JPH0419800A - Voice synthesizing device - Google Patents

Abstract

PURPOSE:To reduce the circuit scale and to decrease the difference between the quality of the synthesized voice of a fixed word and the quality of the synthesized voice of a variable word by performing the synthesis of the fixed word and the synthesize of the variable word by a common vocoder. CONSTITUTION:This device is equipped with a storage means 18 stored with codes of output voices of fixed words which are generated corresponding to vector quantization or matrix quantization according to standard patterns. Further, the device is equipped with the vocoder 8 which decodes and composes the codes of variable words generated with phoneme information and rhythm information obtained by the text composition 6 of character strings of variable words from the correspondence table between the composition units of text composition and standard patterns and codes extracted from the storage means 18. Namely, the composition of the fixed words and the composition of the variable words are performed by the common vocoder 8. Consequently, the circuit scale is reduced and the difference between the quality of the synthesized voice of the fixed words and the quality of the synthesized voice of the variable words can be made small.

Description

[Detailed Description of the Invention] [Field of Industrial Application] The present invention provides a speech synthesis device that outputs a combination of synthesized speech of variable words that can be replaced in fixed sentences and synthesized speech of fixed words that cannot be replaced. It is related to.

[Prior Art] There are three types of speech synthesis: a recording editing method, a parameter editing method, and a text synthesis method.

This recording editing method and parameter editing method (hereinafter, these two methods are collectively referred to as the recording editing method) have a high quality synthesized sound, but the natural sound corresponding to the sound you want to output must be recorded (compressed) in advance. It is necessary to memorize the information (by doing so).

In the case of the text synthesis method, this work is necessary, and any voice can be output by simply inputting a string of characters, but the quality of the synthesized voice is low at the current technological level.

Here, we will first realize a speech synthesis device that synthesizes speech in which the content of the output speech is a fixed phrase and some words in the fixed phrase change in various ways, such as a person's name, using only a recording/editing method. Let's think about it. In this case, the words that change in various ways as mentioned above are called variable words, and the words obtained by removing the variable words from the fixed phrase are called fixed words.The simplest way to memorize the output speech is to record all the words corresponding to each variable word. This method involves memorizing fixed phrases, but there is a problem in that the storage capacity becomes large.

In order to solve this problem, the audio data of each fixed word and the audio data of each variable word are stored in advance in a fixed word audio data file 1 and a variable word audio data file 2, as shown in FIG. Based on this selection signal, the recording/editing synthesis unit 3 receives a selection signal for sequentially selecting fixed words and variable words, and converts the fixed word audio data and the variable word audio data into each audio data file 1. It is conceivable to perform synthesis by reading out the voice from the speaker 4 and output the synthesized voice from the speaker 4.

In this case, the naturalness of the prosody generally deteriorates, but the naturalness can be improved by combining the prosody information of the variable word with the prosody information of the fixed word.

However, the drawback of this method is that only the sounds of variable words that have been stored in advance can be output.

Therefore, as shown in FIG. 5, an output audio instruction signal consisting of a fixed word selection signal and a character string signal of a variable word to be output is input, and the fixed word//' variable word 1llt unit 5 outputs the output audio signal. Based on the selection signal, the recording/editing synthesis unit 3 reads the fixed word audio data from the fixed word audio data file 1 and synthesizes the fixed word into speech. A conceivable method is to input a character string signal corresponding to the text into the text synthesis section 6 to perform text synthesis, and add the voices synthesized by each using the demultiplexer 7 to output the voice of the fixed phrase from the speaker 4.

[Problems to be Solved by the Invention] In the case of the conventional example shown in FIG. There is a problem that the circuit size becomes large, and the difference between the quality of synthesized speech for fixed words and the quality of synthesized speech for variable words is too large.
There is a problem that the synthesized speech becomes unnatural.

The present invention has been made in view of the above-mentioned problems, and its purpose is to reduce the circuit scale, reduce the difference between the quality of synthesized speech for fixed words and the quality of synthesized speech for variable words, and achieve high performance. To provide a speech synthesis device which can perform high quality speech synthesis and is inexpensive to manufacture.

