JP3468184B2 - Voice communication device and its communication method - Google Patents

Description

Detailed Description of the Invention

[0001]

The present invention relates to a linear predictive coding voice communication device using (LPC) analysis and synthesis system (transmitter and receiver), and to a communication method thereof.

[0002]

2. Description of the Related Art A pitch excitation type LPC vocoder has been conventionally known as an LPC analysis / synthesis type voice communication apparatus using an LPC coefficient and a residual signal. FIG. 3 shows a block diagram of an example of a pitch excitation type LPC vocoder as the conventional voice communication device (transmitting device and receiving device).

In the figure, an input voice signal on the transmitter side is limited to a telephone voice band of, for example, 300 Hz to 3.4 kHz by a voice band limiting low pass filter (LPF) 22, and then is input to an A / D converter 23. The audio data is supplied and converted into audio data having a predetermined number of quantization bits sampled at a predetermined sampling frequency.

This voice data is a linear predictive analyzer (LP
C analyzer 24, where it is converted into about 8 to 12 LPC coefficients such as k parameter and α parameter by known linear prediction analysis. The sound source analyzer 25 is this L
The residual signal is extracted by a known method using the PC coefficient, and the squared sum of the extracted residual signal is calculated to calculate the sound source signal (power). Also,
The above voice data is supplied to the pitch extractor 26,
The pitch frequency (vocal chord frequency) of the sound source data is extracted.

L output from the above LPC analyzer 24
The PC coefficient, the sound source signal (power) output from the sound source analyzer 25, and the pitch frequency output from the pitch extractor 26 are respectively supplied to a multiplexer 27, where they are multiplexed and then transmitted. It is sent to the combining side (receiving device side) via the path.

On the receiver side, the LPC coefficient, the sound source signal (power) and the pitch frequency are separated from the signal input by the separator 28. The pulse train generator 29 generates an impulse train corresponding to the pitch frequency from the separator 28. The voiced / unvoiced deciding unit 30 decides a voiced voice or an unvoiced voice based on the pitch frequency and power from the separator 28, and supplies the decision result to the switch circuit 32 as a switching signal.

The switch circuit 32 selects an impulse train extracted at a fixed period (fixed interval) corresponding to the pitch period from the pulse train generator 29 on the basis of the above-mentioned switching signal when the voiced voice is judged, and the sound source demodulator 33. Since there is no vocal cord vibration and there is no pitch frequency at the time of unvoiced voice determination, instead of the impulse train excited in a fixed cycle based on the pitch frequency, the noise generator 31 is supplied based on the switching signal. A random pulse train corresponding to the extracted white noise signal is selected and supplied to the sound source demodulator 33.

The sound source demodulator 33 demodulates the sound source signal based on the pulse train from the switch circuit 32 and the power separated by the separator 28. The LPC synthesis filter 34 is excited by the demodulated sound source signal, the coefficient is controlled by the LPC coefficient separated by the separator 28, and the digital synthesized speech signal is output.

This digital synthesized voice signal is supplied to the D / A converter 35 and converted into an analog signal, and then L
Unnecessary frequency components are removed by the PF 36, and a voice signal in the telephone voice band is output.

As described above, in the conventional voice communication apparatus, when synthesizing the voiced voice, the sound source information is set to the telephone voice band (300
(Hz to 3.4 kHz), the impulse train is generated from the pulse train generator 29 at a fixed cycle corresponding to the pitch frequency obtained as a result.

[0011]

However, the natural natural (voiced) vocal cord vibration is not constant (fixed period), but quasi-periodic with fluctuations in time and frequency and instantaneous fluctuations according to changes in the real voice. Behaves like However, in the above conventional voice communication device, the sound source information (pitch frequency)
In analyzing the above, ignoring the frequency fluctuations and temporal fluctuations of the real voice described above, the telephone voice band (300H
(z to 3.4 kHz) is regarded as a stationary signal, and the whole band is analyzed collectively, so that the temporal change due to the fluctuation of the real voice,
There is a problem that frequency fluctuations cannot be expressed, and therefore only mechanically synthesized speech with poor naturalness can be generated.

In general, a drawback of frequency analysis is that it may detect a signal of a constant multiple of that frequency, and even in the pitch frequency analysis, that frequency (double pitch) or half frequency (half pitch) is detected. ) Is detected, a synthetic voice with deteriorated naturalness of voice quality may be generated.

Conventionally, in-band signals of a voice signal (residual signal) are collectively judged to be voiced / unvoiced, but all in-band signals are not in the same state (voiced / unvoiced) and the frequency is not. Voiced and unvoiced sounds may be mixed in each band. Further, even in the case of voiced voice, the pitch period may differ for each frequency band.

Further, in the above-mentioned conventional voice communication device, L
Since the PC analysis collectively expresses the in-band spectrum, with the normally used 12 to 12 LPC coefficients,
Since it is assigned to the low frequency band where energy is concentrated and the representation accuracy of the high frequency band becomes insufficient, the underestimation of the formant bandwidth and the approximation of the high order (3rd) formant are poor and It may not be possible to reproduce a different spectrum.

Furthermore, the sound source signal is not a stationary signal in practice, but is accompanied by fluctuations, and the fluctuation width is different for each frequency band.

In order to improve the spectrum accuracy,
Although it is possible to increase the LPC coefficient, an increase in the LPC coefficient leads to an increase in the amount of communication information.
It is not desirable to use more than 12 LPC coefficients, and it is currently difficult to use more than 12 LPC coefficients in a voice communication device that actually performs narrow band communication.

The present invention has been made in view of the above points,
Reproduces more natural speech synthesis and LPC
The amount of information is compressed by vector quantization of the coefficient and the sound source information, and only one representative LPC coefficient having a high degree of mutual similarity is selected, and the amount of information is further compressed without increasing the LPC coefficient (information It is an object of the present invention to provide a voice communication device and a communication method thereof capable of reproducing a more faithful spectrum (high quality voice) without increasing the amount).

[0018]

A voice communication apparatus of the present invention.
Inputs a voice signal and outputs it as a coded voice signal to the transmission line.
Voice communication transmitting device and the sound through the transmission line.
Is it an encoded voice signal that is connected to the voice communication transmitter and input?
And a voice communication receiving device for reproducing a synthetic voice signal from
In a voice communication device; the voice communication transmitting device;
Divide the defined voice band into the first number of bands,
Framed with a predetermined period for each of the divided bands
Output a linear prediction coefficient by performing a linear prediction analysis on the input speech signal
Linear predictive analysis means; and output from the linear predictive analysis means
Vector quantization of the linear prediction coefficient for each divided band
First quantizing means for converting; the line for each of the same divided bands
The residual signal is extracted by receiving the shape prediction coefficient and the input speech signal.
Inverse filter means for outputting the residual signal to a second division number
Of each residual signal of the divided band
Sound source analysis method that extracts sound source signals for each divided band based on
And a cepstrum signal from the sound source signal for each divided band.
A cepstrum analysis means for extracting the number;
A sound source correcting means for correcting the sound source signal based on the sound signal;
The correction for each divided band output from the sound source correction means
Second quantizer for vector quantizing the generated source signal
A stage; each divided band output from the first quantizing means
Frame the vector quantized linear prediction coefficients with
Vector accumulation that accumulates in large frame units assembled by a constant number
Means and a plurality of vectors stored in the vector storage means
Toll quantized linear prediction coefficients are similar to each other.
Multiple vectors that are separated from those that are not and are considered to be similar
Represents only one of the quantized linear prediction coefficients
Tol Quantized linear prediction coefficient selected similarity measure
A step; the representative vector output from the similarity determination means
Toll quantized linear prediction coefficient and not considered similar
The remaining vector quantized linear prediction coefficients,
The vector for each divided band output from the second quantization means
Tol quantized sound source signal and encoded sound
Multiplexing means for outputting as a signal;
A signal receiving device; encoded voice from the voice communication transmitting device
Represents a signal and a representative vector for each divided band from the signal
Quantized linear prediction coefficient and not considered to be similar
Remaining vector quantized linear prediction coefficients and
Separated from vector quantized source signal for each region
The representative vector quantized linear prediction coefficient
Vector quantized linear prediction for the number of frames considered
Duplicated as a coefficient and not considered to have the similarities
Large frame with remaining vector quantized linear prediction coefficients
Each vector value to a scalar value.
Original separating means; division band output from the separating means
For all bands from the linear prediction coefficient and sound source signal for each band
And a synthesizing means for reproducing the synthesized voice signal .

