JPH0229235B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a linear predictive speech analysis and synthesis device.

Speech analysis and synthesis equipment has recently been put into practical use with the establishment of the linear predictive analysis method (LPC method).

This LPC analysis method approximates the spectral envelope of speech using an all-pole model, but this method requires an underestimation of the formant bandwidth, which is known empirically, and a third formant with relatively small energy. It is said that there are two drawbacks: poor approximation.

The above-mentioned first drawback is thought to occur because the poles of the spectrum are excessively concentrated at frequencies where the energy of the first formant is concentrated.

Recently, in order to prevent such concentration of poles at specific frequencies, the audio band is divided into multiple subbands, and LPC analysis of an appropriate order is performed on each subband to measure the appropriate dispersion of the poles. A band-splitting linear predictive analysis method is being considered.

By appropriately selecting the band division, this method alleviates not only the first drawback but also the second drawback, and has the potential to become an effective means of contributing to improving the sound quality of LPC vocoders. ing.

However, in general speech, unvoiced sounds, especially fricatives, do not inherently have a steep spectral envelope, so the above-mentioned band division is not only ineffective, but also requires unnecessary band division. In fact, such processing may cause deterioration in sound quality.

SUMMARY OF THE INVENTION An object of the present invention is to provide a linear predictive speech analysis and synthesis device that eliminates the above-mentioned conventional drawbacks.

The device of the present invention is a linear predictive speech analysis and synthesis device that divides an input speech signal into a plurality of speech transmission bands and performs linear predictive analysis on each band.
The apparatus includes means for selectively performing band-split linear predictive analysis if the input audio signal is a voiced sound, and selectively performs a non-band-splitting linear predictive analysis if the input audio signal is an unvoiced sound.

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

Figures 1, 2, and 3 are block diagrams showing one embodiment of the present invention; Figure 1 is a block diagram showing the whole, Figure 2 is a block diagram showing the analysis side, and Figure 3 is a block diagram showing the synthesis side. be.

This embodiment is comprised of an analysis side 1, a synthesis side 2, and a transmission line 3, as shown in FIG.

As shown in FIG. 2, the analysis side 1 further includes a low-pass filter and A/D converter 101, a window processor 102, a Fourier transformer 103, a power spectrum memory 104, and a low-pass autocorrelation coefficient measuring device 105. , low-frequency side linear prediction coefficient analyzer 106,
High-frequency autocorrelation coefficient measuring device 107, high-frequency linear prediction coefficient analyzer 108, full-band autocorrelation coefficient measuring device 109, full-band linear prediction coefficient analyzer 110, voiced/unvoiced discriminator 111, pitch extractor 112 and an encoder 113.

Furthermore, as shown in FIG.
High-frequency side LPC filter 205, high-frequency side interpolator 20
6. Frequency converter 207, high band filter 2
08, low and high frequency synthesizer 209, full band LPC filter 210, pitch generator 211, noise generator 2
12, includes a sound source switch 213, a low-frequency variable gain amplifier 214, a high-frequency variable gain amplifier 215, a full-band variable gain amplifier 216, a mode switch 217, a D/A converter, and a low-pass filter 218. .

Now, the audio waveform to be transmitted is input from the audio input terminal 1000 of the analysis side 1 shown in FIG. The sampled signal is sampled at a sampling frequency of, for example, 8000 Hz, quantized into a digital signal of, for example, 12 bits per sample by an A/D converter, and supplied to the window processor 102.

The window processor 102 temporarily stores the input quantized audio signal in an internal memory. This memory stores, for example, 30 mSEC (240 samples) of the quantized audio signal, and performs window processing by multiplying this by a window function such as a Hamming window or a rectangular window.

Such window processing can be done, for example, by 10mSEC
This is repeated periodically, and this becomes the basic analysis period (hereinafter referred to as the basic frame period).

Now, the window-processed audio waveform data is passed through the Fourier transformer 1 for each basic frame period.
03, the voiced/unvoiced discriminator 111, and the pitch extractor 112.

