JP2535809B2 - Linear predictive speech analysis and synthesis device - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a speech analysis / synthesis apparatus, and in particular, to expanding an operation amplitude range of a speech analysis / synthesis apparatus having means for adjusting a gain of an output waveform of a speech synthesis filter. The present invention relates to a voice analysis / synthesis device.

(Prior Art) A speech analysis / synthesis apparatus analyzes speech on the analysis side to extract characteristic parameters, and resynthesizes the parameters on the synthesis side. Usually, the scale of the apparatus is finite precision with a fixed point and relatively small. A voice synthesis filter that uses computation is provided to synthesize a voice waveform. In speech synthesis by such an analysis-synthesis device, it is not sufficient to maintain the quality of the synthesized speech waveform with high quality, and the dynamic range of the speech synthesis filter, S / N (Signal to No
For the purpose of supplementing the ratio, the gain adjustment corresponding to the strength of the sound source of the input voice is performed on the output side of the voice synthesis filter, not on the input side.

[Problems to be solved by the invention]

However, the conventional voice synthesizing means of this kind has the following problems to be further improved.

In other words, the speech synthesis filter has its spectral transmission characteristics associated with the vocal tract transmission characteristics of the voice speaker, and its filter coefficient also changes from time to time as the spectrum envelope shape of the speech waveform to be analyzed changes. To do. Therefore, the filter gain of the voice synthesis filter, that is, the ratio of the power of the input signal to the power of the output signal also changes from moment to moment.
Therefore, in the conventional voice analysis / synthesis device having means for performing gain adjustment on the output waveform of the voice synthesis filter, the input signal level of the voice synthesis filter is kept constant in consideration of the maximum gain of the synthesis filter known empirically. There is. Therefore, when the gain of the synthesis filter is small, the output waveform level of the speech synthesis filter becomes relatively small, and the dynamic range of the speech synthesis filter and the complement of the S / N ratio, which is the object of the present invention, cannot be sufficiently achieved. There is.

Another object of the present invention is to eliminate the above-mentioned drawbacks and to provide a speech analysis / synthesis apparatus capable of sufficiently achieving the dynamic range and S / N ratio of a speech synthesis filter.

[Means for solving problems]

The apparatus of the present invention has a function of performing gain adjustment on the output waveform of the voice synthesis filter, and is configured to have a means for calculating the normalized prediction residual power on the synthesis side.

〔Example〕

 The present invention will now be described in detail with reference to the drawings.

FIG. 1 is a block diagram showing the configuration of an embodiment of a voice analysis / synthesis apparatus according to the present invention.

The embodiment shown in FIG. 1 comprises an analysis side 1 and a synthesis side 2, and the analysis side 1 further includes a window processor 11, an LPC analyzer 12, a pitch extractor 13, a power calculator 14, a K quantizer 15, Pitch quantizer 1
6, a power quantizer 17, and a multiplexer 18, and the synthesis side 2 includes a demultiplexer 19 and a K decoder 2
0, pitch decoder 21, power decoder 22, normalized prediction residual power calculator 23, 1/2 multiplier (1) 24, multiplier (1) 25, K / α converter 26, LPC synthesis It comprises a filter 27, a multiplier (2) 28, a pitch pulse speaker 29, a noise speaker 30, a switcher 31, and a 1/2 multiplier (2) 32.

In FIG. 1, an input audio signal input via an input line 1001 is subjected to window processing by a window processor 11. This window processing is performed as follows.

The window processor 11 low-pass filters the input audio signal, and then performs quantization with a predetermined number of bits by an A / D (Analog to Digital) converter.

In the case of this embodiment, the low-pass filtering has a cutoff frequency of 3.
A frequency component above this frequency is cut through a 4KHz LPF (Low Pass Filter), and then a sample sampled at a sampling frequency of 8MHz is quantized with 12 bits per sample and then temporarily stored in the internal memory. This internal memory stores a preset time portion of the input quantized signal, for example, 30 mSEC portion, that is, a window time portion of 240 samples, and multiplies this by a window function such as a Hamming function or a rectangular function to cut out a predetermined window processing. The cycle is 10 mSEC cycles in this embodiment, and this is the analysis frame cycle.

In this way, the quantized speech signal for each analysis frame output from the window processor 11 is the LPC analyzer 12 and the pitch extractor 13.
And the power calculator 14 are supplied.

The LPC analyzer 12 performs LPC analysis on the input quantized speech signal for each analysis frame to extract LPC of a predetermined order, in this embodiment, a tenth-order K parameter (partial autocorrelation coefficient) and outputs it to the K quantizer 15. Output. The K quantizer 15 quantizes the supplied K parameter into a predetermined number of bits and outputs it to the multiplexer 18. The pitch extractor 13 determines the voice / unvoice of the quantized voice signal for each analysis frame input by a known method,
Further, the pitch period is extracted and output to the pitch quantizer 16.
The pitch quantizer 16 collectively quantizes the supplied voiced / unvoiced discrimination result and pitch period information into a predetermined number of bits, and outputs it to the multiplexer 18. The power calculator 14 calculates the power of the input quantized speech signal for each analysis frame and outputs it to the power quantizer 17. The power quantizer 17 quantizes the supplied power information into a predetermined number of bits and outputs it to the multiplexer 18.

