JPS61269200A - Voice sythesization filter - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

[Detailed Description of the Invention] [Field of Industrial Application] The present invention relates to a speech synthesis filter, and in particular, the present invention relates to a speech synthesis filter that uses a linear prediction method to extract spectra such as prediction coefficients or perfour coefficients representing the state of the vocal tract of a speaker from a speech signal. Parameters, amplitude information representing the size of the driving source of the vocal tract, voiced/unvoiced information representing the type of driving source, pitch information representing the vocal cord vibration frequency of voiced sounds, etc. are analyzed and extracted by the analysis unit, quantized, and then sent to the synthesis unit. Send it.

The present invention relates to a speech synthesis filter used when a synthesis unit reproduces original speech by synthesis based on the information.

[Conventional technology]

Conventionally, in this type of speech synthesis filter, when voicing occurs, an impulse driven at a pitch period and whose magnitude is proportional to amplitude information is input as a drive signal.

When there is no voice, white noise is used as the drive input signal. And 2
For example, the prediction coefficients etc. (spectral parameters) sent from the analysis section are α1. α2. ......α, the drive input signal is Sn, the output signal one time clock ago is Dn, the output signal two time clocks ago is D, and similarly the output signal p time clocks ago is n-2°. D, the output Dn follows the linear prediction method.

-p Dn=Sn+αI Di-1+α2Dn-2+...
...+αpDn-p. Here, one time clock is a sampling period (for example, 0.125 m5) when audio is analyzed in the analysis section. In other words, the output signal of the speech synthesis filter is applied to a circuit in which delay circuits each having a delay time of one time clock are connected in cascade.

It is configured to add a signal obtained by multiplying the output of the delay circuit of each stage by the spectrum parameter to the drive signal and output the resultant signal. Percoll coefficients are sometimes used instead of prediction coefficients for spectral parameters.

[Problem that the invention seeks to solve]

However, since the spectral parameters sent to the speech synthesis section are usually quantized, even if the system of spectral parameters is stable during the second analysis.

The stability of the spectral parameter system after dequantization in the synthesis section is not guaranteed, and it may become unstable. In such a case, if the above-described synthesis is performed using a conventional speech synthesis filter circuit, the resonance characteristics of the filter may become abnormally sharp, making it difficult to accurately reproduce the original speech. Furthermore, it has the disadvantage of causing a significant oscillation phenomenon.

The purpose of the present invention is to perform stable speech synthesis without causing abnormal resonance etc. even in the case where there is an error in the spectrum information as described above, and to achieve a sound waveform that is closer to the original speech waveform than conventional methods. An object of the present invention is to provide a speech synthesis filter capable of outputting speech having amplitude characteristics.

[Means for solving problems]

In the present invention, an impulse with a pitch period is inputted as a driving input signal when voiced, and white noise is inputted when unvoiced as a driving input signal, and a signal obtained by multiplying a signal obtained by delaying an output signal by a spectral parameter such as a prediction coefficient or a Percoll coefficient is inputted to the driving input signal. A speech synthesis filter that synthesizes and outputs speech by adding and outputting a signal includes a loss adding circuit whose loss is controlled by the predicted gain of the spectral parameter, and when the predicted gain becomes large, synthesis is performed via the loss adding circuit. It is characterized by being made to do.

〔Example〕

Next, two embodiments of the present invention will be described in detail with reference to two drawings.

First, two principles of the present invention will be explained. The present invention takes into account that the resonance characteristics of two synthesis filters become stable due to quantization of the spectral parameters, which are filter coefficients, and controls the resonance characteristics of the synthesis filter by determining the predicted gain of the synthesis filter in advance. It was designed to do so. and.

Speech synthesis is performed using the following equation using the loss parameter al. That is, the output signal Dn is determined by the following equation.

Dn=Sn+aα, Dn-1+a2α2Dn-2+...
・,,+B PαpDn-p The above loss parameter a is 0
Must be a variable with ≦a≦1.

Controlled by the prediction gain of the synthesis filter. and,
When the prediction gain is low, the same synthesis as before is performed with a'=1, but when the prediction gain is high, a' is set according to the prediction gain.
By setting 1 or less to increase the loss of the synthesis filter, the resonance characteristics of the filter are prevented from becoming sharper. What value should be set for the loss parameter a depending on the predicted gain?

For example, an appropriate function can be used, such as setting a = 0.9 for a high prediction gain and approaching 1 as the prediction gain decreases, but for details, it is experimentally possible to synthesize a signal close to the original speech. Just set it so that the output can be obtained.

FIG. 1 is a block diagram showing an embodiment of the present invention.

