JPH0229238B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a band division type vocoder, particularly a band division linear prediction type which divides an input audio signal into two predetermined audio transmission bands, a low frequency band and a high frequency band, and performs linear predictive analysis on each band. Regarding the vocoder.

The input audio signal is analyzed using the linear predictive analysis method (LPC analysis method).
A linear prediction vocoder that analyzes the input audio signal using a Linear Prediction Coefficient analysis method and transmits the resulting spectral distribution information of the input audio signal as a feature parameter in place of the input audio signal is well known these days.

This LPC analysis method uses an all-pole model to approximate the spectral envelope of the voice as spectral distribution information among the spectral distribution information and sound source information that are characteristic parameters of the input voice.
It is known that the results obtained by LPC analysis have the following two drawbacks.

One is what is known empirically as underestimation of the formant bandwidth of the input speech signal, a phenomenon that is particularly common with respect to the first formant, and which often has a formant bandwidth that is a fraction of that of the actual speech signal. This appears as a result of underestimating.

Second, in the LPC analysis method, the approximation of the third formant, which has relatively small energy, is poorer than that of the first formant.

The first and second drawbacks mentioned above, coupled with the fact that human auditory characteristics have logarithmic sensitivity and clarity is significantly impaired unless the third formant is maintained, are analyzed and synthesized using a vocoder. This degrades the quality of the audio.

The first drawback of such an LPC analysis method is thought to occur because the poles are excessively concentrated in a frequency region where energy is concentrated, such as the first formant of input speech. In this way, in order to prevent poles from concentrating in a specific frequency region, the audio band is divided into multiple subbands and LPC analysis is performed on each of the subbands to appropriately disperse the poles. Partitioned linear predictive analysis methods have been recently investigated. According to this method, by appropriately selecting the number of band divisions, etc., in addition to the first drawback mentioned above, the second drawback is alleviated, and it becomes an effective means for improving the sound quality to be analyzed and synthesized. It has potential.

As described above, experimental results have revealed that the band-splitting LPC analysis and synthesis method, which can improve the shortcomings of the conventional LPC analysis method, has the following characteristics.

In other words, using two as the basic number of band divisions can yield better analytical results than other numbers of division, and this eliminates the two drawbacks mentioned above that are unique to the conventional LPC analysis method. is significantly improved. In addition, in this type of band-splitting LPC analysis method, it is necessary to reduce the sampling rate from the sampling rate before dividing so that each band is represented by the minimum necessary sampling frequency, which is called decimate. , and this decimation has various characteristics such that it is advantageous for calculation processing to perform it in the frequency domain.

In addition, a drawback of the band-splitting LPC analysis method, especially the two-split band LPC analysis method, is that the K parameter obtained as a result of LPC analysis is the second-order K parameter, especially in the low range of the two-split band, ₂ has a significantly biased distribution compared to other K parameters in the low and high ranges, and numerous actual measurement results show that the appearance distribution of K ₂ parameter values is −
Concentrate near 1. To explain the reason for this in detail, in the band-splitting LPC analysis method, the spectral envelope of the input voice is expressed by a combination of a finite number of polar frequencies, and the boundary frequency is Set. Further, the second formant and subsequent formants do not exist in the low frequency range, but have a high proportion in the high frequency range. Therefore, there is usually only one formant on the low frequency side. One formant can be expressed (approximated) by one pole. In this case, the first-order K parameter K ₁ is the center frequency of the first formant (when the boundary frequency is πrad.)
It becomes the cosine of the angular frequency. Also, the K ₂ parameter depends on the bandwidth of the first formant. When the bandwidth is zero or line spectrum
K ₂ =-1, K ₂ =0 when the bandwidth is infinite, ie, there are no poles, and 0<K ₂ <-1 when the bandwidth is finite. 0<K ₂ <-1 is the distribution range of the K ₂ parameter when only the first formant exists on the low frequency side. The first formant usually has a sharp peak in the spectrum and a narrow bandwidth. Therefore, the distribution of the K ₂ parameter in the low range is concentrated near −1.

As is well known, the K parameter is used as information to set the coefficients of the digital synthesis filter used for speech synthesis on the synthesis side, and is a necessary condition for the stable operation of this synthesis filter |Kn|
It is necessary to satisfy <1, (n=1, 2, . . . P). Here, P is the order of the synthesis filter.

