JPH0441840B2 - - Google Patents

Description

[Detailed description of the invention] [Industrial application field] The present invention relates to a pattern matching vocoder.

[Conventional technology]

A pattern matching vocoder that analyzes and synthesizes an input audio signal by comparing a vector pattern of feature parameters obtained by LPC analysis of the input audio signal with a standard pattern regarding the distribution of feature parameters of audio materials is well known. ing.

This LPC analysis obtains the characteristic parameters of the input audio signal by approximating its spectral envelope with an all-pole model. In this case, the input audio signal and audio material are assumed to be stationary, and the α parameter is calculated for each analysis frame LPC coefficients such as and K parameters are extracted.

In a pattern matching vocoder, the analysis side transmits only the specified number of the standard pattern that best matches the input audio signal to the synthesis side, and the synthesis side analyzes the characteristic parameters that should be used as coefficients of the speech synthesis filter based on this information. The transmission bit rate is reduced by preparing the data for each frame.

[Problem that the invention seeks to solve]

However, it is known from many sources that the feature parameters obtained by this LPC analysis are approximate values of the spectral envelope of the input audio signal or audio material, and are likely to cause underestimation of the formant bandwidth. This underestimation of the formant bandwidth is particularly likely to occur in the first formant, and this is thought to be due to excessive concentration of poles at frequencies where energy is concentrated, such as in the first formant. Regarding the spectral envelope, this means that its polar frequency bandwidth, especially the first polar frequency band in the low frequency region, can often be expressed as extremely narrow, and therefore such characteristics Pattern matching performed via parameters has the disadvantage that it is not always possible to select the best pattern, and the quality of reproduced sound deteriorates accordingly.

SUMMARY OF THE INVENTION An object of the present invention is to provide a pattern matching vocoder which eliminates the above-mentioned drawbacks and which makes it possible to always select the best pattern by using normalized prediction residual power as a determination condition.

[Means for solving problems]

In a pattern matching vocoder, the device of the present invention calculates the ratio between the normalized predicted residual power that can be calculated based on the LPC parameter of the input signal or in conjunction with LPC analysis and the normalized predicted residual power regarding the standard pattern. In addition, the apparatus is configured to include a selected pattern matching means for performing pattern matching only on standard patterns whose ratios are within a preset range.

〔Example〕

Next, the present invention will be explained in detail with reference to the drawings.
FIG. 1 is a block diagram showing the structure of a first embodiment of a pattern matching vocoder of the present invention.

The configuration of the embodiment shown in FIG. 1 consists of an analysis side 1 and a synthesis side 2, and the analysis side 1 includes an LPC analyzer 11,
The composition includes a sound source analyzer 12, a pattern matching device 13, a standard pattern memory 14, and a multiplexer 15, and the synthesis side 2 includes a demultiplexer 21,
The device includes a pattern decoder 22, a standard pattern memory 23, a sound source synthesizer 24, a speech synthesis filter 25, a D/A converter 26, and an LPE 27.

An input audio signal input via an input line 101 is supplied to an LPC analyzer 11 and a sound source analyzer 12.

The LPC analyzer 11 converts the input audio signal into an LPF (Low
Pass Filter) to cut off unnecessary high frequencies,
Next, the A/D converter samples the data at a predetermined sampling frequency and quantizes it with a predetermined number of bits. This quantized audio signal is further subjected to window processing by multiplying it with a window function at every predetermined analysis cycle, and then this analysis cycle is used as an analysis frame unit using a known processing technique such as Auto-
Perform LPC analysis using correlation method, PARCOR method, etc. and extract LPC coefficients as spectral envelope characteristics.

In this embodiment, K of a predetermined order is used as the LPC coefficient.
parameter (partial autocorrelation function), but this is based on other α parameters or these
Other feature parameters derived from the LPC parameters may also be used.

The LPC coefficients extracted from the LPC analyzer 11 for each analysis frame are checked in real time against a standard pattern by the pattern checker 13.

The standard pattern memory 14 includes a standard pattern regarding the distribution of the LPC coefficients of the audio material prepared in advance, in this case the K parameter, and the pattern matching device 1
3 measures the spectral distance between the LPC coefficient read from the standard pattern memory 14 and the standard pattern group, selects the standard pattern with the minimum value using this spectral distance as a measure, and sends its designated number to the multiplexer 15. In addition, in selecting this standard pattern, either the LPC analyzer 11 calculates it using the LPC coefficients extracted by the LPC analyzer 11 for each analysis frame, or the LPC analyzer 11 calculates it in the process of LPC analysis. While calculating the ratio between the obtained normalized predicted residual power and the normalized predicted residual power regarding the standard pattern, only the standard patterns whose ratio is within a preset range are compared.

The normalized prediction residual power is defined by U shown in the following equation (1).

U= _p π ^j=1 (1−k ² _j ) ...(1) In equation (1), k represents the K parameter, and p is
This is the analysis order of the K parameter in LPC analysis. The LPC coefficient to be matched is the α parameter, etc.
If it is another LPC coefficient, it is sufficient to convert it into a K parameter and then calculate U according to equation (1). Also, the Levinson method to find the α parameter, etc.
Depending on the processing method, U in equation (1) may be added to LPC analysis.
may be calculated. Note that the standard pattern memory 14 may store the normalized predicted residual power for each built-in standard pattern together with the standard pattern in advance, as in this embodiment, or store the normalized predicted residual power for each built-in standard pattern in advance (1).
There is no problem in using the method in which the data is calculated based on a formula and then supplied to the pattern matching unit 13.

This normalized predicted residual power corresponds inversely to the sharpness of each peak value waveform shown by the spectral envelope, that is, the polar frequency bandwidth of the spectral envelope, and is the reciprocal of the selectivity Q of the spectral envelope waveform in which the polar frequency appears. /Q has almost the same physical meaning. Therefore, the smaller U is, the sharper the spectral envelope waveform centered on the polar frequency becomes. Taking this condition into account, we can eliminate the problem of underestimating the polar frequency bandwidth of the spectral envelope obtained by LPC analysis. It's summery.

That is, the LPC analyzer 11 uses this K parameter as well as the K parameter for each analysis frame to calculate U according to equation (1), and the standard pattern memory 14 also stores U calculated in advance for each standard pattern together with the standard pattern. In advance, each time pattern matching is performed, the pattern matching unit 13 calculates the ratio of the two normalized predicted residual powers prior to pattern matching,
Only those whose ratio is within a predetermined judgment range are determined to be standard patterns for pattern matching. In this case, the decision threshold value is set from the perspective of limiting the difference in the normalized predicted residual power of the two patterns to be matched to be within the error range that should be allowed for LPC synthesis, and within such a decision threshold value. By performing matching only with the standard pattern in
It is possible to perform efficient matching processing that fundamentally eliminates problems in pattern matching using feature parameters based on LPC analysis, that is, problems based on underestimation of the polar frequency bandwidth of the spectral envelope. In the present embodiment, the normalized predictive residual power U of the standard pattern is obtained in advance using equation (1), and is stored in the standard pattern memory 14 in an array sorted in the ascending order of U.

Now, the sound source analyzer 12 extracts sound source information from the input audio signal using a known method. This sound source information includes pitch period information, voiced/unvoiced/silent discrimination information, and information regarding the strength of the sound source, which are each extracted for each analysis frame and supplied to the multiplexer 15. In this embodiment, sound source information modeled using this information is sent from analysis side 1 to synthesis side 2 as sound source information, but multi-pulse or residual excitation type that uses waveform information as sound source information, etc. Of course, there is no problem with it being in the form of .

The multiplexer 15 encodes the input standard pattern designation number and sound source information, multiplexes them in a predetermined format, and sends them out to the synthesis side 2 via the transmission line 102.

On the synthesis side 2, the multiplexed signal supplied from the analysis side 1 is demultiplexed and decoded by a demultiplexer 21, and the standard pattern designation number is supplied to the pattern decoder 22, and the sound source information is supplied to the sound source synthesizer 24. do.

The pattern decoder 22 reads a corresponding standard pattern from the standard pattern memory 23 based on the input standard pattern designation number and supplies the LPC parameter designated by this standard pattern to the speech synthesis filter 25. In this embodiment, the standard pattern memory 23
The same content as the standard pattern stored in the standard pattern memory 14 is stored in the standard pattern memory 14, but the standard pattern prepared for the synthesis side does not necessarily have the same content and the same content as the standard pattern prepared for the analysis side. It is not necessary to use an array, but it is also possible to use an array in which the spectral envelope of the same analysis frame is expressed using other LPC coefficients, or an array convenient for pattern decoding.

Now, the speech synthesis filter 25, which is constituted by an all-pole digital filter of the same order as the analysis order, uses the LPC parameters supplied from the pattern decoder 22 as filter coefficients.
It is driven by a synthesized sound source received from a digital input audio signal.

The sound source synthesizer 24 uses pitch period information, voiced/unvoiced/silent discrimination information, and sound source strength information to model the sound source with pitch period information when voiced and unvoiced, and with white noise when silent. The synthesized sound source is synthesized with a level corresponding to the strength of the sound source and is supplied to a speech synthesis filter 25.

The digital playback input audio signal output from the voice synthesis filter 25 is then converted into an analog signal by the D/A converter 26, and after cutting off unnecessary high-frequency components by the LPF 27, it is sent to the output line 201.
sent to.

FIG. 2 is a block diagram showing the structure of a second embodiment of the pattern matching vocoder of the present invention.

