JPS6032720Y2 - speech synthesizer - Google Patents

Description

[Detailed explanation of the idea] The present invention relates to a speech synthesis device using phoneme synthesis.

In general, single sounds in Japanese and English, etc., are as shown in Figure 1.
The waveform has a consonant part C at the front and a vowel part V at the back.

Note that a voice consisting only of vowels will have a waveform of only the vowel part V.

The vowel part of a single note sampled in a predetermined frequency band is
A typical vowel waveform (with a predetermined pitch period)
It is known that the representative vowel sound waveform is repeated within the vowel part.

By quantizing such a representative vowel waveform at a predetermined sampling period and storing it in a memory, and repeatedly reading out the quantized data from this memory and converting it into an analog signal,
2. Description of the Related Art A phoneme synthesizer that synthesizes speech has been known in the past.

However, since the conventional phoneme synthesizer synthesizes based on the regular repetition of representative vowel sound waveforms, the synthesized speech feels unnatural.

The purpose of the present invention is to provide a voice threat device using phoneme synthesis to obtain synthesized speech close to natural speech.

According to this invention, a first repetitive waveform having a predetermined period appearing in a single waveform and at least one second repetitive waveform having a mutually different period near the said period are quantized at a predetermined time interval. a storage unit having a data group; a means for randomly reading out quantized data of the first and second repetitive waveforms from the storage unit; and creating an audio signal based on the read quantized data. A voice threat device is obtained, characterized in that it has means.

The present invention focuses on the fact that, for example, human voice has fluctuations in its pitch cycle (the above-mentioned repetitive waveform cycle) while it is transmitted from the sound source through the vocal tract and generated from the oral cavity.

In this invention, in order to give this fluctuation to the synthesized speech, we prepare second repetitive waveforms with different periods near the period of the representative first repetitive waveform (phoneme), and then combine them with quantum The converted digital data is stored in memory.

As a result, by reading out the prepared quantized data indicating the first and second repetitive waveforms and performing phoneme synthesis based on these, synthesized speech close to natural speech with pitch period fluctuations taken into account can be obtained. Can be done.

Furthermore, by reading out the first and second repetitive waveforms irregularly, a person who constantly listens to the synthesized sound can feel as if a human being is speaking.

An embodiment of the present invention will be described in detail below with reference to the drawings.

FIG. 2 is a model diagram of a first repetitive waveform having a predetermined pitch T1 as a representative phoneme included in one single vowel, and FIGS.
Different pitches T2. T3. FIG. 3 is a schematic diagram of a second repetitive waveform group having T.

FIG. 4 is a block diagram of the main parts of the speech synthesizer used in this embodiment.

Its configuration consists of the first and second
A memory (e.g., ROM) 1 in which a group of digital data obtained by quantizing a repetitive waveform at a sampling period of, for example, 10 KHz is stored in a predetermined area, and a read control unit 2 that controls reading data from the memory 1.
, a processing section 3 that performs synthesis processing based on the read data, and an output section 4 that converts the synthesized digital audio into an analog signal and outputs it to the speaker 5.

First and second repeated waveform quantized data arranged for each waveform in a predetermined address space in the memory 1 (for example, 8 bits per sampling point (of which 1 bit is code data indicating the sign of the amplitude) (normalized data)
at~an (period T1), b,~bn (period T2),
C1 to Cn (period T3) and d1 to dn (period T, ) (see FIGS. 2 and 3) are arbitrarily read out by the address signal output from the readout control section 2 and sent to the processing section 3.

Now, for example, a waveform with period T1 is read out twice in succession,
Next, the waveform of period T is read out three times in a row, and then the waveform of period T1 is read out twice, and the processing unit Data strings input to 3 are a1 to an, a1 to an1, and d1 to d.
n, d□~dn, b1~bn, b□~bn, a
□, an, *e*-e-.

As a result, the processing unit 3 performs necessary synthesis processing (for example, multiplication processing with envelope data of the vowel waveform, that is, envelope data connected to the maximum amplitude of each repeated waveform), and further processing of the volume (sound intensity). The vowel waveform shown in FIG. 5 is input to the loudspeaker 5 by performing multiplication processing with amplitude magnification data, etc. in order to capture the amplitude, and transmitting it to the D/A conversion circuit of the output section 4.