The Fourier transformer 103 uses the input audio waveform data that has been subjected to the window processing described above, performs a Fourier transform on this to obtain the spectrum component of each frequency, and further calculates the power at each frequency by taking the square of this absolute value. It is converted into spectrum components and stored in the power spectrum memory 104. The data representing the power spectrum components thus stored in the memory 104 are freely read out by various autocorrelation coefficient measuring devices 105, 107, and 109, and used to measure the autocorrelation coefficients as described below. used.

Now, the low-frequency side autocorrelation coefficient measuring device 105 measures the low-frequency side of the power spectrum, for example, from 0 to
Memory 10 of the power spectrum components of 1333Hz
4 and performs inverse Fourier transform to measure the autocorrelation coefficient at each delay time within the necessary range, and supply this to the low-band linear prediction coefficient analyzer 106.

At the same time, the measuring device 105 supplies the measured autocorrelation coefficient with a delay time of 0 to the encoder 113 via the output line 1050 as the low-frequency side short-time average power in this basic frame period.

The analyzer 106 converts the K parameter from the supplied set of autocorrelation coefficient data to a predetermined order, for example, by autocorrelation (AUTO).
CORRELATION) method or the like, and the extracted low-frequency parameters are supplied to the encoder 113.

On the other hand, the high-frequency side autocorrelation coefficient measuring device 107 measures the high-frequency side of the power spectrum, in the above example.
The components of the power spectrum from 1333Hz to 3333Hz are read out from the memory 104 and subjected to inverse Fourier transform to measure the autocorrelation coefficient at each delay time within the required range, which is then calculated by the high-frequency side linear prediction coefficient analyzer. 108. However, when calculating the inverse Fourier transform mentioned above, from 1333Hz to 3333Hz
The Hz power spectrum component is frequency-shifted by 1333 Hz to a lower frequency, treated as a power spectrum from 0 to 2000 Hz, performs inverse Fourier transform, and measures the autocorrelation coefficient.

At the same time, the measuring device 107 supplies the measured autocorrelation coefficient with a delay time of 0 to the encoder 113 via the output line 1070 as the high-frequency side short-time average power in this basic frame period.

The analyzer 108 calculates the K parameter to a predetermined order from the supplied autocorrelation coefficient data set.
For example, it is extracted by a linear predictive analysis method such as an autocorrelation method, and the extracted high-frequency side K parameter is supplied to the encoder 113.

Furthermore, full-band autocorrelation coefficient measuring device 109
reads out the power spectrum components from the memory 104 over the entire band of the power spectrum, from 0 to 3333 Hz in the above example, and calculates the autocorrelation coefficient at each delay time within the required range by performing inverse Fourier transform on this. and supplies it to the full-band linear prediction coefficient analyzer 110. Furthermore, the measuring device 109 supplies the measured autocorrelation coefficient with a delay time of 0 to the encoder 113 via an output line 1090 as the entire band short-time average power in this basic frame period.

Now, the analyzer 110 calculates the K parameter to a predetermined order from the supplied set of autocorrelation coefficient data.
For example, it is extracted by a linear predictive analysis method such as an autocorrelation method, and the extracted full-band K parameter is supplied to the encoder 113.

For details on the above-mentioned autocorrelation method, see, for example, John Matsukoll.
Makhoul): “Linear Preday Crayon (Linear
Prediction: A Tutorial Review”, Proceedings of the
IEEE, Vol.63, No.4, pp.561-580 April, 1975
Please refer to

Furthermore, when the analyzers 106 and 108 perform linear prediction coefficient analysis on the low-frequency side and the high-frequency side, the highest frequency of these input audio signals is
As mentioned above, the low frequency side ranges from 0 to 1333Hz, and the high frequency side ranges from 0 to 2000Hz, and the highest frequency is 4000Hz, which is the maximum frequency determined by the original sampling period.
However, when performing linear predictive analysis, the sampling period is decimated to 3 times and 2 times the original sampling period, respectively. This is equivalent to using .