In the voice communication method of the present invention, the transmitting side sends the input voice signal as an encoded voice signal to the transmission line, and the receiving side synthesizes the encoded voice signal received through the transmission line. In a voice communication method for reproducing as a voice signal;
On the transmitting side; a predetermined voice band is divided into bands of a first division number, and a linear prediction analysis is performed on the framed input voice signal of a predetermined cycle for each of the divided bands to obtain a linear prediction coefficient; Vector quantizing the linear prediction coefficient for each divided band output from the linear prediction analysis means; extracting a residual signal from the linear prediction coefficient for each same divided band and the input speech signal; Dividing into a band of a second number of divisions, extracting a sound source signal for each divided band based on each residual signal of the divided band; extracting a cepstrum signal from the sound source signal for each divided band; the cepstrum signal Correct the excitation signal on the basis of the above; vector-quantize the corrected excitation signal for each of the divided bands; frame the vector quantized linear prediction coefficient for each of the divided bands. Accumulated in large frame units aggregated by a constant; multiple accumulated vector quantized linear predictive coefficients are separated into those with similarities and those with no similarity, Only one of the linear predictive coefficients is selected as the representative vector quantized linear predictive coefficient; the representative vector quantized linear predictive coefficient and the remaining vector quantized linear predictive coefficients that are not regarded as having similarity. , The vector quantized excitation signal is multiplexed and output to the transmission line as a coded voice signal; on the receiving side; the coded voice signal is received via the transmission line, and a divided band is generated from the signal. Representative vector quantized linear prediction coefficient for each and the remaining vector quantized linear prediction coefficient that is not regarded as having similarity, and vector quantized sound for each divided band. Signals are separated from each other; the representative vector quantized linear prediction coefficient is duplicated as vector quantized linear prediction coefficients for the number of frames considered to have similarity, and the remaining vector not considered to have similarity After restoring to a value for a large frame together with the quantized linear prediction coefficient, each vector value is restored to a scalar value; a synthesized speech signal for all bands is restored from the restored linear prediction coefficient and excitation signal for each divided band. There is a step of reproducing.

[0020]

[0021]

[0022]

[0023]

[0024]

DESCRIPTION OF THE PREFERRED EMBODIMENTS First, the outline of the present invention will be described.
A voice communication device of the present invention relates to a voice communication device that generates a spectral envelope information and a sound source signal and transmits the sound signal as an encoded voice signal, and includes a plurality of types of frequency analyzers, inverse filters, sound source correctors, and vector quantizers. And a voice communication receiver including a sound source signal demodulator connected to the voice communication transmitter via a transmission path.

The voice communication transmitting device performs a linear predictive analysis (LPC analysis) on a framed input audio signal of a predetermined period for each of the first divided bands of the voice band, and outputs a linear predictive coefficient (LPC coefficient). Prediction analysis means, first quantization means for vector-quantizing linear prediction coefficients,
Inverse filter means for extracting the residual signal between the linear prediction coefficient and the input speech signal, sound source analysis means for extracting the sound source signal for each second divided band from the residual signal, and cepstrum for extracting the cepstrum signal from the sound source signal. Analysis means and sound source correction means for correcting the sound source signal based on the cepstrum signal,
Second quantizing means for vector quantizing the corrected sound source signal, vector accumulating means for accumulating the vector quantized linear prediction coefficients in a unit of a large frame in which a predetermined number of frames are assembled, and a plurality of accumulated vectors Quantized linear prediction coefficients are separated into those with similarity and those with no similarity, and only one of multiple vector quantized linear prediction coefficients that are regarded as having similarity is the representative vector quantized linear prediction Similarity determination means selected as coefficients, vector quantized linear prediction coefficients, remaining vector quantized linear prediction coefficients not regarded as having similarity, and vector quantized excitation signal are multiplexed and coded. And a multiplexing means for outputting as an audio signal.