As is clear from looking at this audio analog waveform, the repetitive waveform contained within a single syllable is a series of waveforms with different pitches, so the voice has subtle frequency fluctuations that occur in the human vocal tract. The signal can be made of synthetic leather.

Therefore, the sound heard through the speaker sounds to the listener as if it were being produced by a human being.

Furthermore, as shown in FIG. 6, the readout control unit 2 includes a random address generation unit 6 and the type of vowel (for example, in Japanese, a, i, u, e, o; in English, a,
i, 11. e, o), a type specifying section 7 for specifying
It may also be configured to include a waveform specifying section 8 for specifying sampling data within a repetitive waveform.

In this case, in the memory 9, for example, in Japanese, the address space is divided into five tables 1, 2, and 3, and a plurality of repetitive waveforms with different pitches are provided in each table.

At the time of synthesis, the type of vowel to be synthesized is specified in the type specifying section 7 (upper bits of the address), and any one of the different pitch groups is specified depending on the contents of the random address section 6 (the middle bit of the address). bit) The waveform specifying section 8 controls the sampling data within the waveform to be read out sequentially.

If a counter that counts a number according to the number of samples is used as the waveform specifying section 8, and the contents of the random address section 6 are changed according to the end of the count of the counter, a repeating waveform of an arbitrary pitch can be successively created. It can be read out.

Furthermore, the random address section 6 changes the contents randomly according to the number of repeated waveforms within one vowel.
For example, if you use a random number generation program or a polynomial counter, the array of waveforms will change repeatedly each time you synthesize the skin, so even regular listeners will be able to feel subtle differences between the audible sounds each time they listen. You can listen to it as a natural sound without any resistance.

This has the effect of giving an extremely stable feeling to the listener's mental structure.

Note that since the timbre of a vowel is determined by the shape of the repeated waveform, the first and second repeated waveforms described above may be prepared by extracting approximate waveforms with different pitches from each vowel; The pitch may be varied within a predetermined range (that is, to the extent that subtle pitch fluctuations can be felt) based on one representative phoneme, and a waveform approximation may be performed in accordance with the pitch.

Furthermore, the pitch of the repetitive waveform of the voice to be synthesized may be changed by changing the access speed to the memory or the synthesis processing speed using a control signal such as a clock signal.

Note that when sampling a repetitive waveform at intervals of 100 μs, for example, it is better to set the sampling start point and sampling end point to match the end point and start point of the consecutive waveforms before and after them. This is desirable because it prevents skipping.

In this example, the amplitude may be set to zero at the start and end.

Furthermore, the data to be quantized in memory is PCM, ADP
There is no problem in adopting an optimal data storage method such as CM.

In addition, we have presented an example in which the pitch of a repetitive waveform is varied as a frequency fluctuation in the vocal tract. Optimally modified waveforms including these may be prepared in memory, or amplitude modification processing, waveform distortion processing, etc. may be performed at the time of synthesis.

Further, address control for the memory may be either hardware control or software control.

Furthermore, it is disclosed that this invention can be fully applied not only to single vowels but also to cases in which consonants accompanied by a noise signal in front of them are synthesized, or to synthesize words or sentences that are grouped together.

[Brief explanation of the drawing]

Figure 1 is a single sound (noise + vowel) waveform diagram, Figure 2 is a repetitive waveform diagram with a predetermined pitch representative of a certain vowel,
Figures 3a and by c are waveform diagrams with different pitches that approximate the waveforms in Figure 2, Figure 4 is a block diagram of the main part showing an embodiment of this invention, and Figure 5 is a synthesized vowel. FIG. 6 is a block diagram of a main part showing another embodiment. 1... Readout control unit, 2, 9... Memory, 3... Processing unit, 4... Output unit,
5... Speaker, 6... Random address section, 7... Type specification section, 8... Waveform specification section.

Claims (1)

[Scope of utility model registration request] sampling data of a first audio signal waveform having a predetermined pitch; and sampling data of at least one second audio signal waveform having a waveform shape similar to the first audio signal waveform and having different pitches. the means by which it occurs;
It is characterized by comprising means for successively reading sampling data of each of the first and second audio signal waveforms in an arbitrary order, and means for creating an analog audio signal based on the read sampling data. Speech synthesis device.