The second embodiment shown in FIG. 2 is similar to the first embodiment shown in FIG.
The LPC analyzer 11 in the embodiment is changed to the LSP analyzer 16, and the standard pattern memories 17 and 28 are changed to the LSP analyzer 16.
The only difference is that parameters are registered, and all other contents are the same, so a detailed explanation of these common contents will be omitted.

The second embodiment shown in FIG. 2 uses LSP parameters derived from LPC parameters as LPC parameters used for pattern matching. As is well known, the LSP parameter is a spectral envelope expressed by multiple line spectral pairs, and can be determined from the LPC coefficients using known methods such as solving higher-order equations using Newton's iterative method or zero point search method. It is widely used in speech synthesis due to its excellent interpolation characteristics in the spectral domain.

However, since this LSP parameter is derived from the LPC parameter, it has the property that the polar frequency bandwidth of the spectral envelope is often underestimated and becomes extremely narrow, and as a result, it is not always the best There is a problem in that it often becomes impossible to select a pattern.

LSP type pattern matching vocoders perform matching through real-time comparison between the distribution of LSP parameters found by analyzing the input audio signal and the LSP parameters stored as standard patterns. Dij expressed by the following equation (2) is used as the distance, and the standard pattern that minimizes this Dij is selected as a pattern matching pattern.

Dij _p 〓 ^K=1 W _K (ω _iK −ω _jK ) ² ...(2) In equation (2), i, j are processing frame numbers in LSP analysis, ω _ik and ω _jk are frames i, j
The P-order LSP coefficient W _k is determined by the P-order LSP spectral sensitivity corresponding to the LSP frequencies ω ₁ , ω ₂ , . . . ω _k , etc.

LSP parameters underestimate the frequency bandwidth of the spectral envelope calculated from this when the distance between adjacent frequencies, for example, ω ₁ and ω ₂ , becomes extremely narrow.
Pattern matching using equation (2) cannot select the best pattern.

In order to avoid such problems, the LSP analyzer 15 calculates the normalized predicted residual power U shown in equation (1) for each analysis frame, and the standard pattern registered in the standard pattern memory 16 LSP
Standard patterns regarding the distribution of parameters are prepared, and normalized predicted residual power for each of these standard patterns is also calculated and stored in advance. Prior to pattern matching, the pattern matcher 13 calculates the ratio of these two normalized predicted residual powers, and uses LSP only for standard patterns within a preset decision threshold value.
The parameter distribution is verified using the spectral distance in equation (2) as a measure.

Calculation of the normalized prediction residual power for LSP parameters is performed by first calculating the α parameter from the LSP parameters ω ₁ , ω ₂ ...ω _k , and then converting the α parameter into the K parameter to calculate the K parameter. Calculate.

Other operations are almost the same as in the first embodiment.
Analysis and synthesis of the input audio signal is performed by pattern matching via LSP parameters.

In this way, efficient pattern matching can be performed that avoids the problem of underestimation of the polar frequency bandwidth of the spectrum envelope.

Note that in the first and second embodiments described above, analysis and synthesis are performed using fixed-length frames, but it is clear that this can also be easily performed using variable-length frames.

〔Effect of the invention〕

As explained above, according to the present invention, in a pattern matching vocoder, the ratio of the normalized predicted residual power regarding the input audio signal and the standard pattern is determined before the pattern is determined, and the ratio is indirectly calculated from the polar frequency bandwidth in the spectral envelope. By providing a means to perform pattern matching only on standard patterns whose corresponding ratios are within a preset threshold, the shortcoming in LPC analysis of underestimating the polar frequency bandwidth of the spectral envelope is greatly improved. This has the advantage that a pattern matching vocoder can be realized.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing the structure of a first embodiment of a pattern matching vocoder of the present invention, and FIG. 2 is a block diagram showing the structure of a second embodiment of the pattern matching vocoder of the present invention. 1... Analysis side, 2... Synthesis side, 11... LPC analyzer, 12... Sound source analyzer, 13... Pattern collation device, 14... Standard pattern memory, 15... Multiplexer, 16... LSP analyzer , 17... Standard pattern memory, 21... Demultiplexer, 22...
...Pattern decoder, 23...Standard pattern memory, 2
4... Sound source synthesizer, 25... Sound source synthesis filter,
26...D/A converter, 27...LPF, 2
8...Standard pattern memory.

Claims (1)

[Scope of Claims] 1. In a pattern matching vocoder, based on LPC (Linear Prediction Coefficient) parameters of input speech,
After calculating the ratio between the normalized predicted residual power that can be calculated along with the analysis and the normalized predicted residual power regarding the standard pattern, pattern matching is performed only for standard patterns for which this ratio is within a preset range. A pattern matching vocoder comprising a selected pattern matching means for performing selected pattern matching. 2 Pattern matching is performed using LSP (Line Spectrum Pairs).
2. The pattern matching vocoder according to claim 1, wherein the pattern matching vocoder is implemented via a line spectrum pair) parameter.