Now, the voiced/unvoiced discriminator 111 receives the window-processed audio waveform data, determines whether the audio signal in this frame is a voiced sound or an unvoiced sound for each basic frame period, and encodes the determination result. It is supplied to the converter 113. As this voiced/unvoiced discriminator, for example, Japanese Patent Application Laid-Open No. 54-94212 or
It can be easily configured by applying 54-151303.

Further, the pitch extractor 112 extracts pitch frequency data in each basic frame from the supplied window-processed audio waveform data,
This is supplied to the encoder 113.

The encoder 113 encodes the various data thus supplied to create transmission frames, and sends out one transmission frame for each basic frame to the combining side 2 via the transmission line 3. However, when creating this transmission frame, the following processing is particularly included.

That is, each transmission frame includes the voiced/unvoiced discrimination information from the voiced/unvoiced discriminator 111, and in the transmission frame including voiced information, the low-frequency side K parameter and the high-frequency side K parameter are included as the K parameters to be transmitted. The band-side K parameters are encoded as a set, and in this case, the full-band K parameters are not used.

On the other hand, in a transmission frame including unvoiced information, the full-band K parameter is used for encoding as the K parameter to be transmitted, and the high-band K parameter and the low-band K parameter are not used.

Regarding short-time average power, similarly, in a transmission frame including voiced information, the low-frequency short-term average power and the high-frequency short-term average power are encoded as a set as the short-term average power to be transmitted. , in this case, the full-band short-time average power is not used.

On the other hand, in a transmission frame containing unvoiced information, the above-mentioned all-band average power is used for encoding as the short-term average power to be transmitted, and the high-band short-term average power and the low-band short-term average power are used. not present.

Thus, the band-divided K parameter and short-time average power are sent to the synthesis side 2 for the voiced sound section, and the band-unsplit K parameter and short-time average power are sent to the synthesis side 2 for the unvoiced sound section.

Now, on the combining side 2, as shown in FIG. 3, the transmission frames transmitted via the transmission path 3 are successively supplied to the decoder 201.

The decoder 201 reproduces each data on the analysis side by decoding these transmission frames, and supplies these data as follows.

First, the voiced/unvoiced discrimination information is determined, and if the transmitted frame is voiced information, the reproduced low-frequency side K parameter is sent to the low-frequency side LPC filter 202, and the reproduced high-frequency side K parameter is transmitted to the low-frequency side LPC filter 202. are respectively supplied to the high frequency side LPC filter 205. On the other hand, in the case of a transmission frame whose discrimination information is unvoiced information, the reproduced full-band K parameter is supplied to the full-band LPC filter 210.

The decoder 201 further reproduces the transmitted information specifying the pitch frequency and supplies this to the pitch generator 211 as a pitch frequency control signal.

Further, the decoder 201 reproduces the voiced/unvoiced discrimination information and supplies it to the sound source switch 213 as a voiced/unvoiced switching signal via an output line 2010, and also supplies this as a mode switching signal to the mode switch 213.
Supply to 7.

Furthermore, in the case of a transmission frame in which the voiced/unvoiced discrimination information is voiced information, the decoder 201 converts the reproduced low frequency side short time average power and high frequency side short time average power into the low frequency side Output line 2 is used as gain control information and high frequency side gain control information, respectively.
011 and 2012, respectively, to a low-band variable gain amplifier 214 and a high-band variable gain amplifier 215. On the other hand, in the case of a transmission frame that is unvoiced information, the reproduced full-band short-time average power is supplied as full-band gain control information to the full-band variable gain amplifier 216 via an output line 2013.

Now, the pitch generator 211 generates pitch pulse data of a specified frequency, and the sound source switch 213
supply to.