The K parameters, denoted by Kn, are each assigned a predetermined number of bits and quantized. This K parameter, which is sent as a quantized transmission code from the analysis side of the vocoder to the synthesis side, is used to suppress quantization errors, From the viewpoint of improving transmission efficiency in terms of the required number of quantization bits and therefore compressing the transmission band, or from the viewpoint of appropriateness of spectral sensitivity, the value should be approximately equal to the positive and negative values around 0. It is desirable to have a distribution characteristic that is distributed. Conversely, if the distribution of the K parameter of a specific order has a bias concentrated near the limit value such as +1 or -1, spectral distortion, so-called quantization distortion, occurs when encoding this K parameter. This increases the sensitivity to quantization, which causes deterioration in the quality of the speech to be synthesized.Increasing the number of quantization bits to avoid this has various drawbacks, such as increasing the bit rate, that is, expanding the transmission band.

Therefore, in order to avoid such drawbacks due to quantization of K parameters that include biased distributions such as the _K2 parameter, it is necessary to quantize with consideration to the spectral sensitivity and distribution of specific K parameters that exhibit such biases. Become. Each K parameter has a different spectral sensitivity depending on its order, and this spectral sensitivity is the conversion rate of the spectrum that corresponds to a _minute change in the K parameter. If the sensitivity is sensitive, quantization must be performed to compress the bias of this distribution by some means to suppress quantization distortion, reduce the bit rate, and compress the transmission band. Must be.

FIG. 1 is a K parameter distribution diagram showing an example of the distribution when K parameters up to the sixth order are used in a band division type linear predictive vocoder for low band A and high band B. Both low range A and high range B are K ₁
We use K parameters from to K ₆ , but
As is clear from the figure, it is clearly seen that the distribution of the K ₂ parameter, especially in the low range A, is concentrated near -1.

The purpose of the present invention is to eliminate the above-mentioned drawbacks, and to provide a band division _type vocoder that divides an input audio signal into predetermined transmission bands of low and high frequencies and performs linear predictive analysis. Band division can greatly reduce the effects of quantization distortion by providing a simple means of nonlinear quantization and transmission on the analysis side, and can also significantly improve the reduction of the required quantization bit rate and therefore the transmission band. Its purpose is to provide a type vocoder.

The vocoder of the present invention is a band division type vocoder that divides an input audio signal into two predetermined audio bands, a low frequency band and a high frequency band, and performs linear predictive analysis on each band. K ₂ parameter nonlinear quantization transmission means for nonlinearly quantizing and transmitting a second-order K parameter K ₂ of the partial autocorrelation coefficient K parameters obtained by linear predictive analysis of the low frequency audio transmission band obtained by analysis; Prepared and configured.

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

FIG. 2 is a block diagram showing an embodiment of the analysis side of the band division type vocoder of the present invention, and FIG. 3 is a block diagram showing an embodiment of the synthesis side of the band division type vocoder of the present invention.

When an input audio signal 1001 is received from an input terminal 1000, it is passed through a low-pass filter LPF1 with a high-frequency cutoff frequency of 3333Hz to an LPF that cuts off high frequencies.
An output 101 is obtained, which is further sampled by the A/D converter 2 at a sampling frequency of 8 KHz, converted into a digital quantity of a predetermined number of bits, and output as an A/D converter output 201, which is sent to the window processor 3. The high cutoff frequency of LPF1 mentioned above is the upper limit of the audio signal to be analyzed.
It is set to 3333Hz.

The window processor 3 temporarily stores the input A/D converter output 201 in an internal buffer memory. This memory stores 30 mSEC of the input signal, ie, 240 samples, and performs window processing by multiplying this by a Hamming function. This window processing is performed at a cycle of 10 mSEC, and this cycle is the basic analysis cycle, or so-called basic frame cycle.

Windowed audio waveform data 301
is the DFT for each basic frame period mentioned above.
(Discrete Fourier Transformation) circuit 4,
The signal is sent to the voiced/unvoiced discriminator 5 and the pitch extractor 6.

DFT circuit 4 input audio waveform data 301
is Fourier-transformed into a frequency domain spectral component, which is sent to the power spectrum measurement circuit 7 as a DFT output 401.