A drive signal Sn is applied to the input terminal 1, and the output signal of the adder circuit 2 is sent to the delay circuit 4. Coefficient addition circuit 5 and loss addition circuit 6
It is again applied to the adder circuit 2 via the . That is, the output signal is delayed for a certain period of time and added to the drive signal via a coefficient addition circuit and a loss addition circuit to obtain a composite output. The coefficient addition circuit 5 is a circuit that multiplies the output signal of the delay circuit 4 by the spectrum parameter (given as a digital value, for example, from the analysis section). Further, the loss addition circuit 6 is a circuit that multiplies the aforementioned loss parameter ai controlled by the loss control circuit 7. The prediction gain calculation circuit 8 calculates a prediction gain based on the spectrum parameters input to the input terminal 9.

The loss control circuit 7 determines the loss parameter ai according to the predicted gain and controls the loss adding circuit 6. With the above configuration, when the predicted gain is high, the loss of the loss addition circuit 6 becomes large, so the signal added to the drive signal becomes small, and the loss of the entire filter increases, so abnormal resonance can be prevented. is possible.

FIG. 2 is a diagram showing an example of a specific circuit configuration of the embodiment shown in FIG. 1, and shows a case where p delay circuits are connected in cascade. That is, the output signal of the adder circuit 2 is sent to the delay circuit 43.
Square multiplier 63. Delay circuit 42° multiplier 62.・・・・・・
, delay circuit 41. and a cascade connection circuit of multiplier 61 (
p stage). Then, the output signal of the square multiplier 63 is multiplied by the spectrum parameter α1 input from the input terminal 93 by the multiplier 53, and the resultant signal is applied to the addition circuit 2. The output signal of multiplier 62 is also output to terminal 9 by multiplier 52.
2 is multiplied by the spectral parameter α2 inputted from 2 and is applied to the adder circuit 2. Similarly, the final stage (pth) I
The output signal of the multiplier 61 is multiplied by the spectrum parameter α by the multiplier 51, and the resultant signal is applied to the addition circuit 2.

The adder circuit 2 adds these signals to the drive signal input from the input terminal 1 and outputs the resultant signal, and outputs a synthesized speech output from the output terminal 3. Multipliers 63 , 62 .

. . , 61 is a circuit that is controlled by the loss control circuit 7 and multiplies the loss parameter a. That is, it constitutes a loss adding circuit 6 (FIG. 1). The loss control circuit 7 determines a loss parameter a according to the output of the predicted gain calculation circuit 8, and controls each multiplier 63, 62 . . . . controls the multiplication value of 61. Here, input terminals 93, 92 . ...91 is used. The loss parameter a is.

When the prediction gain is low, it is set to 1, and as the prediction gain increases, it is set to a value smaller than 1. Details may be determined experimentally as appropriate. The number of stages p of the delay circuit is equal to the order of the spectrum parameter α.

With the above configuration, when the predicted gain is high, a signal with a large loss is added to the drive signal.

Abnormal resonance can be prevented. Also, the loss parameter a
By appropriately setting the value of , it is possible to output synthesized speech that is close to the original speech signal.

FIG. 3 shows each delay circuit 44 of a conventional Percoll lattice type speech synthesis filter. . . , 45 are provided with multipliers 64 . ..., 65 is inserted. In this case as well, the multiplication value a of the multipliers 64 and 65 is controlled by the loss control circuit 7. The input terminals 94, . . . , 95 are supplied with a Percoll coefficient of 1. ..., kp is input. In this case as well, the input terminal 94. ....

1 to the Percoll coefficient input from 95. . . . The multiplication value is controlled by the loss parameter a determined by the loss control circuit 7 in accordance with the predicted gain calculated by the predicted gain calculation circuit 8, and the loss is given in the same manner as in the previous embodiment. The same effects as described above are achieved in terms of prevention of reeds and abnormal resonance.

〔Effect of the invention〕

As described above, in the present invention, when an output signal is added to a drive input signal via a delay circuit.

Since the loss is added in accordance with the predicted gain after quantization, it is possible to prevent abnormal resonance from occurring and perform stable operation. Furthermore, by appropriately setting the loss parameters, it is possible to output synthesized speech with good quality that is close to the original speech waveform.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing one embodiment of the present invention. FIG. 2 is a block diagram showing an embodiment in which p delay circuits are connected in cascade, and FIG. 3 is a block diagram showing an embodiment in which the present invention is applied to a Percoll grid filter. In the figure, 2.21 to 24...addition circuit, 4°41
~45...Delay circuit, 5...Coefficient addition circuit, 6...
- Loss addition circuit, 7... Loss control circuit, 51-57° 61-65... Multiplier, 8... Predicted gain calculation circuit.

Claims (1)

[Claims] 1. When voiced, an impulse with a pitch period is input, and when unvoiced, white noise is input as a driving input signal, and a signal obtained by multiplying a signal obtained by delaying the output signal by a spectral parameter such as a prediction coefficient or a Percoll coefficient is used as the driving input signal. 1. A speech synthesis filter that synthesizes and outputs speech by additionally outputting it, comprising a loss adding circuit whose loss is controlled by a predicted gain of the spectral parameter.