The sound source switcher 213 selects the input data on the pitch generator 211 side when the voiced/unvoiced switching signal supplied via the line 2010 specifies a voiced sound, and selects the input data on the pitch generator 211 side when the voiced/unvoiced switching signal supplied via the line 2010 specifies a voiced sound, and selects the input data on the pitch generator 211 side when the voiced/unvoiced switching signal supplied via the line 2010 specifies a voiced sound. 2
12 side input data is selected and supplied to each variable gain amplifier 214, 215 and 216.

Each variable gain amplifier 214, 215 and 216
was thus provided. Pitch pulse data or noise signal data are sent to the line 201, respectively.
1, 2012, and 2013 to create variable amplification and source excitation data, which is applied to excitation signal lines 2140, 2150, and 2, respectively.
160 to the low frequency side LPC filter 202, the high frequency side LPC filter 205, and the full band LPC filter 210, respectively, as sound source excitation data.

Now, the low-frequency side LPC filter 202 that has received the reproduced low-frequency side K parameter converts the supplied K parameter into an α parameter, and uses this α parameter as a filter coefficient of the LPC filter. and synthesizes the sound source excitation data supplied via the line 2140, this filter coefficient, and the audio waveform data on the lower frequency side,
This is supplied to the low frequency side interpolator 203.

Due to the aforementioned decimation on the analysis side, the sampling period of the low-frequency audio waveform data synthesized from the low-frequency K parameters is three times the normal sampling period. Low-frequency interpolator 203
interpolates the supplied audio waveform data by passing it through a 1333 Hz low-pass filter to create audio waveform data with a normal sampling period, and supplies this to the low-side band filter 204.

The low band filter 204 removes frequency components in unnecessary bands by passing the supplied data through a band filter having a band from 300Hz to 1333Hz, for example, to generate low band audio waveform data.
It is supplied to one input side of the low and high frequency synthesizer 209.

On the other hand, the high-frequency side LPC filter 205, which has received the reproduced high-frequency side K parameter, internally converts the supplied K parameter into an α parameter, and uses this α parameter as a filter coefficient of the LPC filter. Then, the sound source excitation data supplied via the line 2150, this filter coefficient, and higher frequency audio waveform data are synthesized, and this is supplied to the higher frequency interpolator 206.

Due to the processing on the analysis side mentioned above, the high-frequency side K parameter changes from 1333Hz of the original audio signal to 3333Hz.
from 0 to 2000 by frequency shifting the components of
This is the K parameter for the voice waveform, which is transmitted in the band up to Hz and decimated to a sampling period twice the normal rate. Therefore, the audio waveform data synthesized from this K parameter has a sampling period twice the normal sampling period, and has a waveform whose frequency has been shifted lower by 1333 Hz. Therefore, the high frequency side interpolator 206
interpolates the supplied audio waveform data by passing it through a 2000 Hz low-pass filter to create audio waveform data with a normal sampling period, and supplies this to the frequency converter 207.

The frequency converter 207 multiplies the supplied audio waveform data by a 1333 Hz sine wave to shift the frequency of the audio waveform by 1333 Hz, and supplies this to the high band filter 208 .

The high frequency side band filter 208 removes frequency components in unnecessary bands by passing the supplied data through a band filter having a band from 1333Hz to 3333Hz, generates high frequency side audio waveform data, and generates high frequency side audio waveform data. The other input side of combiner 209 is supplied.

The low frequency high frequency synthesizer 209 adds and synthesizes the supplied low frequency audio waveform data and high frequency audio waveform data. Thus, the output 2090 is a synthesized speech obtained by performing the band division type linear predictive analysis, and this is supplied as one input of the mode switch 217.

On the other hand, the full-band LPC filter 210 that has received the reproduced full-band K parameter converts the supplied K parameter into an α parameter, and uses this α parameter as a filter coefficient of the LPC filter. , line 216
The sound source excitation data supplied via the output line 2100 and the filter coefficients are used to synthesize audio waveform data of all bands, and this is supplied as the other input of the mode switch 217 via the output line 2100.

In addition, each of the above-mentioned LPC filters 202, 205
The conversion from the K parameter to the α parameter, which is performed inside the filter 210, can be easily performed by applying the autocorrelation method described above, and the LPC filter itself can be easily configured as a recursive filter. be able to.