The power spectrum measurement circuit 7 uses the input DFT
The output 401 is squared and converted into power spectrum components at each frequency, which are stored in a built-in memory. The power spectrum data 701 stored in the memory is freely read out by the low-frequency autocorrelation coefficient measuring device 8 and the high-frequency autocorrelation coefficient measuring device 9, and the low-frequency and high-frequency bands specified by the audio waveform data 301 are It is used to measure the autocorrelation coefficient in

The low frequency autocorrelation coefficient measuring device 8 measures the low frequency range of the power spectrum data 701, from 0 to 1333 in this embodiment.
The power spectrum data up to Hz is read out, subjected to discrete inverse Fourier transform to measure the autocorrelation coefficient at each delay time within the required range, and this low frequency autocorrelation coefficient data 801 is sent to a low frequency linear prediction coefficient analyzer. Send on 10.

Furthermore, the low-frequency autocorrelation coefficient measuring device 8 sends the measured autocorrelation coefficient at a delay time of zero, that is, the short-time average power in this basic frame period, to the encoder 11 as low-frequency short-term average power data 802. .

The low-frequency linear prediction analyzer 10 extracts K parameters up to a predetermined order from a set of input low-frequency autocorrelation coefficient data 801 using the well-known autocorrelation method. Low range K parameter data 1010
It is sent to the encoder 11 as However, of these
Only the K ₂ parameter is sent to the K ₂ parameter memory 12 as K ₂ parameter data 1020 without being sent as is to the encoder 11 .

The K ₂ parameter data 1020 input to the K ₂ parameter memory 12 is read out one after another by the nonlinear conversion circuit 13, and prior to quantization by the encoder 11, the data is subjected to nonlinear conversion as follows.

The non-linear conversion circuit 13 uses _K2 parameter data 10 that is read out one after another by a BOM containing a control program in advance and stored at a predetermined address.
20 to log area ratio (Log Area
Ratio) conversion to compress the bias in the distribution.

The K parameter has a value corresponding to the reflection coefficient of a vocal tract acoustic filter that equivalently regards the inside of the vocal tract as a filter, and this vocal tract acoustic filter has a value corresponding to the reflection coefficient of a vocal tract acoustic filter that treats the vocal tract as a filter. It is well known that long acoustic tubes can be regarded as connected one after the other, and that the K parameter has a value corresponding to such a vocal tract cross-sectional area. The cross-sectional area of each part within the vocal tract changes continuously over time, and in speech analysis, this is expressed by a cross-sectional area ratio representing each cross-sectional area on both sides of a sampling point within the vocal tract. It is also well known that from this relationship, the cross-sectional area ratio Dn is expressed by the following equation (1).

Dn=(1-Kn)/(1+Kn)...(1) In equation (1), the subscript n indicates the order of the K parameter. log Dn of equation (1), that is, log(1-Kn)/(1
+Kn) is called the log area ratio of the K parameter, and this is often used in speech analysis as a logarithmic conversion form corresponding to the value of the K parameter. By using this log area regio instead of the K parameter, the K parameter becomes one whose variance is significantly compressed.

Log area ratio, which is one of the means to compress variance, is expressed by the following equation (2) in the case of K ₂ parameter.

log D ₂ = log(1-K ₂ )/(1+K ₂ ) ...(2) In this example, log D ₂ in equation (2) is substituted for the K ₂ parameter in order to compress the distribution of the K ₂ parameter. We are using.

The non-linear conversion circuit 13 in FIG. 2 applies arithmetic processing according to equation (2) to the input K ₂ parameter data 1020 under the control of a built-in program to obtain the log area ratio log D ₂ shown in equation (2). Nonlinear transformation K ₂ parameter data 1301 by
and sends it to the encoder 11, which encodes the K parameter data 1 excluding the _K2 parameter output from the low-band linear prediction coefficient analyzer 10.
010 is output from the nonlinear conversion circuit 13. The K ₂ parameters compressed by nonlinear transformation, that is, the K ₂ log area registers, are each quantized with a specific number of bits and converted into a transmission code, and then sent to the combining side via a transmission line 1101.

Now, the high-frequency autocorrelation coefficient measuring device 9 reads out the components from 1333Hz to 3333Hz as a high-frequency band out of the power spectrum data 701 output from the power spectrum measurement circuit 7, and performs an inverse Fourier transform on this to obtain The autocorrelation coefficient at each delay time within the required range is measured, and this high frequency autocorrelation coefficient data 901 is sent to the high frequency linear prediction coefficient analyzer 1.
At the same time, the high-frequency autocorrelation coefficient data at zero delay time is transmitted to the encoder 11 as high-frequency short-term average power data 902 in the aforementioned basic frame period.