The voice communication receiving device receives the encoded voice signal, and from the signal, the representative vector quantized linear prediction coefficient for each divided band and the remaining vector quantized linear prediction coefficient not regarded as having similarity. , The vector quantized sound source signal for each divided band is separated, and the representative vector quantized linear prediction coefficient is duplicated as the vector quantized linear prediction coefficient for the number of frames considered to have similarity, and the similarity is calculated. With the remaining vector quantized linear prediction coefficients that were not considered to be present, a large-frame value is restored, and a separation means that restores each vector value to a scalar value, and a synthetic speech signal from the linear prediction coefficient and the excitation signal is generated. And a synthesizing means for reproducing.

In the present invention, the input speech signal is subjected to a linear prediction analysis for each of the divided bands to obtain a linear prediction coefficient (LPC).
(1) underestimation of formant bandwidth, which was said to be a drawback of the conventional LPC analysis.
(2) Poor approximation of the third formant can be improved, and a highly accurate residual signal can be extracted by an inverse filter using this LPC coefficient.

Further, since the high-precision residual signal is further band-divided and the sound source signal is obtained by the sound source analysis means for each of them, the mixed voiced voice / unvoiced voice is separated for each band, For voiced speech, optimal sound source information can be extracted for each band.

Further, since the fluctuation of the sound source signal is measured and corrected for each band, it is possible to extract the sound source signal having the fluctuation close to that of natural voice. Therefore, since it is possible to extract the sound source information with high accuracy according to the individual characteristics, it is possible to perform voice analysis / synthesis faithful to the original human vocalization mechanism.

In addition, instead of transmitting all vector quantized linear prediction coefficients, a representative vector sending method in which only one representative vector value of similar vector values is transmitted is used, and therefore communication is performed while maintaining sound quality. The amount of information can be reduced.

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

FIG. 1 is a block diagram showing an embodiment of a voice communication transmitting apparatus 100 in the voice communication apparatus of the present invention. In the figure, an input audio signal is, for example, 300H by an audio band limiting low pass filter (LPF) 110.
After being limited to a telephone voice band of z to 3.4 kHz, it is supplied to an analog / digital converter (A / D converter) 120 and is sampled at a predetermined sampling frequency (for example, 8 kHz). It is converted into digital audio data having a bit number (for example, 16 bits). This audio data is composed of consecutive frames in a predetermined cycle (22.5 msec in this example), and in the following processing, it is processed in units of this frame.

This audio data is supplied to the bandpass filter group 130 and divided into a plurality of bands for each predetermined frequency band. In this example, it is divided into two around 1.2 kHz. That is, the audio data is the bandpass filter (BPF) 1 that constitutes the bandpass filter group 130.
31 and 132, respectively, taken out from the BPF 131 with a band of 300 Hz to 1.2 kHz and from the BPF 132 with a band of 1.2 kHz to 3.4 kHz.

The respective band-divided voice data output from the BPFs 131 and 132 are linear prediction analyzers (LPC analyzers) 141 and 142 provided corresponding to the respective divided bands.
To the LSP by known linear prediction analysis.
After being converted into LPC coefficients (linear prediction coefficients) such as parameters and α parameters, A / P is applied to the LPC inverse filter 160.
It is input together with the output voice data of the D converter 120.

L from each LPC analyzer 141 and 142
The PC coefficients (in this example, the α parameter is used) are input to vector quantizers 151 and 152 for compressing and quantizing, respectively, and the vector quantized value of each LPC coefficient (vector quantized LPC coefficient) is calculated. Vector storage 1
81 and 182 respectively.