Now, the mode switch 217 is connected to the line 2010.
When the transmission frame synthesizes voiced sound, the input on the output line 2090 side is selected, and when the transmission frame synthesizes unvoiced sound, the input on the output line 2100 side is selected. is output to the D/A converter and low-pass filter 218.

The D/A converter and low-pass filter 218 converts the supplied audio data into an analog audio signal using a D/A converter, and further converts the supplied audio data into an analog audio signal using a low-pass filter.
Components of 3333 Hz or higher are blocked and output as a synthesized audio signal from the output terminal 2000.

As is clear from the above explanation, if the input speech is a voiced sound, the speech analysis and synthesis device of this embodiment transmits the data resulting from the band-splitting linear predictive analysis on the analysis side to the synthesis side, and synthesizes it. The synthesis side performs corresponding speech synthesis, and if the input speech is unvoiced, the analysis side transmits the data resulting from band non-splitting linear predictive analysis to the synthesis side, and the synthesis side synthesizes the corresponding speech. Perform synthesis.

As a result, the disadvantages of the conventional LPC analysis method mentioned earlier, such as underestimation of the formant bandwidth and poor approximation of the third formant, which has relatively small energy, can be alleviated. Through this process, linear predictive speech analysis and synthesis can be performed without the possibility of deteriorating sound quality.

In addition, in the above-mentioned embodiment, when performing band division type operation, the audio band is divided into two into a high frequency side and a low frequency side, but this is just an example, and the number of divisions can be further increased. .

Furthermore, the division frequency of 1333Hz used in the case of two divisions is just an example and is not limited to this.

Similarly, the decimate ratio is just an example.

In addition, in this example, when performing band division on the analysis side, the power spectrum of the entire band was first obtained and divided into each band, but instead of this, the input waveform was handled on the time axis. It is also possible to adopt a configuration in which the waveform is divided using a bandpass filter and then converted into a base band by frequency shifting, and this waveform is decimated according to the bandwidth, and then linear predictive analysis is performed.

As described above, by using the present invention, it is possible to improve the sound quality of a linear predictive speech analysis and synthesis device.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing the entire embodiment of the present invention, FIG. 2 is a block diagram showing the analysis side of this embodiment, and FIG. 3 is a block diagram showing the synthesis side of this embodiment. In the figure, 1...analysis side, 2...synthesis side, 3
...Transmission line, 101...Low pass filter and A/
D converter, 102... Window processor, 103
...Fourier transformer, 104...Power spectrum memory, 105...Low band side autocorrelation coefficient measuring device, 106...Low band side linear prediction coefficient analyzer, 10
7...High frequency side autocorrelation coefficient measuring device, 108...High frequency side linear prediction coefficient analyzer, 109...Full band autocorrelation coefficient measuring device, 110...Full band linear prediction coefficient analyzer, 111... ...voiced/unvoiced discriminator, 112...
...Pitch extractor, 113... Encoder, 201...
...Decoder, 202...Low band side LPC filter,
203...Low band interpolator, 204...Low band filter, 205...High band LPC filter, 2
06...High band interpolator, 207...Frequency converter, 208...High band filter, 209...
Low and high frequency synthesizer, 210... Full band LPC filter, 211... Pitch generator, 212... Noise generator, 213... Sound source switcher, 214... Low band side variable gain amplifier, 215... High Band side variable gain amplifier, 216...Full band variable gain amplifier, 217...
...Mode switch, 218...D/A converter and low-pass filter.

Claims (1)

[Claims] 1. In a linear predictive speech analysis and synthesis device that divides an input audio signal into a plurality of audio transmission bands and performs linear predictive analysis on each band, if the input audio signal is a voiced sound, the band division linear predictive analysis is performed on an unvoiced sound. A linear predictive speech analysis and synthesis device comprising means for selectively performing band non-splitting linear predictive analysis, if any.