In addition, in the Fourier inverse transform mentioned above,
The power spectrum component from 1333Hz to 3333Hz shifts this to the 1333Hz low frequency region, equivalently shifting from 0 to 3333Hz.
The autocorrelation coefficient is measured as a 2000Hz power spectrum component.

The high-frequency linear prediction coefficient analyzer 14 extracts K parameters up to a predetermined order from the input high-frequency autocorrelation coefficient data set 901 in exactly the same way as the low-frequency linear prediction coefficient analyzer 10, and extracts K parameters up to a predetermined order. It is sent to the encoder 11 as K parameter data 1401, whereby each K parameter data is quantized with a specific number of bits, converted into a transmission code, and sent via transmission line 1101 to the combining side shown in FIG. .

In addition, in the case of each linear prediction coefficient analysis in the above-mentioned low frequency linear prediction coefficient analyzer 10 and high frequency linear prediction coefficient analyzer 14, these low frequency bands from 0 to 1333Hz and the high frequency band 0
These linear predictions The sampling period in the analysis is three times and twice the original sampling period, respectively, which is equivalent to using a so-called decimated sampling period.

By the way, the audio waveform data 301 output from the window processor 3 is also supplied to the voiced/unvoiced discriminator 5 and the pitch extractor 6.

The voiced/unvoiced discriminator 5 uses voice waveform data 30 that has been window-processed by the window processor 3.
1, it determines whether the data is voiced or unvoiced for each basic frame, and sends the result to the encoder 11 as voiced/unvoiced discrimination data 501.

The method of determining voiced, unvoiced, or silent states is well known as a technology that is similar to pattern recognition, and this is discussed in B.S. Attal et al. Silence Classification with Application to Speech Recognition (BSAtal etal: “A Pattern Recognition
to Voiced－Unvoiced－Silence Classification
with Application to Speech Recognition”，
IEEE TRANS-ACTION ON ACOUSTIC,
SPEECH, AND SIGNAL PROCESSING,
VOL.ASSP-24, NO.3, June 1976.) and many other documents.

Further, the pitch extractor 6 extracts pitch frequency data 601 in the basic frame period from the input audio waveform data 301, and transmits it to the encoder 11.
Send to. To extract this pitch frequency data 601, the autocorrelation coefficient of the input audio waveform data 301 is calculated. A well-known pitch extraction method is used, which focuses on the fact that the autocorrelation coefficient takes the maximum value at a delay time that is the same as the pitch period.

The encoder 11 encodes the type of data supplied as described above to create a transmission frame, and sends each basic frame as one transmission frame to the synthesis side shown in FIG. 3 via the transmission line 1101. Send.

Next, the operation on the synthesis side will be explained with reference to FIG.

The decoder 21 decodes and reproduces various encoded data sent from the analysis-side encoder 11 shown in FIG. 2, and supplies each of these decoded data as follows.

The reproduced low-range K parameter data 1010 excluding the non-linearized _K2 parameters are processed by low-range LPC (Linear
The high-frequency K parameter data 1401 that is supplied to the Prediction Coefficient (Linear Prediction Coefficient) filter 22 and reproduced is sent to the high-frequency LPC filter 23.
Furthermore, low-frequency nonlinear conversion K ₂ parameter data 1
301 supplies each to the buffer memory 24. In addition, the reproduced voiced/unvoiced discrimination data 501
is supplied to the sound source information switch 25, and the reproduced pitch frequency data 601 is supplied to the pitch generator 26.
supply to.

The decoder 21 also supplies the reproduced low-frequency short-time average power data 802 and high-frequency short-term average power data 902 to the low-frequency variable gain amplifier 27 and the high-frequency variable gain amplifier 28, respectively.

In this way, among the various reproduced data supplied by the decoder 21, the pitch frequency data 601 is used to generate pitch pulse data 2601 of a pulse sequence of a repetition frequency corresponding to the pitch frequency information by the pitch generator 26. and sends this to the sound source information switch 25.