The multiplexer 18 multiplexes the input thus supplied in a predetermined format and transmits it to the combining side 2 via the transmission line 101.

On the synthesizing side 2, the demultiplexer 19 demultiplexes the input, the LPC code is sent to the K decoder 20, and the pitch information (including voiced / unvoiced discrimination information) is sent to the pitch decoder 21.
Then, the power information is supplied to the power decoder 22 and decoded respectively.

The K decoder 20 is a K / α converter for the decoded K parameter.
26 and the normalized prediction residual calculator 23. The normalized prediction residual power calculator 23 calculates the normalized prediction residual power U by the following equation (1).

In equation (1), K _p (p = 1.2 ... 10) is a K parameter. U calculated by the normalized prediction residual power calculator 23
Output to 1/2 multiplier (1) 24. 1/2 multiplier (1) 24 is based on the table backup method Is calculated and output to the multiplier (1) 25.

The K / α converter 26 calculates an α parameter (α ₁ , α ₂ ... α ₁₀ ) from the K parameter using a known method and outputs it to the LPC synthesis filter 27.

The 1/2 multiplier (2) 32 is based on the power information P supplied from the power decoder 22 by the table lookup method. Is output to the multiplier (2) 28 which has obtained.

The pitch decoder 21 outputs the decoded pitch period information to the pitch pulse voice generator 29 and the voiced / unvoiced discrimination information to the switch 31. Pitch pulse utterer 29 utters a pitch pulse train having a period that matches the pitch period information and outputs it to switch 31. The noise utterer 30 utters white noise using, for example, a 15th-order M sequence and outputs it to the switch 31. Based on the voiced / unvoiced information, the switch 31 selects the pitch pulse train when voiced and the white noise when unvoiced and supplies it to the multiplier (1) 25. Multiplier (1) 25 is the above The pitch pulse train or the amplitude of white noise using The result is multiplied and output to the LPC synthesis filter 27. LPC synthesis filter 27
Is supplied from the multiplier (1) 25 using the α parameter supplied from the K / α converter 26 as a filter coefficient. The pitch pulse train or the white noise thus generated is used as an input to synthesize a voice (output to the multiplier (2) 28).

By the way, the gain of the LPC synthesis filter is roughly It is empirically known to be. This is because the ratio between the power of the impulse response waveform of the LPC synthesis filter and the power of the input impulse is accurate. It can be easily inferred from Therefore, if the power of the pitch pulse train or the white noise train of the multiplier (1) 25 is made constant, the output power of the LPC synthesis filter 27 becomes almost constant, and the amplitude of the synthesis waveform becomes close to the upper limit of the operating range of the LPC synthesis filter 27. Is extremely easy to set.

The voice waveform synthesized by the LPC synthesis filter 27 is output by the multiplier (2) 28. Output line 2001 after being associated with analysis side power
Sent to

〔The invention's effect〕

As described above, according to the present invention, by having the function of performing gain adjustment on the output waveform of the speech synthesis filter and the function of calculating the normalized prediction residual power on the synthesis side, the speech synthesis filter is always operated. Since it can be used near the upper limit of the dynamic range, there is an effect that it is possible to realize a voice analysis / synthesis device with significantly improved synthetic sound quality.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of an embodiment of a voice analysis / synthesis apparatus according to the present invention. 1 ... Analysis side, 2 ... Synthesis side, 11 ... Window processor, 12 ...
LPC analyzer, 13 …… Pitch extractor, 14 …… Power calculator, 1
5 ... K quantizer, 16 ... pitch quantizer, 17 ... power quantizer, 18 ... multiplexer, 19 ... demultiplexer, 20 ... K decoder, 21 ... pitch decoder ,twenty two…
… Power decoder, 23 …… Normalized predicted power calculator, 24 ……
1/2 multiplier (1), 25 ... multiplier (1), 26 ... K / α converter, 27 ... LPC synthesis filter, 28 ... multiplier (2), 29
...... Pitch pulse generator, 30 …… Noise generator, 31 …… Switcher, 32 …… 1/2 multiplier (2).

Claims (1)

(57) [Claims] 1. A calculation means provided on a synthesis side for calculating a normalized prediction residual power, and a normalized prediction residual power value calculated by the calculation means for an input waveform of a speech synthesis filter on the synthesis side. It has a first adjusting means for adjusting the gain and a second adjusting means for adjusting the gain of the output waveform of the voice synthesis filter with the power value analyzed on the analysis side and sent to the synthesis side. A linear predictive speech analysis / synthesis device characterized by comprising: