JPH052158B2 - - Google Patents

Description

Detailed Description of the Invention (Industrial Application Field) The present invention relates to a band division type vocoder that uses a linear prediction analysis method, and in particular, to a band division type vocoder that uses a linear prediction analysis method. This invention relates to a band division type vocoder that extracts linear prediction parameters and synthesizes speech.

(Prior art) The purpose of the band-splitting vocoder is to overcome two drawbacks of linear predictive analysis methods: underestimation of formant bandwidth and poor approximation in high-order formants.
However, it is known that as a result of this improvement, discontinuities at the band boundary frequencies appear on the spectral envelope of the synthesized speech. In order to remove this discontinuity, conventionally, linear prediction parameters covering all frequency bands are extracted from linear prediction parameters for each divided frequency band transmitted from the analysis side to the synthesis side, and audio corresponding to all frequency bands is generated. A method has been used in which a linear predictive analysis means for controlling a synthesis filter is provided on the synthesis side. The above content is published in Patent Publication No. 58-220199, "Band division type vocoder"
is described in detail.

(Problems to be Solved by the Invention) However, when linear prediction analysis over all bands is performed on the synthesis side in order to remove discontinuities between bands, the following new drawbacks occur. In other words, this method attempts to remove the spectral discontinuity caused as a result of band division by linear predictive analysis of the entire band, and of course, the discontinuous spectral envelope analyzed on the analysis side and the discontinuous spectral envelope analyzed on the synthesis side are It is obvious that a difference occurs in the calculated spectrum envelope, and from the known properties of the linear prediction method, this difference is spread over the entire frequency band. That is, the conventional method has the disadvantage that distortion occurs in the entire spectral envelope as a result of removing the discontinuity between bands.

SUMMARY OF THE INVENTION An object of the present invention is to provide a band-splitting vocoder that solves discontinuities between bands and does not cause distortion of the entire spectral envelope due to the removal of discontinuities.

(Means for Solving the Problems) The band-splitting vocoder of the present invention is a band-splitting vocoder that uses a linear prediction method, and on the analysis side, overlapping frequency sections are provided between the bands to be analyzed. The synthesis side calculates the spectral envelope of each band from these linear prediction coefficients without creating an overlapping frequency interval between bands, and from these spectral envelopes, the spectrum of the linear prediction coefficients of all bands is calculated. The apparatus is configured to include means for recalculating coefficients expressing the envelope.

(Example) Next, an example of the present invention will be described in detail with reference to the drawings. FIG. 1 is a block diagram showing an embodiment of the present invention. In FIG. 1, the part surrounded by a dashed line 100 is the analysis side of the present invention, and the part surrounded by the dashed dot line 101 is the composition of the synthesis side of the present invention.

The audio signal is sent to the A/D converter 2 via the input terminal 1.
supplied to A/D converter 2 converts the audio signal into
After band-limiting to 3.4KHz, it is sampled at 8KHz and output to window processor 3. Window processor 3 has 256 audio samples
Each set is multiplied by a Hamming coefficient and output to Fourier transformers 1 and 4. Note that the multiplication of the Hamming coefficient is performed by moving every 160 samples, which results in a basic frame period of 20 mS=160/8. The Fourier transformers 1 and 4 perform Fourier transform on the 256 window-processed audio samples using the FFT method and output the obtained complex spectrum data to the power spectrum calculator 5. Power spectrum calculator 5 calculates a power spectrum from a plurality of spectra using a known method, and outputs it to autocorrelation coefficient calculators 1 and 6, autocorrelation coefficient calculators 2 and 7, and logarithm calculator 8.

The autocorrelation coefficient calculators 1 and 6 calculate the unnormalized autocorrelation coefficient φ _L (i), i=0, 1,..., using the following equation (1).
Find 6.

φ _L (i)=p(5)+2 ₃₈ 〓 ^j=6 p(j)cos [(j-5)iπ/128] (1) However, p(k), k=1, 2,..., 256 is a discrete power spectrum when 8KHz is divided into 256, and p(k) is a power spectrum at (k-1)×4000/128Hz. In other words, autocorrelation coefficient calculators 1 and 6 are 125
Discrete Fourier transform is applied to the frequency spectrum in the range of ~1156.25Hz, and the unnormalized autocorrelation coefficient φ _L (i), i=
We are looking for 0, 1,..., 6.

Furthermore, the autocorrelation coefficient calculators 1 and 6 are connected to the output line 6.
01, outputs φ _L (o) corresponding to power information to the quantizers 1 and 10. Next, the autocorrelation coefficient calculators 1 and 6 calculate the normalized autocorrelation coefficients ρ _L (i), i=1, 2, ..., 6 using the following equation (2), and the linear prediction analyzers 1 and Output to 9.