The sound source information switch 25 selects the pitch pulse data 2601 when the separately inputted voiced/unvoiced discrimination data 501 specifies voiced, and selects the pitch pulse data 2601 when the voiced/unvoiced discrimination data 501 specifies voiced. A switching circuit selects and outputs the white noise signal 2901 that corresponds to the voiced/unvoiced discrimination data 501, and outputs either the pitch pulse data 2601 or the white noise signal 2901 via the output line 2501 in accordance with the voiced or unvoiced information. The signal is sent to the low frequency variable gain amplifier 27 and the high frequency variable gain amplifier 28 described above.

Either the pitch pulse data 2601 or the white noise signal 2901 inputted in this manner is input separately to the low-frequency variable gain amplifier 27 and the high-frequency variable gain amplifier 28, respectively, depending on whether the sound source is voiced or unvoiced. The low-frequency short-time average power data 802 and the high-frequency short-term average power data 902 are variably amplified so as to receive weighting corresponding to the values of the respective short-term average power data, and then the low-frequency and high-frequency excitation signals are generated, respectively. Low frequency LPC as data 2701 and 2801
The signal is sent to filter 22 and high-frequency LPC filter 23.

Now, the low-pass LPC filter 22 that receives K parameter data 1010 other than K ₂ from the decoder 21 converts these K parameters into α parameters using a built-in K parameter/α parameter conversion circuit, and converts these K parameters into α parameters. Let be the filter coefficient of the LPC filter.

In addition, nonlinear transformation _K2 parameter data 130
1 is once stored in the buffer memory 24 for each basic frame, and is successively read out via the output line 2401 under the control of a program pre-built in the ROM of the _K2 parameter linear conversion circuit 30. Moreover, logs performed on the analysis side
Inverse conversion operation for area ratio conversion, that is, nonlinear conversion operation to obtain D ₂ shown by equation (1) from log D ₂ according to equation (2) mentioned above, and this D ₂
The K ₂ parameter data is reproduced by calculating K ₂ using the four arithmetic operations shown in equation (1). All the programs necessary for these calculations are preset in the built-in ROM of the K ₂ parameter linear conversion circuit 30, and as will be described later, the nonlinear conversion of the K ₂ parameters performed on the analysis side is performed as in this embodiment. In addition to nonlinear transformation using a log area ratio function for logarithmic compression, various nonlinear transformation functions can be used depending on the desired K _{2 parameter} dispersion compression characteristics. The nonlinear operation to reproduce the _K2 parameter is basically performed on the analysis side.
Nonlinear processing may be performed by inverse calculation of the function used for nonlinear transformation of the K ₂ parameter.

Now, the K ₂ parameter data 1020 before nonlinear conversion thus reproduced is sent to the low-pass LPC filter 22.

When the low-pass LPC filter 22 receives the K ₂ parameter data 1020 restored in this way, it converts it into an α parameter using a built-in K parameter/α parameter conversion circuit, and converts it into an α parameter.
parameter along with the other α parameters mentioned above.
Let it be the filter coefficient of the LPC filter. The low-pass LPC filter 22 that has obtained the necessary LPC filter coefficients in this way uses these filter coefficients and the low-pass excitation signal data 2 sent out from the low-pass variable gain amplifier 27.
701 to synthesize low-frequency audio waveform data,
This low frequency audio waveform data 2201 is transferred to the low frequency interpolator 3.
Send to 1. This low frequency audio waveform data 2201
Because of the decimating process on the analysis side shown in FIG. 1, the sampling period is three times the normal sampling period.

The low-frequency interpolator 31 interpolates input low-frequency audio waveform data 2201 through a low-pass filter with a cutoff frequency of 1333 Hz, converts it into audio waveform data with a normal sampling period, and converts this low-frequency interpolator output data 3101 into a low-frequency filter. Send to area BPF32.

The low-frequency BPE 32 is a band-pass filter for extracting only the necessary band from the input low-frequency interpolator output data 3101. In this embodiment, the high-frequency cutoff frequency is 1333Hz, and the low-frequency cutoff frequency is 1333Hz.
Although it is set to 300Hz, this low cutoff frequency can be arbitrarily set as desired. The output of the low frequency BPF 32 is sent to the low frequency high frequency synthesizer 33 as low frequency audio signal data 3201.

On the other hand, the high-frequency LPC filter 23 converts the input high-frequency K parameter data 1401 into an α parameter using a built-in K parameter/α parameter conversion circuit, uses this α parameter as a coefficient of the LPC filter, and uses this α parameter as a coefficient of the LPC filter. Area excitation signal data 28
01 to synthesize high-frequency audio waveform data,
This high frequency audio waveform data 2301 is transferred to the high frequency interpolator 3.
Send to 4.

As explained with reference to FIG.
Frequency shift the Hz component and convert it from 0 to 2000Hz
, and is equivalent to the K parameter for audio waveform data decimated to twice the normal sampling period. Therefore, the high frequency audio waveform data 230 synthesized using this K parameter
1, the sampling period is twice that of the normal one, and the frequency is also shifted lower by 1333Hz. The high-frequency interpolator 34 is
The input high-frequency audio waveform data 2301 is interpolated by passing it through a low-pass filter with a high-frequency cutoff frequency of 2000 Hz to obtain audio waveform data with a normal sampling period, and this high-frequency interpolator output 3
401 to the frequency converter 35.

The frequency converter 35 multiplies the input high-frequency interpolator output 3401 by the 1333Hz conversion signal 3501, and converts the input high-frequency interpolator output from 0 to 2000Hz.
After performing frequency conversion to shift by 1333 Hz, this frequency converter output 3502 is sent to the high frequency BPF 36.

The high frequency BPF 36 is the high frequency required from the input frequency converter output 3502, that is, 1333Hz.
This is a band filter for extracting components from 3333 Hz to 3333 Hz, and the components obtained by this are sent to the low frequency high frequency synthesizer 33 as high frequency audio signal data 3601.

The low frequency high frequency synthesizer 33 combines the thus inputted low frequency audio signal data 3201 and high frequency audio signal data 3.
601 are added and synthesized by an adder circuit, and this is sent to the D/A converter 37 as synthesized speech waveform data 3301.

The D/A converter 37 converts the input synthetic audio waveform data 3301 from a digital amount to an analog amount, and sends this analog audio signal output 3701 to the low-pass filter LPF 38.

LPF38 is a low-pass filter with a high cutoff frequency of 3.4KHz, and the analog audio signal 3 that passes through it
701 is output as a synthesized audio signal 3801.

The present invention provides a band division _type vocoder that divides an input audio signal into predetermined low and high frequencies and performs linear predictive analysis on each of them. The basic feature is that the data is nonlinearly quantized and transmitted from the analysis side to the synthesis side, and various modifications of this embodiment can be considered.

For example, in this example, K ₂
When nonlinearly quantizing parameters, log this
Quantization is performed by applying a logarithmic nonlinear transformation to an area ratio, but this is just an example of a nonlinear quantization method, and basically the input audio signal to be processed is The degree of bias in the distribution of the _K2 parameter that differs depending on the type, the degree of compression required and the degree of smoothing of the spectral sensitivity, or the desired number of bits to be added in quantization, that is, the number of quantization steps, etc. Approximate function data using quadratic or cubic functions, trigonometric functions, or other desired functions having nonlinear conversion characteristics that almost satisfy these conditions is stored in the ROM of the nonlinear conversion circuit 13 in advance, taking various conditions into consideration. By setting the values in advance, it is possible to easily perform desired nonlinear transformation and quantization.

In addition, in the above embodiment, 1333 Hz is used as the boundary frequency between the low and high range sides, but this can be arbitrarily selected as desired, and in connection with this, the decimation ratio can also be arbitrarily selected. It is clear that there is a choice.

Furthermore, in the above-mentioned embodiment, the analysis side shown in FIG. 2 measures the power spectrum over the entire band of the input audio signal and then divides it into low and high frequencies for processing. The input audio signal was divided using a bandpass filter having the desired low and high frequency band pass characteristics, and then the high frequency band was shifted so that the baseband was approximately the low frequency band. Moreover, it is clear that exactly the same results can be obtained by performing linear predictive analysis by performing decimation processing according to each band.

In addition, in this example, as shown in FIG.
The low-pass and high-pass LPC filters 22 and 23 are
These composite filters are separately formed as coefficients and excitation signals for each LPC filter by independently using the low and high frequency K parameters and the low and high frequency short-term average powers transmitted from the analysis side. The method used was to obtain the output waveform on the time axis and synthesize these on the time axis.
It is also possible to easily synthesize this in the frequency domain and then form a waveform on the time axis as shown below.