Each of the vector accumulators 181 and 182 has 100 msec to 22 which does not cause a sense of discomfort as a vocal tract cycle of voice.
A plurality of audio data frames are assembled into a large frame so as to correspond to a time length of about 0 msec (in this example, 1 frame is 22.5 msec and 10 frames are 1 large frame). The vector value (vector quantized LPC coefficient) corresponding to each of the constituent frames is accumulated, and the similarity determiner 191
And 192 respectively.

Each of the similarity determiners 191 and 192 extracts signals having similarity to each other from the vector signals (vector quantized LPC coefficients) for 10 frames which form the above-mentioned large frame, and extracts them. Only one is selected as the representative vector. Experimentally, in the case of voiced speech, it is often calculated that 2 to 5 frames out of 10 are similar. Therefore, the linear prediction vector quantization information amount 2 to
Representative vector value (representative vector quantized LPC) selected as data for five frames as data for one frame
The coefficient) and the vector value (vector quantized LPC coefficient) determined to have no similarity are sent to the multiplexer 10. In this way, since the data of 2 to 5 frames which are similar to each other is set to the data of 1 frame, it is possible to compress the information to 1/2 to 1/5.

On the other hand, the LPC inverse filter 160 uses the A / D
The LPC analyzer 1 for the output voice data of the converter 120
Using the LPC coefficients of the two divided bands from 41 and 142, a filter characteristic opposite to the spectral envelope characteristic of the linear prediction analysis is added and a residual signal is output.

This residual signal is used as a bandpass filter group 1
It is supplied to 70 and divided into a plurality of bands. In this example,
The band is divided into three at the boundaries of 1.5 kHz and 2.5 kHz. That is, the residual signal is the bandpass filter group 1
70 is supplied to the BPFs 171, 172, and 173 of the BPF 171, respectively, and 300 Hz to 1.5 Hz from the BPF 171.
A frequency component of kHz is extracted, a frequency component of 1.5 kHz to 2.5 kHz is extracted from the BPF 172, and a frequency component of 2.5 kHz to 3.4 kHz is extracted from the BPF 173.

A plurality of (two in this example) sound source analyzers 211 to 212 are provided for the band-divided residual signals extracted from the BPFs 171, 172, and 173, respectively, for each sub-band. 221-222, and 2
31 to 232, respectively, where they are summed squared and extracted as a sound source signal (power) for each divided band.

Each sound source analyzer 211-212, 221-2
The sound source signals for each band from 22 and 231 to 232 correspond to the corresponding cepstrum analyzers 311 to 312, 32.
1 to 322 and 331 to 332, respectively.

Each cepstrum analyzer 311 to 312, 3
21-322 and 331-332 are cepstral values (keflency) on the frequency axis of the sound source signal for each band.
Is calculated, and the corresponding sound source correctors 411 to 412 and 421 are calculated.
To 422 and 431 to 432, respectively.

Each sound source corrector 411-412, 421-4
22 and 431 to 432 measure the dispersion value of the cepstrum value on the cefency, and when the dispersion is small, it is determined as a stationary signal, and when the dispersion is large, it is determined as a signal with fluctuation. The sound source signal determined to have fluctuation is corrected by giving fluctuation (fluctuation) synchronized with the pitch cycle (vocal cord signal). Furthermore, each sound source corrector 411-412, 421-4
The outputs of 22 and 431 to 432 are passed through decision units 501, 502, and 503 corresponding to each band, which will be described below.
Vector quantizer 51, which is the second vector quantizer
1, 512, and 513, respectively.

A sound source analyzer, a cepstrum analyzer,
The reason for providing two sound source correctors for each band is as follows. In the sound source analysis, the output signal divided into each band by the BPF group 170 is further divided into two so as to include a double pitch or a half pitch in each band, and one of the divided results is obtained. It is determined for each band whether the side (one of two) is more correct.
This is because the analysis route indicating the value closer to the input signal is determined as the correct value, and the value obtained from the route is determined as the correct analysis information for each band.

The vector quantizers 511, 512, and 513 have the sound source compensators 411 to 412, 421 to the side selected by the determiners 501, 502, and 503, respectively.
The vector quantized value (vector quantized sound source signal) of the sound source signal corrected in 422 and 431 to 432 is calculated.

The multiplexer 520 is a vector quantizer 51.
Vector quantized sound source signal (power) for each divided band extracted from each of 1, 512, and 513,
The vector quantized LPC coefficients of the two divided bands extracted from the other vector quantizers 151 and 152 (via the vector accumulator and the similarity determiner) are multiplexed and transmitted as a coded voice signal. Road 70
Output to 0.

FIG. 2 is a block diagram showing an embodiment of a voice communication receiving apparatus 800 in the voice communication apparatus of the present invention. In the figure, a separator 810 is a vector-quantized excitation signal (power) of three divided bands similar to that on the transmission side, from an encoded voice signal input from the voice communication transmitting device 100 via a transmission path 700. And the vector quantized LPC coefficients of the two divided bands are separated from each other, and the selected representative vector value (vector quantized LP
C coefficient) is duplicated as a vector value for the number of frames considered as having similarity, and is restored to a value for a large frame (10 frames) together with the vector value determined as having no similarity, and then is subjected to normal linear calculation. Restore to a scalar value that can be handled.

Of these, the LPC coefficient is the LPC interpolator 820.
The LPC values (vector values) supplied to the upper and lower sides of the voice band, respectively, are changed to scalar values that can be handled by linear calculation, and then both LPC coefficients (scalar values) are simply overlapped. , LPC coefficients that represent the entire audio band are reproduced, where a fixed period (for example, 22.5
(msec), the linear interpolation value that uses the previous input value and the current input value of the LPC coefficient is corrected to, for example, an LPC coefficient in units of 5.625 msec (in other words, changes in units of 22.5 msec). LPC coefficient is converted to an LPC coefficient that changes in units of 5.625 msec).

Further, the sound source signal (power) for each divided band separated by the separator 810 is supplied to the sound source demodulator 830, where the sound source signal (power) for each band is subjected to interpolation processing and the entire band (300 Hz). Up to 3.4 kHz). Here, the sound source demodulator 830 has three band pass filter groups, and uses sound source information (scalar value) for each band as a filter coefficient for sound source reproduction in the low band, the middle band, and the high band, respectively. The pitch information corresponding to is driven as energy. Then, the linear sum of the outputs of these three filters becomes the sound source information representing the entire audio band.

The LPC synthesis filter 840 uses the modified LPC coefficient output from the LPC interpolator 820 as a filter coefficient, and the digital synthesized speech data with the demodulated pitch information output from the sound source demodulator 830 as input energy. Reproduce.

This digital synthesized voice data is supplied to a digital / analog converter (D / A converter) 850, subjected to digital / analog conversion and taken out as a synthesized voice signal of an analog signal, and an LPF 860 at the next stage unnecessary frequency. After the components are removed, the re-synthesized speech signal is output.

As described above, in this embodiment, the LPC coefficient obtained by dividing the audio signal band into two is used for LPC.
The high-precision residual signal extracted by the inverse filter 160 is further divided into three bands, and the sound source analyzer 2
By obtaining the sound source signals at 11-212, 221-222, and 231-2232, the mixed voiced / unvoiced sound for each band is separated, and accurate sound source information matched to each characteristic is obtained. It is possible to extract, and the fluctuation of the sound source signal is measured and corrected for each band, so that the sound source signal having the fluctuation close to that of natural speech can be extracted. Therefore, it is possible to perform voice analysis / synthesis faithful to the original human vocalization mechanism.

Further, in order to correct the double period detection which is a defect in frequency analysis, the sound source analyzer is divided into two, and the BPF group 1
Since the calculation is performed by including the double pitch or the half pitch in each band divided by 70, it has a feature that the analysis error can be limited.

Further, since not all vector quantized linear prediction (LPC) coefficients, which are vocal tract information, are transmitted but only one representative value having similarity is transmitted, the communication information is transmitted without deterioration in sound quality. The amount can be reduced.

The present invention is not limited to the above embodiment, and for example, the bandpass filter group 130.
Although the characteristics of the BPFs of 170 and 170 are described as fixed,
The center frequency can be varied based on the input information, and the number of divisions can be from 2 divisions to 4 divisions. The determination of the dispersion value of the cepstrum is fixed in the present embodiment, but the pitch frequency is described. It has been experimentally confirmed that it may be changed according to The conventional LPC vocoder transmits 40 bits for LPC analysis, 7 bits for pitch, 6 bits for sound source, and 1 bit for synchronization as a total of 54 bits per frame of 22.5 msec (54 / 22.5m = 24).
00 bps).

In this method, the bit quantizers of the vector quantizers (151, 152) in the LPC analysis are assigned 10 bits, and a total of 20 bits are assigned to the LPC coefficients. Further, the similarity determination information for the representative value selection information is set to 3 bits, and a total of 6 bits are allocated. Further, the vector quantizer (511, 512, 513) is also used as the sound source information.
Each bit of 4 is set to 4 bits, and a total of 12 bits are assigned to the sound source information. Further, a total of 15 bits of 5 bits are assigned as pitch information for each sound source in each band. (L
(PC coefficient 20 bits, similarity determination information 6 bits, sound source information 12 bits, pitch information 15 bits, and synchronization 1 bit can be realized with a total of 54 bits). It can be seen that it can be realized without increase.

[0057]

As described above, according to the present invention,
The input speech signal is subjected to linear prediction analysis for each of the divided bands, a high-precision residual signal is extracted by an inverse filter using a linear prediction coefficient (LPC coefficient), and then the high-precision residual signal is further band-divided. , By obtaining the sound source signal for each of them, the mixed voiced voice / unvoiced voice is separated for each band.
It is possible to extract accurate sound source information (optimal for each frequency band) according to individual characteristics, and prevent analysis errors due to double pitch and half pitch, which can be said to be a drawback in frequency analysis. Since the fluctuation of the sound source signal is measured and corrected for each band, the sound source signal having fluctuation close to that of natural speech is extracted, and voice analysis / synthesis faithful to the original human vocalization mechanism can be performed, resulting in higher naturalness. It is possible to obtain synthetic speech. Further, since not all vector quantized linear prediction (LPC) coefficients are sent but only one representative value having similarities is sent, there is no increase in LPC coefficients (reduction of communication information amount) and more faithful It is possible to reproduce a spectrum (without deterioration of sound quality) and perform narrowband voice communication with a higher degree of information compression.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of a voice communication transmitting device in a voice communication device of the present invention.

FIG. 2 is a block diagram showing an embodiment of a voice communication receiving device in the voice communication device of the present invention.

FIG. 3 is a block diagram showing a conventional voice communication transmitting device and voice communication receiving device.

[Explanation of symbols]

100 voice communication transmitter 110,860 low pass filter (LPF) 120 A / D converter 130,170 band pass filter group 131,132,171-173 band pass filter (BPF) 141,142 LPC analyzer 151,152,511 ˜513 vector quantizer 160 LPC inverse filter 211,212,221,222,231,232
Sound source analyzer 311,312,321,322,331,332
Cepstrum analyzer 411, 412, 421, 422, 431, 432
Sound source correctors 501 to 503 Determinator 520 Multiplexer 800 Voice communication receiving device 810 Separator 820 LPC interpolator 830 Sound source demodulator 840 LPC synthesis filter 850 D / A converter

Claims (2)

(57) [Claims] 1. A voice signal is input and as an encoded voice signal
A voice communication transmitting device for outputting to a transmission line, and
And input encoding connected to the voice communication transmitter
Audio communication receiver for reproducing synthesized audio signal from audio signal
A voice communication transmitting device comprising: a voice communication transmitting device;
Divide into as many bands as the number of divisions, and
Predicts a framed input speech signal with a predetermined period
Linear prediction analysis means for analyzing and outputting a linear prediction coefficient;
Before each divided band output from the linear prediction analysis means
First quantizing means for vector quantizing the linear prediction coefficient
And; the linear prediction coefficient and the input sound for each of the same divided bands
Inverse filter means for receiving the voice signal and extracting the residual signal
And dividing the residual signal into a second number of bands,
For each split band based on each residual signal of the split band
Sound source analysis means for extracting a sound source signal into
Cepstrum that extracts the cepstrum signal from the sound source signal
Lamb analysis means; the sound source based on the cepstrum signal
Sound source correction means for correcting the signal; from the sound source correction means
The corrected sound source signal for each output divided band is
Second quantizing means for quantizing Kuttle; and the first quantum.
Vector quantum for each divided band output from the digitizing means
A large frame composed of a predetermined number of frames of the converted linear prediction coefficients.
Vector accumulating means for accumulating in units of units;
Multiple vector quantized linear predictions stored in storage means
The coefficients are separated into those with similarities and those with no similarity,
Multiple vector quantized linear predictions considered similar
Representative vector quantized linear prediction for only one of the coefficients
Similarity determination means selected as a coefficient; the similarity determination
The representative vector quantized linear prediction output from the means
Coefficients and remaining vectors not considered alike
The quantized linear prediction coefficient and the second quantization means
Vector quantized source signal for each divided band
And multiplex and output as encoded audio signals.
And a voice communication receiving device;
Receives the coded speech signal and uses the signal for each divided band.
Table vector quantized linear prediction coefficient and assumed to be similar
The remaining vector quantized linear prediction coefficients that were not
And vector quantized sound source signal of the divided bands your capital it
Separate each of the representative vector quantized linear prediction coefficients
Vector quantized for the number of frames regarded as having similarity
Duplicated as a linear prediction coefficient and considered to have the similarity
With the remaining vector quantized linear prediction coefficients that were not
After restoring to large frame values, each vector value is a scalar
-Separating means for restoring the value; output from the separating means
From the linear prediction coefficient and source signal for each divided band
And a synthesizing unit for reproducing a synthesized voice signal for a band.
A voice communication device characterized by the above . 2. A voice in which a transmitting side sends an input voice signal as a coded voice signal to a transmission line, and a reception side reproduces a coded voice signal received via the transmission line as a synthetic voice signal. In the communication method, on the transmitting side; a predetermined voice band is divided into bands of a first division number, and a linear prediction analysis is performed by linearly predicting a framed input voice signal of a predetermined cycle for each of the divided bands. Obtaining a coefficient; vector-quantizing the linear prediction coefficient for each divided band output from the linear prediction analysis means; extracting a residual signal from the linear prediction coefficient for each same divided band and the input speech signal; The residual signal is divided into bands of a second division number, and a sound source signal is extracted for each divided band based on each residual signal of the divided band; a cepstrum signal is extracted from the sound source signal for each divided band. Extracting; correcting the excitation signal on the basis of the cepstrum signal; vector-quantizing the corrected excitation signal for each of the divided bands; a predetermined number of frames of the vector quantized linear prediction coefficient for each of the divided bands; Accumulated in a unit of large frames that have been assembled; multiple accumulated vector quantized linear predictive coefficients are separated into those with similarity and those with no similarity, and multiple vector quantized linear that are regarded as similar Only one of the predictive coefficients is selected as the representative vector quantized linear predictive coefficient; the representative vector quantized linear predictive coefficient and the remaining vector quantized linear predictive coefficients that are not regarded as having similarity; The vector quantized excitation signal is multiplexed with each other and output to the transmission path as an encoded audio signal; on the receiving side; the encoded audio via the transmission path Vector quantized linear predictive coefficient for each sub-band and the remaining vector quantized linear predictive coefficients that are not regarded as having similarity and the vector quantized sound source for each sub-band. Separate the signal and the signal respectively;
The representative vector quantized linear prediction coefficient is copied as a vector quantized linear prediction coefficient for the number of frames that are regarded as having similarity, and the remaining vector quantized linear prediction coefficients that are not regarded as having similarity are included. After restoring to a value for a large frame, each vector value is restored to a scalar value; a synthesized speech signal for all bands is reproduced from the restored linear prediction coefficient and excitation signal for each divided band. Voice communication method.