JPH0414813B2 - - Google Patents

Abstract

PURPOSE:To obtain a good sentence intelligibility even in case of encoding with a low bit rate by dividing the time series of spectrum parameters, which are extracted in frame units of an input voice, to segments each of which consists of plural frames and encoding them. CONSTITUTION:The voice input from a terminal 11 passes an LPF12 and is sampled periodically and is converted to a digital signal by an A/D converter 13 and is inputted to an LPC analyzing part 14. In the analyzing part 14, the spectrum parameter time series of the input voice is calculated in frame units and is divided to segments on a basis of the change rate of the spectrum by a segmentation part 15, and segments are resampled by a means 16 so that the number of samples of the parameter time series in each segment is fixed. Resampled segments are matched with standard pattern from a memory 23 in segment unit by a matrix quantizing part 22, and the standard pattern approximating most the segment is outputted, and this pattern and a signal quantized by a quantizing part 27 are synthesized by a multiplexer 26, and this voice code output is restored in an LPC synthesizing part 36 and is subjected to D/A conversion by a D/A converter 37 and is outputted through an LPF38.

Description

DETAILED DESCRIPTION OF THE INVENTION "Field of Industrial Application" The present invention relates to a speech encoding method for extracting spectral parameters of input speech and encoding them with low bits.

"Prior Art" Conventionally, there are two methods for encoding audio at a low bit rate of 1000 bps or less: vector quantization and variable frame rate encoding. The vector quantization method quantizes the spectral parameter information per frame to about 8 bits while keeping the frame unit (sound analysis unit) constant, and is characterized by treating the parameters as one vector. However, this method only removes spatial, ie, frequency, redundancy, and when it becomes less than 500 bps, the quality deteriorates rapidly because the frame unit is constant.

On the other hand, the variable frame rate encoding method changes the frame unit (frame length) in response to temporal changes in the spectrum, and removes temporal redundancy, but the average transmission speed is about 1/3. There is little deterioration in quality even if the amount is reduced. However, this method inherently relies on the (linear) interpolation characteristics of the parameters, resulting in a transmission rate of 25 frames per second (overall
600 bps) or less, rapid quality deterioration occurs.

An object of the present invention is to provide a speech encoding method that can reproduce speech with good text intelligibility even at a low bit rate of 600 bps or less.

"Structure of the Invention" This invention describes the spatial (frequency)
This eliminates not only the redundancy of the data, but also the temporal redundancy.
Therefore, in this invention, the spectral parameters of the input audio are extracted frame by frame, the time series of this spectral parameter is divided into segments consisting of multiple frames based on the rate of change of the spectrum, and each divided segment is , the spectral parameter time series is resampled so that the number of sample points is the same. In other words, in order to obtain a sequence pattern of spectral parameters that repeatedly appears temporally and preferably spatially (frequency) in speech, the spectral parameter time series of speech is divided into segments, for example, at phoneme or syllable boundaries. . For each segment,
The parameter time series is resampled so that the number of sample points becomes the same, the time length of each segment is normalized, and temporal redundancy is removed. The parameter time series resampled in this way is encoded by matching it with a standard pattern for each segment.

Embodiment FIG. 1 shows an embodiment of the speech encoding method of the present invention. The audio input from the input terminal 11 is band-limited by the low-pass filter 12 and input to the AD converter 13, where it is periodically sampled and converted into a digital signal. The output of this AD converter 13 is LPC
The analysis unit 14 extracts the spectral parameters of the input audio on a frame-by-frame basis. The parameter time series calculated by LPC analysis is divided into segments each consisting of a plurality of frames based on the rate of change of the spectrum of the spectral parameter time series in the segmentation unit 15. In addition to LPC coefficients, LAR, PARCOR coefficients, LPC cepstrum coefficients, LSP, etc. may be used as spectral parameters extracted from input speech, but in this example, weighted least squares approximation coefficients of LPC cepstrum coefficients are used. It was segmented into an average of 12 frames per second. For details of this segmentation, see, for example, Sagayama and Itakura: Personal information included in dynamic measures of speech, Acoustical Society of Japan Spring Conference Proceedings 3-
2-7 (1979). In other words, the frequency (spectrum) components included in a speech waveform are generally not uniform over time, but are based on the phonology (content of the speech) being spoken.
is heavily dependent on. This situation of heterogeneity also occurs when the phonemes and words are the same. The environment before and after speaking, the speed of speech, etc. change greatly.
On the other hand, it is known that even in such highly non-uniform audio spectrum time series, the reproducibility of the maximum point of change rate is high. Therefore, the spectral parameter time series of speech is divided into segments based on the rate of change of the spectrum, that is, divided at phoneme and syllable boundaries to obtain highly reproducible spectral parameters.