That is, in FIG. 3, the spectral envelopes in the low and high ranges are calculated using the low range K parameter data and the high range K parameter data reproduced by the decoder 21, and this result and the decoder 21
The autocorrelation coefficient of the entire band is measured using low-frequency and high-frequency short-term average power data that are input separately from . In this way, the autocorrelation coefficients for all bands including low and high bands are obtained, and linear predictive analysis is performed on all bands to obtain α parameters for all bands. At this time, the low-frequency K parameter data is transmitted alone and reproduced by the decoder 21,
It is clear that the low-band K-parameter data obtained via the buffer memory 24 and the K2 _- parameter linear conversion circuit 30 can be easily incorporated into the low-band spectral envelope calculation operation described above.

After obtaining the α parameter for the entire band in this way, use this α parameter as the filter coefficient.
Configure LPC filter. Furthermore, the normalized predicted residual power for all bands is measured using the low-frequency short-term average power data, the high-frequency short-term average power data, the low-frequency spectrum envelope information, and the high-frequency spectrum envelope information, and this is applied to the LPC filter. The excitation signal is assumed to be a gain control signal of a variable gain amplifier to be supplied.

The variable gain amplifier described above covers the entire band, and therefore only one is required. This variable gain amplifier requires sound source information as input, and this sound source information is the sound source information in the embodiment, that is,
Either the pitch pulse data 2601 or the white noise signal 2901 output from the sound source information switch 25 is input in accordance with whether the voiced/unvoiced discrimination signal 501 is voiced or unvoiced. In this way, it is also possible to easily perform speech synthesis in which low-frequency and high-frequency K parameter information is processed in the frequency domain.

All of the above descriptions can be easily implemented without departing from the spirit of the present invention.

As explained above, according to the present invention, in a band division type vocoder that divides an input audio signal into two predetermined transmission bands, a low frequency band and a high frequency band, and performs a linear predictive analysis, a linear predictive analysis of the low frequency band is performed. K obtained with ₂
By providing a simple means of nonlinearly quantizing and transmitting parameters, the quantization distortion of the audio signal to be transmitted from the analysis side to the synthesis side can be significantly reduced, and the required number of quantization bits and therefore the transmission band can be reduced. This has the effect of realizing a band-splitting vocoder that can significantly improve the compression of the spectral sensitivity of the K ₂ parameter and greatly improve the smoothing of the spectral sensitivity of the K 2 parameter, making it possible to synthesize high-quality speech.

[Brief explanation of drawings]

FIG. 1 is a K-parameter distribution diagram showing an example of the K-parameter distribution of a band-splitting linear predictive vocoder.
FIG. 2 is a block diagram showing an embodiment of the analysis side of the band division type vocoder according to the present invention, and FIG. 3 is a block diagram showing an embodiment of the synthesis side of the band division type vocoder according to the invention. 1...LPF, 2...A/D converter, 3...
...Window processor, 4...DFT circuit, 5...
Voiced/unvoiced discriminator, 6... Pitch extractor, 7...
Power spectrum measuring device, 8...Low frequency autocorrelation coefficient measuring device, 9...High frequency autocorrelation coefficient measuring device, 10
...Low band linear prediction coefficient analyzer, 11...Encoder, 12...K ₂ parameter memory, 13...Nonlinear conversion circuit, 14...High band linear prediction coefficient analyzer, 21...Decoder, 22...Low frequency LPC filter, 23...High frequency LPC filter, 24...Buffer memory, 25...Sound source information switch, 26...
...Pitch generator, 27...Low-range variable gain amplifier,
28... High-frequency variable gain amplifier, 29... Noise generator, 30... K _2- parameter linear conversion circuit, 31
...Low frequency interpolator, 32...Low frequency LPF, 33...
Low-frequency high-frequency synthesizer, 34... High-frequency interpolator, 35...
Frequency converter, 36...High frequency BPF, 37...
D/A converter, 38...LPF.

Claims (1)

[Claims] 1 Low range and high range with predetermined input audio signal 2
In a band division type vocoder that performs linear predictive analysis on each band divided into two audio transmission bands, the second-order K parameter among the partial autocorrelation coefficient K parameters obtained by the linear predictive analysis of the low-frequency audio transmission band. K ₂ that nonlinearly quantizes and transmits K ₂
A band division type vocoder comprising a parameter nonlinear quantization transmission means.