the number of multi-pulses can be set by calculating the difference D (n) of the cross-correlation coefficient power.

The multi-pulse type speech code decoding apparatus according to the present invention comprises:
A cross-correlation coefficient power calculating means for calculating the cross-correlation coefficient power according to the equation (1), a power difference calculating means for calculating the power difference according to the equation (2), and an optimum number of multi-pulses based on the power difference. And a multi-pulse number determining means for determining.

EXAMPLES Next, the present invention will be described in detail with reference to the drawings.

FIG. 1 is a block diagram showing the configuration of a multi-pulse type speech code decoding apparatus according to an embodiment of the present invention. First
The multi-pulse type speech coding / decoding device of the embodiment shown in the figure is composed of an analyzing side, that is, a multi-pulse type speech coding device and a synthesis side, that is, a multi-pulse type speech decoding device.

On the analysis side, the waveform memory 1, the LPC analyzer 2, the K / α converter 3, the attenuation α calculator 4, the impulse response calculator 5, the autocorrelation coefficient calculator 6, the perceptual weighting waveform calculator 7, the mutual phase relationship Number calculator 8, cross-correlation coefficient corrector 9, multi-pulse searcher 10, cross-correlation coefficient power calculator 11, difference power calculator 12, pulse number determiner 13,
The quantizer pattern output device 14, the K parameter quantizer 15, the pulse quantizer 16, and the multiplexer 17 are included.

Also, on the combining side, the demultiplexer 18, the decoding pattern output device 19, the K parameter decoding device 20, the pulse decoding device 21, K /
It is composed of an α converter 22 and an LPC synthesis filter 23.

Input voice waveform is in units of analysis frame, eg 20msec
Is cut out and stored in the waveform memory 1. The voice waveform accumulated in the waveform memory 1 is supplied to the LPC analyzer 2, and LPC analysis is performed by the well-known Levinson method or the like to calculate the K parameter. The obtained K parameter is converted into an α parameter by the k / α converter 3, and then sent to the attenuation α calculator 4 to calculate the attenuation α parameter. The attenuation α parameter is obtained for a perceptual weighting process called noise shaping, and is intended for a so-called masking effect. In the impulse response calculator 5, the impulse response waveform of the LPC filter represented by the attenuation α parameter is, for example, 5 msec.
Minutes are calculated.

The perceptual weighting waveform calculator 7 calculates the perceptual weighting waveform using the α parameter, the attenuation α parameter, and the input voice waveform. The cross-correlation coefficient calculator 8 calculates the cross-correlation coefficient between the impulse response waveform and the perceptual weighting waveform. The multi-pulse searcher 10 searches for the maximum point of the absolute value of the cross-correlation coefficient and sets a pulse at that point. The pulse amplitude is calculated from the autocorrelation coefficient of the impulse response waveform obtained by the autocorrelation coefficient calculator 6 and the crosscorrelation coefficient of the pulse setting position. The cross-correlation coefficient corrector 9 calculates the correction amount using the set pulse amplitude / position and the auto-correlation coefficient, and corrects the cross-correlation coefficient in the cross-correlation coefficient calculator 8. This series of multi-pulse search processing is performed a number of times corresponding to a predetermined maximum number of multi-pulses.

On the other hand, the cross-correlation coefficient power calculator 11 calculates the sum of squares of the cross-correlation coefficient, that is, the cross-correlation coefficient power, each time a pulse is set and the cross-correlation is corrected. The difference power calculator 12 calculates the difference between adjacent powers obtained by the cross-correlation coefficient power calculator 11. The pulse number determiner 13
The most appropriate number of pulses is determined based on the differential power.

Specifically, the difference power corresponding to a predetermined minimum pulse number and maximum pulse number is observed, and the pulse number obtained by subtracting 1 from the pulse number at which the difference power sharply decreases is taken as the optimum pulse number. This is because the fact that the differential power has decreased sharply means that the degree of improvement due to the setting of the pulse has decreased.

There are various conceivable methods for determining the sudden decrease of the differential power, but here the determination is made based on whether the differential power falls below a certain threshold. If it does not fall below the threshold, the maximum number of pulses is set in advance.

The quantization pattern output unit 14 outputs a predetermined quantization pattern corresponding to the number of pulses determined by the pulse number determiner 13 to K.
The labels are output to the parameter quantizer 15 and the pulse quantizer 16, respectively, and the label corresponding to the quantized pattern is sent to the multiplexer 17.

The K parameter quantizer 15 is the K obtained by the LPC analyzer 2.
The parameter is quantized based on the quantization rule input from the quantization pattern output device 14. The pulse quantizer 16 is
The set amplitude and position of the multi-pulse is quantized based on the quantization rule as in the K parameter. The multiplexer 17 multiplexes the quantized K parameter, the multi-pulse amplitude / position and the label of the quantized pattern, and sends out to the transmission path.

On the other hand, on the synthesis side, the demultiplexer 18 separates the multiplexed data into a quantized pattern label, K parameter, and multipulse. Based on the decoding pattern rules output from the decoding pattern output device 19, the K parameter and the multi-pulse are respectively K parameter decoding device 2
0, decoded by the pulse decoder 21. The multipulse decoded by the pulse decoder 21 is an LPC synthesis filter.
Sent to 23. The K parameter decoded by the K parameter decoder 20 is converted into an α parameter by the K / α converter 22 and sent to the LPC synthesis filter 23. LPC synthesis filter 23
Uses the α parameter as the filter coefficient and multipulse as the input of the filter to synthesize and output speech.

[Effects of the Invention] As described above, according to the present invention, there is an effect that the sound quality can be improved by determining the optimum number of pulses by using the difference in cross-correlation coefficient power.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of a multi-pulse type speech code decoding apparatus according to an embodiment of the present invention. 1 …… Waveform memory, 2 …… LPC analyzer, 3 …… K / α converter, 4
…… Attenuation α calculator, 5 …… Impulse response calculator, 6 …… Autocorrelation coefficient calculator, 7 …… Audition weighting waveform calculator, 8 ……
Cross-correlation coefficient calculator, 9 …… Cross-correlation coefficient corrector, 10 ……
Multi-pulse searcher, 11 ... Cross-correlation coefficient power calculator, 12
…… Differential power calculator, 13 …… Pulse number determiner, 14 …… Quantization pattern output device, 15 …… K parameter quantizer, 16 …… Pulse quantizer, 17 …… Multiplexer, 18 …… De Multiplexer, 19 ... Decoding pattern output device, 20 ... K parameter decoder, 21 ... Pulse decoder, 22 ... K / α converter, 23
...... LPC synthesis filter.

Claims (2)

(57) [Claims] 1. An LPC coefficient extracted by performing LPC analysis on an input speech signal for each analysis frame is used as spectrum envelope information, and sound source information that constitutes voice information of the input speech signal together with this spectrum envelope information is analyzed for each analysis frame. What is claimed is: 1. A multi-pulse type speech encoding device which expresses a plurality of multi-pulses having an occurrence time position and an amplitude corresponding to the characteristics of the sound source information to analyze the input speech signal, wherein Means for determining the optimum number of pulses of the multi-pulse for each analysis frame based on the difference in power of the cross-correlation coefficient between the weighted waveform subjected to the perceptual weighting process and the impulse response waveform represented by the LPC coefficient Having the spectral envelope information, the excitation information, and pulse number information representing the optimum pulse number as an encoded signal. A multi-pulse type speech encoding device which is characterized by transmitting. 2. A multi-pulse type speech decoding device comprising a synthesizing means for synthesizing the input speech signal from the coded signal transmitted from the multi-pulse type speech coder according to claim 1. Device.