JPS595917B2 - Onseigouseisouchi - Google Patents

Abstract

PURPOSE:Voice synthetic device which uses residual waveform as sound source.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to an improvement in a speech synthesis device, particularly in a speech synthesis device using linear prediction coefficients or equivalent partial autocorrelation coefficients.

As voice response devices become more practical, the types of words that can be answered (
There has been a strong demand for the development of a response device with a large number of words.

In response to this, speech synthesis methods using linear prediction coefficients or equivalent partial autocorrelation coefficients have been developed. In this method, the quality is degraded by replacing the sound source with pulses and random noise.

SUMMARY OF THE INVENTION An object of the present invention is to improve the above-mentioned drawbacks and to provide a speech synthesis device that minimizes distortion of its output speech.

In order to achieve the above object, the present invention uses a waveform after extracting a linear prediction coefficient or a partial autocorrelation coefficient from a speech waveform, that is, a so-called prediction residual waveform, as a sound source.

An example of the residual waveform is shown in Figure 1 (lower side).
In a voiced sound section where the original speech (shown at the top) is approximately periodic, it consists of a pulse-like component corresponding to the period and an aperiodic component (pure prediction error component) in the background.

Furthermore, in a voiceless consonant section where the original speech is aperiodic, it consists of almost random prediction error components only.

Now, in voiced sound sections with periodic components, the periodicity is more pronounced than in the original speech waveform. This means that the period (
This is evidenced by the fact that pitch period detection based on the correlation of residual waveforms is currently the most reliable pitch period detection method. Therefore, consider the following. C Analysis time width T
Analyze speech waveforms in seconds and extract linear prediction coefficients or partial autocorrelation coefficients.

2) Extract a residual waveform as a prediction error according to the extracted coefficients.

3) Detect the periodicity from the correlation of the residual waveform.

When the shift time is set to , the time at which the normalized residual correlation is the next largest after 0 = 0. shall be. When the residual correlation at ro is greater than a preset threshold value θ, it is regarded as a voiced sound. Let be its period. The correlation value is θ
When it is smaller, it is considered a voiceless sound. (θ is 0.1
4) R based on the residual waveform.

Find and record a representative residual waveform between the two. 5) During synthesis, use this R in voiced sound sections.

The second residual waveform is used repeatedly for T seconds. R. According to the definition of O, if it is a voiced sound, this T seconds has a period R.

It will be synthesized using the sound source. In the unvoiced section, random noise is used as input, as in normal synthesis. 6) It goes without saying that this unvoiced sound source is controlled by the sound source strength extracted at the time of analysis.

Assuming that T is 10 to 20 milliseconds, RO is 6 to 8 milliseconds on average for men, and 3.3 milliseconds for women, by the above procedure, the information of the residual waveform for men is: 20 milliseconds /
(6-8) ~1/3-1/2 For women: (20-10
)/3.3 to 1/6 to 1/3.

Experiments on direct pitch period segment transmission of voice waveforms have also shown that it is possible to compress the voice waveform to about 1/6.

Therefore, it is thought that almost the same degree of compression is possible for the residual waveform as well. An embodiment for extracting a residual waveform based on the above principle will be explained with reference to the block diagram of FIG.

The speech waveform 100 of the word to be synthesized is amplified to an appropriate level and band-limited through a preamplifier 1 low frequency or band preamplifier 2, and then converted into a digital signal by an A/D converter 3. , and stored in the buffer memory 4. When the digital input of one word waveform is completed, Tf! from the beginning of the data is completed. > is read out to the analysis register 5, and the linear prediction coefficient analysis circuit 6 extracts linear prediction coefficients {α1}1=1 to p and stores them in the control information memory 12.

It is sufficient to set p to 8 to 12. The linear prediction coefficient is given by solving the following simultaneous linear equations (Yule-Walker equation) whose coefficients are the correlation coefficients {δj} (j=0 to p) of the input data.

However, in the above equation, it is possible to set δ-k=δk (k=1, ~p).

Here, δk is the data to be analyzed {Xt+1}, i=1
It is given when ~n.

Based on the extracted preliminary line coefficients {α,}, a prediction residual Rt is calculated by the following equation and stored in the residual waveform register 7.

On the other hand, the correlation circuit 8 calculates the correlation from this residual waveform, and the maximum value is detected by the maximum value detection circuit 9 to determine the pitch period τ.

The threshold circuit 10 determines whether the analysis section is voiced or unvoiced, and if voiced, τ from the beginning of the residual waveform. Data for seconds is stored in the control information memory 12. At the same time, the average intensity of the residual waveform is calculated by the intensity extraction circuit 1.
2 and stored in the memory 12 together with the output (voiced/unvoiced) of the threshold circuit 10. Thereafter, this analysis process is repeated to convert the waveforms of all word sections into control information necessary for synthesis and sequentially store it in the memory 12.

As a result, information as shown in FIG. 4 is stored in the control information memory 12 every T seconds.

Section number N: 400 Voiced/unvoiced: 401 (1: voiced, 0: unvoiced) Sound source strength A: 402 (O for a voiced section) Pitch period τ.

:403 linear prediction coefficient {α1}: 404, i=1~p residual waveform {Rt}: 405, t=1, ~τ (none for unvoiced sound) The amount of information is on average 8+1+8+8+8X12+8X30=361 sets,
If T is 10 milliseconds, then T is 36.1 kbits/sec.
If it is 20 milliseconds, then it is 18.5k bits/second.

Therefore, compared to direct recording of audio waveforms, the former is about 1/2
2.5, the latter compresses information by 1/5.

The pitch period can also be detected on the synthesis side by recording a special end mark at the end of the recorded residual waveform and detecting it. This method saves 800 bits per second. Next, an embodiment of the present invention will be described based on FIG.

21 is the same as the control information memory 12. An output command 200 from the main computer is stored in a buffer memory 22, and the contents are decoded and necessary information is read out from the memory 21 via a read control circuit 23. First, the linear prediction coefficient {α1} is stored in the register 24 at the pitch period τ.

is stored in register 25, and the residual waveform {r1} is stored in register 2.
At step 7, the sound source intensity A is set in the register 28. In the case of voiced sounds, A=0 and the output of the random noise generator 30 is set to O by the amplitude controlled multiplier 29;
Only residual waveform 201 is applied as an input to synthesis circuit 31.

The number of times the residual waveform set in the shift register 27 is repeated and applied to the synthesis circuit 31 is determined by the pitch period τ. This control is determined by the relationship between and the analysis time T, and this control is based on the residual reading control circuit 26 and the contents of the register 25 τ. controlled by. In the case of unvoiced sounds, there is no residual waveform,
The output of the random noise generator 30 is controlled by the sound source intensity A and is applied as a waveform 202 to the synthesis circuit 31.

The output 203 of the synthesis circuit 31 is converted into an analog signal by the D/A converter 32, excess frequency components are removed by the low frequency converter 33, and the output 203 is amplified by the amplifier 34 and sent to the line to which it should respond. It is output as an audio waveform 300. Various control information is updated every 10 to 20 msec, and synthesis is performed continuously.

Details of the synthesis circuit 31 using linear prediction coefficients are shown in FIG.

A residual waveform 201 or intensity-controlled random noise 202 as a sound source is added to an adder 80 . On the other hand, the sample value of the past output waveform is 1 sample period △T seconds (1
In the case of 0kHz sampling, the linear prediction coefficients α1, α2 are added to the amplitude control circuits 51, 52, . . . 59, 5n through the delay circuits 41, 42, . ,...α. :71,72,...79
, 7n, are added together by adders 62, . To express this relationship mathematically, the explanation above has been given using the linear prediction coefficient as an example, but the equivalent partial autocorrelation coefficient can also be considered in the same way, with the only difference being the configuration of the analysis and synthesis circuit. be able to.

An example of the results of implementation according to the present invention is shown below.

Figure 5 shows an example of p-12 partial autocorrelation coefficient analysis and synthesis, where (a) is the original speech waveform and (b) is the residual waveform (
(Example of voiced sound), (c) is the result obtained by repeating the residual waveform for one cycle, and (d) is the synthesis result from (c). Comparing (a) and (d), there is almost no difference, which indicates that high quality audio is being reproduced.

In the above explanation, the pitch period is determined from the correlation of the residual waveform, but it is clear that it may be detected from the correlation of the original speech.

In this case, separation of voiced sounds and unvoiced sounds may be performed according to the method described above. According to the present invention, since a predetermined residual waveform stored in advance is used as a sound source in voiced sound synthesis, it is possible to synthesize speech of arbitrary content in a natural form with less distortion. .

[Brief explanation of drawings]

Fig. 1 is an example of a residual waveform as a prediction error in linear prediction coefficient analysis, Fig. 2 is a block diagram for calculating the residual waveform, Fig. 3 is a block diagram of a speech synthesis circuit according to an embodiment of the present invention, and Fig. 4 is a block diagram of a speech synthesis circuit according to an embodiment of the present invention. The figure is an explanatory diagram of the storage contents of control information for synthesis.
The figure is an example of the results of implementing the present invention.

Claims (1)

[Claims] 1. A speech synthesis device using linear prediction coefficients or partial autocorrelation coefficients, comprising means for storing a predetermined residual waveform for a pitch interval, and random noise generation means, and capable of performing voiced speech synthesis. A speech synthesis device characterized in that the residual waveform is used as a sound source in the case of unvoiced sound synthesis, and the output of the random noise generating means is used as a sound source in unvoiced sound synthesis. 2. In the above claim 1, the means for storing the residual waveform divides the waveform into a voiced sound section and an unvoiced sound section based on the correlation between the residual waveform or the source speech at regular time intervals, and stores the residual waveform at regular intervals. A speech synthesis device characterized in that a pitch period is detected in a voice section, and only the first pitch period section is stored from the residual waveform.