These divided segments may be long or short, but so that the number of sample points of the parameter time series in each segment is constant (same number), a predetermined number is set at equal intervals in time for each segment. The time length of each segment is normalized to the same length by resampling by the resampling unit 16, and temporal redundancy is removed. The greater the number of resamples in one segment, the smaller the spectral distortion will be when restored.As shown in Figure 2, if the number of resamples in each segment is 10 or more, the spectral distortion will be reduced to ^1dB2 or less. Being held down. In Figure 2, the horizontal axis shows the number of resamples per segment, and the vertical axis shows the spectral distortion when restored, and the spectral parameters are LSP
It is. In the embodiment of the present invention, the resampling unit 16 uses LSP as the parameter and 10 as the number of resamples.

FIG. 3 shows a specific example of the resampling unit 16. The segmented spectrum time series from the terminal 17 is separated by the signal separation unit 18 into a spectrum time series and a segment length (continuation length). The continuation length is input to the proportionality constant section 19, and a desired proportionality constant is calculated from the predetermined number of resamplings (10 in the embodiment) and the input continuation length. The resampling period is determined by division and sent to the linear interpolation section 21. The linear interpolation section 21 resamples the segmented spectrum time series from the signal separation section 18 by a predetermined number at the resampling period. At this time, since the resampling time does not match the sampling time of the spectral time series before resampling, the spectral time series is linearly interpolated to obtain the resampled value. In other words, the resampled value [L] _i of the spectral parameter time series at the resampled time T i is the sampled value [L _{] j} at the original sample time t _j immediately before the time T _i , and the original sample time t _j+1 immediately after the resampled time T i. From the sample value [L] _j+1 , [L] _j + {[L]
_j+
1 − [L] _j }×T _i −t _i /t _j+1 −t _j .

Returning to the explanation of Figure 1, the training speech is input, segmented as described above, and resampled, and the resulting time length is normalized to create a highly reproducible spectral pattern similar to Standard patterns are created by grouping and classifying items, and these standard patterns are stored in a memory 23. In this way, a standard pattern with frequency (spatial) redundancy removed is obtained. Next, when normal audio input is quantized, the matrix quantizer 22 uses the resampled segment as a unit to match it with a standard pattern created in advance from the memory 23, and matches it with the standard pattern that is the most similar. Outputs the number of The same scale (distance) calculation is performed in both standard pattern generation and matching. That is, a segment is treated as a matrix (parameter time series arranged), and the distance between the matrices is defined as a weighted Euclidean distance. In this example, the 12th order
The 10 parameters of LSP (L ₁ , L _{2 .} . . L ₁₂ ) and audio power (logarithm) P ₁ are arranged horizontally to form a 13×10-order (segment) matrix.

P ₁ L ₁₁ L ₂₁ 〓 〓 L ₁₂₁ P ₂ L ₁₂ L ₂₂ 〓 〓 L ₁₂₂ …… …… …… …… P ₁₀ L ₁₁₀ L ₂₁₀ L ₁₂₁₀ How to make a standard pattern, for example, in addition to A-Buzo “Speech Coding based upon Vector
Quantization” IEEE, ASSP−283Vol5.pp562−
See pp574 (1980).