ρ _L (i) = φ _L (i) / φ _L (o) (2) The autocorrelation coefficient calculators 2 and 7 calculate the denormalized autocorrelation coefficient using the following formulas (3) and (4). φ _H (i), i=0,1,
..., 6, and the normalized autocorrelation coefficient ρ _H (i), i=
Find 1, 2,..., 6, and output φ _H (o) to the output line 70.
1 to the quantizers 2 and 12, ρ _H (i), i=1,
2, . . . , 6 are output to the linear predictive analyzers 2 and 11.

φ _H (i)=p(35)+2 ₁₀₁ 〓 ^j=36 p(j)cos [(j−35)iπ/128] (3) ρ _H (i)=φ _H (i)/φ _H (o )(i=1,2,...,6) (4) That is, autocorrelation coefficient calculators 2 and 7 are ^1062.5 to 3125
It uses frequency components in the Hz range. ^1062.5Hz
~1156. ²⁵ Hz part is autocorrelation coefficient calculator 1, 6
One of the key points of the present invention is that the autocorrelation coefficient calculators 2 and 7 are used for analysis.

The linear predictive analyzers 1 and 9 and the linear predictive analyzers 2 and 11 are supplied with ρ _L (i), ρ _H (i), (i=1,
..., 6), each of 6 K parameters K _L (i), K _H (i),
(i = 1, 2, ..., 6) using a known method, for example, the autocorrelation method, and send K _L (i) to the quantizers 1, 10 via the output line 901 or 1101,
K _H (i) is output to the quantizers 2 and 12.

Quantizers 1 and 10 are supplied with φ _L (o) and K _L (i),
(i)=1, . . . , 6) is quantized with a predetermined number of bits and output to the multiplexer 13. Similarly, quantizers 2 and 12
quantizes the supplied φ _H (o) and K _H (i) (i=1, . . . , 6) using a predetermined number of bits and outputs it to the quantizer 13.

On the other hand, the logarithm calculator 8 calculates the logarithm value of the supplied power spectrum by table pickup and outputs it to the Fourier transformers 2 and 14. The Fourier transformers 2 and 14 perform Fourier transform on the logarithmic power spectrum, that is, calculate the cepstral coefficients, Pitch,
Output to the V/uv discriminator 15. Pitch, V/uv discriminator 15 uses a known method using cepstral coefficients as input, for example, A. Michael Noll "Short-Time".
Spec−trum and “Cepstrum” Techniques for
Vocal-Pitch Detection”, The Journal of
Acoustic Society of America, vol.36,
Pitch period and V/uv are determined using the method described in No. 2, Feb. 1964, pp. 296-302, and the results are quantized to a predetermined number of bits by quantizers 3 and 16 and multiplexed. output to the device 13.

Now, on the analysis side, quantizers 1 and 10, quantizer 2,
12. The data quantized into predetermined bits by the quantizers 3 and 16 is sent to the multiplexer 13 and the transmission line 1.
7. Demultiplexer 1, 1 via demultiplexer 18
9, the signal is supplied to the decoders 2 and 20 and the decoders 3 and 21 for decoding. φ _{′ L} (o) and K _{′ L} (i), (i)=1,...,6) decoded by the decoders 1 and 19 are sent to the envelope calculators 1 and 22, and then sent to the envelope calculators 1 and 22. φ _{′ H} (o) and K′ _H (i), (i=1, . . . , 6) decoded in step 20 are supplied to envelope calculators 2 and 23, respectively.

The envelope calculators 1 and 22 calculate K′ _L (i), (i=1,...,
6), the normalized predicted residual power u _L and the α parameters α _L (i), (i=1,..., 6) are obtained, and the power spectrum envelope is calculated using the following equation (5).

However, A _L (o)= ₆ 〓 ^t=0 α ² _L (t), A _L (r)=2 _6-r 〓 ^t=0 α _L (t)・α _L
(t+r), α _L (o)=1.

Similarly, the envelope calculators 2 and 23 use K′ _H (i), (i=1,
..., 6), find the normalized predicted residual power u _H and the α parameter α _H (i), (i=1,..., 6), and use the following (6).
Calculate the power spectrum envelope using the formula.

However, A _H (o)= ₆ 〓 ^t=0 α ² _H (t), A _H (r)(r)=2 _6-r 〓 ^t=0 α _H
(t)・α _H (t+r), α _H (o)=0.

