JPS5993498A - Voice synthesizer - 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] The present invention relates to a speech synthesis device.

In a speech synthesizer such as the PARCOR method, a digital filter is used to realize vocal tract resonance characteristics.

If this digital filter has certain frequency characteristics,
Even if there is no input sound source signal, the output signal continues to be sent out by feeding back the error of the digital filter. That is, there is a drawback that oscillation occurs during the silent period, resulting in an extremely noisy sound. The conditions under which this oscillation occurs are determined by the filter constant and past input sound sources, and are therefore very difficult to predict.

An object of the present invention is to provide a speech synthesis device that prevents oscillation during silent periods through simple operation.

Other objects of the invention will become apparent from the following description and drawings.

Hereinafter, this invention will be explained in detail together with examples.

A block diagram of an embodiment of the invention is shown in the drawings. Each circuit block surrounded by a broken line in the figure is formed on a semiconductor substrate such as a silicon single crystal by a known semiconductor integrated circuit manufacturing technique.

Although the semiconductor integrated circuit device IC for speech synthesis is not particularly limited, the Hefei pulse train generator 1. The following circuits are added to the system which is composed of a sound source section consisting of a noise generator 2, a digital filter 4, a D/A converter 5, a parameter decoding circuit 3, etc.

The delay circuit 6 starts operating upon receiving the silence signal from the decoding circuit 3, and forms a switching control signal for the multiplexer 7 after a certain period of time delay as described later. This multiplexer 7 selectively transmits 1 to the feature parameter output from the decoding northern TiR3 and the feature parameter ki-0 fixedly formed internally to the digital filter 4.

That is, basically, multiplexer 7 is activated during the sound period.
from the t1 encoding circuit 3 through the ! characteristic parameter ki
is supplied to the digital filter 4, and during the silent period, the feature parameter set to kI=0 is transmitted to the digital filter through the multiplexer 7.

For example, a PARCOR type digital filter 4 as shown in FIG. 2 is a lattice type filter, and feedback due to the error occurs, causing oscillation. Therefore, as mentioned above, the feature parameter kl-00
When replaced as shown in FIG. 3, in the equivalent circuit of the digital filter 4, the feedback path is cut off as shown in FIG. 3, so no oscillation occurs.

In this embodiment, the delay circuit 6 does not immediately switch the multiplexer 7 even when there is a silent period, but switches the multiplexer 7 after a certain period of time. The reason for this will be explained next.

If the above feature parameter ki=o is replaced immediately after the sound source input disappears (the end of utterance), the digital filter 4
Because information remains in the output (the output of the digital filter 4 is not 0), abnormal sounds will occur in the output audio.Therefore, the above feature parameters should be changed! It is necessary to pay attention to the timing of replacing with =O. The attenuation characteristic of the filter output is the most Sogi (smallest in band 1I9i!B)
In the formant, it is determined.

As an example, the filter characteristics (female vowel/a
Let's consider a case where voice production ends with /). From the same figure, for the formant with the steepest characteristic, 1Zl=
0.9812, frequency f-10B4Hz, bandwidth B=
48.33Hz, damping constant α=151. 87 seconds
, the time constant τ=1/α=6.583m5 is found.

FIG. 5 shows the response when an impulse is applied to this filter as a sound source input. It can be seen from the figure that the above time constant exhibits a damping characteristic equal to -. In reality, the minimum bandwidth of a digital filter for speech synthesis is 30 to 100 nm in order to prevent abnormal amplitudes due to resonance and improve sound quality.
Frequently set to 00Hz or higher 8 For example, B-30Hz
Even so, the attenuation becomes 32.7 dB 40 m5 after the end of the utterance, and 49.1 dB after 601 IIs. Therefore, in this embodiment, the delay time formed by the delay circuit 6 is set to 40 to 60 m5. This makes it possible to prevent oscillation during silent periods without affecting the synthesized speech.

Although not particularly limited, the delay circuit uses a counter constituted by a flip-flop or the like that counts the frame period. Immediately upon arrival of the next utterance period, it is reset and the multiplexer 7 is switched to the decoding circuit 3 side.

In the block diagram of FIG. 1, the voice synthesis operation of the PARCOR method is such that an audio data signal such as an audio data ROM (read only memory) is inputted every frame period, and the parameter restoration circuit 3 processes, for example, 48
In the bit/frame case, the feature parameters are each decoded into 10 bit data. Furthermore, in a parameter interpolation circuit (not shown), quasi-linear interpolation is performed every 2.5 ms with the feature parameter value of the previous frame, and smoothing is performed for discrete changes in each feature parameter.

Then, the pitch information flP connects the sound source section to the amplitude information NA and P.
The ARCOR coefficient ki controls the digital filter 4. As a result, the original voice is restored and synthesized, and the D/A converter 5
and is uttered via an external speaker SP.

The invention is not limited to the above embodiments.

In other methods using digital filters, methods using line spectral coefficients, and methods using feature parameters r and f3 representing formant Q, these feature parameters can be mutually converted. Therefore, in the case of using line spectrum coefficients to prevent oscillation during the silent period, the constants are set at equal intervals, and in the case of using characteristic parameters r and e representing formant Q, r All you have to do is replace it with Bori Zero.

Furthermore, in the block diagram of FIG. 1, multiplexers, delay circuits, etc. for replacing characteristic parameters in order to prevent oscillation during the silent period may be provided outside the semiconductor integrated circuit device for speech synthesis. good.

[Brief explanation of drawings]

Fig. 1 is a circuit diagram showing an embodiment of the present invention, Fig. 2 is an equivalent circuit diagram of a PARCOR type digital filter, and Fig. 3 is an equivalent circuit diagram of a digital filter when the parameter kl-Q is set. , FIG. 4 is a filter characteristic diagram, and FIG. 5 is a characteristic diagram showing the response of the filter. 1...Pulse train generation circuit, 2...Noise generation circuit, 3...
Decoding circuit, 4...Digital filter, 5...D/A
Converter, 6... Delay circuit, 7... Multiplexer money - 655

Claims (1)

[Scope of Claims] 1. A speech synthesis device characterized in that oscillation in a digital filter is prevented by detecting a silent period, delaying it for a certain period of time, and manipulating a filter constant. 2. The filter constant has a partial autocorrelation coefficient as a characteristic parameter, and the constant ki is replaced with one zero during the silent period. The described speech synthesizer. 3. The filter constant according to claim 1, wherein the filter constant has a line spectrum coefficient as a characteristic parameter, and the constant is replaced by equal division during the silent period. Speech synthesizer. 4. The above filter constant uses formant information as a feature parameter, and the formant Q during the silent period
2. The speech synthesis device according to claim 1, wherein the constant r representing .times..times..times..times..times..times..times.