In FIG. 4, the horizontal axis represents the amount of information per segment, and the vertical axis represents CD (cepstral distance). Curve 24 represents the power and spectrum when the power and spectrum are quantized separately, and curve 25 represents the power and spectrum described above. is regarded as one vector. It can be seen from FIG. 4 that the curve 25 including the power compresses information by 3 bits/segment more than the curve 24 that separates the power.

The spectral order sequence is encoded in the matrix quantization section 22 as described above, and this is combined with the pitch information of the input voice and the duration information of each segment in the multiplexer 26 and output. Here, the pitch information is the fundamental frequency of the voice, which determines the pitch of the voice. The pitch information is detected, for example, by the method described in the document "Fundamentals of Speech Information Processing" by Saito and Nakata (Published by Ohmsha, 1981, pp. 88-89). In this example, after smoothing the pitch information,
The point pitch is set to one point per segment, and the quantization unit 2
7 and quantized with 3-bit ADPCM. As shown in FIG. 5, the audio digital signal train from the terminal 28 is subjected to LPC analysis in the LPC analysis unit 29, and the LPC parameters, voicing,
The unvoiced determination sequence is output to the audio power multiplexing section 31, and the pitch is supplied to the pitch smoothing section 32, where it is smoothed and then supplied to the multiplexing section 31.

Also, the duration of the segment should be determined by considering the frequency.
The signal is quantized to 2.5 bits by the quantizer 27. In the specific example explained above, the average is 12 segments per second, the standard putter is 10 bits, the pitch is 3 bits,
The continuation length is quantized to 2.5 bits. 3000 samples (4 minutes of audio) obtained as above
The curve 33 in FIG. 6 shows the result of matrix quantization and matching with the generated standard pattern. In FIG. 6, the horizontal axis is LSP bits/second, the vertical axis is spectral distortion, and curve 34 is for the conventional vector quantization method. From FIG. 6, it can be seen that in order to suppress spectral distortion to the same level, this invention requires only one-fourth the information amount of the vector quantization method, and therefore one-tenth of the information amount of the conventional scalar quantization method. The amount of information is sufficient. In the above example, 12× per second
(10+3+2.5)=186 bps, and the intelligibility of the text was good.

Note that the audio encoded output from the multiplexer 26 is transmitted or stored and decoded by the LPC synthesis unit 3.
In step 6, a standard pattern is obtained from the matrix quantization code by referring to a dictionary, and a pattern corresponding to the original LPC analysis output is restored from the segment duration information and pitch information, and this is converted into analog by the DA converter 37. , outputs an analog audio signal to an output terminal 39 through a low-pass filter 38.

"Effects of the Invention" As explained above, according to this invention, approximately
Good text intelligibility can be obtained even at extremely low speeds such as 200bps, so it has the advantage of being useful for making effective use of transmission paths and configuring communication paths with high privacy.

The sending of pitch information and segment continuation information is not limited to the above quantization, but other quantization may be applied.
Does not need to be quantized.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing an example of the present invention, FIG. 2 is a diagram showing the relationship between the number of resamplings and spectral distortion, and FIG. 3 is a block diagram showing a specific example of the parameter resampling unit 16. Figure 4 shows the relationship between the number of bits per segment and spectral distortion.
The figure is a block diagram showing an example of the LPC analysis section 14.
FIG. 6 is a diagram showing the relationship between bits/second and spectral distortion. 14...LPC analysis unit, 15...Segmentation unit, 16...Resampling unit, 22...Matrix quantization unit, 23...Standard pattern memory.

Claims (1)

[Claims] 1 A means for extracting spectral parameters of input speech in frame units, and a time series of the extracted spectral parameters divided into intervals corresponding to phoneme or syllable boundaries based on the rate of change of the spectrum to create segments consisting of multiple frames. means for resampling the parameter time series so that the number of sample points of the parameter time series within that segment becomes the same predetermined number for each divided segment; means for encoding the converted parameter time series using a standard pattern of spectral parameter time series in units of segments, and means for encoding pitch information of the input voice and duration information of each segment. Speech encoding method.