The filter coefficient calculator 24 calculates the above p _L (i), (i=5,
. . , 36) and p _H (i), (i=37, . . . , 101), the filter coefficients representing the spectrum envelope of the entire band are calculated and output to the speech synthesis filter 25.

Now, the supplied p _L (i), (i=5,...,36) is
Discrete spectral envelope from 125 to ^1093.75 Hz and p _H
(i), (i=37,..., 101) are discrete spectral envelopes from 1125 to 3125 Hz. Therefore, 31, 25 = 1125−1093.
⁷⁵ Hz is the frequency sample interval. That is, the frequency components used on the analysis side are ^1093.75 to ^1156.25 Hz and
^1062.5 to 1125 Hz is not used by the filter coefficient calculator 24. This point is another key point of the present invention.

As described above, in the band analysis type vocoder of the present invention, the analysis side includes and analyzes components near the band boundaries that are extra than a predetermined band, and the synthesis side removes the extra components. This significantly reduces spectral envelope discontinuities caused by band boundary extraction. That is, the distortion of the spectral envelope caused by boundary cutting tends to concentrate near the boundary, and this is because the vicinity of this boundary is cut off on the synthesis side.

The filter coefficients calculated by the filter coefficient calculator 24 may be those of an all-pole filter or all-zero filter. When determining the coefficients of an all-pole type filter, refer to the above-mentioned published patent publication Sho.
For example, the 16th-order α parameter may be calculated by the method of determining the linear prediction coefficient of the entire band through the autocorrelation coefficient described in 58-220199, page 644. Also, when calculating the coefficients of an all-zero type filter, √ _L (i), (i=5,...,36), and √ _H (i),
(i=37,...,101) is Fourier transformed, for example
All you have to do is calculate the coefficients of the 129-point transversal filter.

On the other hand, the pitch decoded by the decoders 3 and 21,
The V/uv signals are supplied to a pitch generator 26 and a switch 27, respectively. The pitch generator 26 generates a pitch pulse train based on the pitch information and the switch 27
Output to. The noise generator 28 generates white noise and outputs it to the switch 27. Switch 27 is V/
The uv signal generates a pitch pulse train when voiced and white noise when unvoiced, and sends it to the voice synthesis filter 25.
Output to. The speech synthesis filter 25 synthesizes speech and outputs it to the D/A converter 29. D/A converter 2
Reference numeral 9 reproduces analog audio and outputs an audio waveform via an output terminal 30.

(Effects of the Invention) As described above, in a band division type vocoder, the analysis side includes and analyzes the components near the band boundaries that are extra than the predetermined band, and the synthesis side removes the extra components. This has the effect that discontinuities in the spectral envelope caused by band boundary extraction can be significantly reduced and a very high quality vocoder can be realized.

[Brief explanation of the drawing]

FIG. 1 is a block diagram for explaining an embodiment of the present invention, and reference numbers 1 to 16 indicate the analysis side, 17 indicates the electrical transmission path, and 18 to 30 indicate the synthesis side. 2...A/D converter, 3...Window processor, 4...
Fourier transformer, 1, 5...power spectrum calculator, 6...autocorrelation coefficient calculator, 1,7...autocorrelation coefficient calculator, 2,8...logarithm calculator, 9...
Linear prediction analyzer, 1, 10...Quantizer, 1, 1
1... Linear prediction analyzer, 2, 12... Quantizer,
2, 13... Quantizer, 14... Fourier transformer, 2, 15... Pitch, V/uv discriminator, 16...
...Quantizer, 3, 18... Demultiplexer, 19...
Decoder, 1, 20...Decoder, 2, 21...
Decoder, 3, 22... Envelope calculator, 1, 23...
...Envelope calculator, 2, 24...Filter coefficient calculator, 25...Speech synthesis filter, 26...Pitch generator, 27...Switch, 28...Noise generator, 29...D/A converter.

Claims (1)

[Claims] 1. In a band division type vocoder that uses a linear prediction analysis method, the analysis side includes means for providing overlapping frequency intervals between bands to be analyzed and calculating linear prediction coefficients for each band, and a means for calculating linear prediction coefficients for each band, and There is a means for sending the prediction coefficients to the analysis side, and a means for the synthesis side to calculate the spectral envelope for each band from these linear prediction coefficients calculated on the analysis side without creating overlapping frequency sections between bands, and from these spectral envelopes. What is claimed is: 1. A band-splitting vocoder comprising means for calculating coefficients of a speech synthesis filter corresponding to the spectral envelope of all bands, and a speech synthesis filter for synthesizing speech based on the coefficients.