[Means for Solving the Problems] In order to achieve the above-mentioned object, the present invention according to claim 1 combines fixed words whose words do not change and variable words consisting of words that can be changed. A speech synthesis device that outputs speech of fixed phrases is equipped with a storage means for storing codes formed by vector quantization or matrix quantization of output speech of fixed words based on a standard pattern, and also includes a storage means for storing codes formed in accordance with vector quantization or matrix quantization of output speech of fixed words. A bogorder that decodes and synthesizes the code of the variable word generated from the correspondence table between units and standard patterns and the phonological information and prosody information obtained by text synthesis of the character string of the variable word and the code extracted from the storage means. It is characterized by the fact that it is equipped with

The invention as set forth in claim 2 is provided with a storage means that can store codes of variable words generated in advance, and extracts the codes of variable words from this storage means and provides them to the bogoda.

[Function] According to the invention as claimed in claim 1, the synthesis of fixed words and the synthesis of variable words can be performed by a common bogorder, and therefore the circuit scale can be reduced, and the quality of the synthesized speech of fixed words and the synthesis of variable words can be improved. The difference between the quality of the synthesized speech and the quality of the synthesized speech can be reduced, and as a result, high quality synthesized speech can be obtained and costs can be reduced.

The invention according to claim 2 eliminates the need for a text synthesis section in the speech synthesis device, making it possible to further reduce the circuit scale.
As a result, the speech synthesis device can be made smaller and lighter, and the cost of the − layer can be reduced.

[Example] The present invention will be explained below with reference to Examples.

FIG. 1 shows the overall configuration of this embodiment. In this embodiment, both fixed words and variable words are synthesized using a common matrix quantized LPG bogorder 8. In the method using the Bogoda 8, first, as shown in Fig. 2 <a), during learning, the time series of spectral parameters obtained by performing linear predictive analysis on the speech waveform with the linear predictive analysis means 9 is used to analyze temporal changes in the time series of spectral parameters. Based on the segmentation means 10, each segment is resampled by the resampling means 11, and the matrix quantization means 12
Create a codebook 13 using this matrix representation as a standard pattern.
At the time of encoding, the encoding means 14 performs encoding based on the spectrum indicating the time series of the standard pattern of the codebook 13, the segment length, and the pitch.

For decoding and synthesis, as shown in FIG. 2(b), the generated code is taken into the decoding means 15 and decoded, and then the linear predictive synthesis means 16 obtains a synthesized speech waveform. This matrix quantization LPC Bogoda method is a well-known method as shown in "Extremely low bit speech coding using r time-space spectral patterns" by Yoshihisa Shiraki and Masaaki Honda (1984). In this embodiment, the original speech for the synthesized speech of fixed words and the original speech for the synthesis unit stored in the text synthesis unit 6 for variable words are the speech of the same speaker, and first, the large amount of speech generated by this speaker is The code book 13 is created by extracting the standard pattern according to the above method from the speech, and then a file of synthesis units (CV syllables) to be used in text synthesis is generated from the same speech. Then, each composite unit is encoded using only the spectrum, that is, the time series of the standard pattern for each composite unit is determined, and a composite unit-standard pattern correspondence table 17 is created.

On the other hand, in the fixed word synthesis system, the output speech is encoded based on the standard pattern and the above method, and the generated code is stored in advance in a storage means as a fixed word speech data file 18 as appropriate.

The text synthesis unit 6 provided in the variable word synthesis system performs matrix quantization using the above method according to the phonological information and prosody information generated from the input character string signal by the sentence analysis unit 19 and the prosody information generation unit 20. A code is generated by an LPC bogorder code generation unit 21. Regarding the spectrum, the time series of the standard pattern is determined from the above-mentioned synthesis unit-standard pattern correspondence table 17 based on the sequence of synthesis units obtained from the phonological information, and regarding the pitch, the pitch pattern generated based on the prosody information is determined. Convert into a form that can be expressed by each segmentation. Regarding segment length,
The length of each segment is determined based on the expansion/contraction method of each synthesis unit based on the data regarding phoneme length in the prosody information. Normally, the constant part of each composite unit is expanded or contracted, so the length of the segment (standard pattern) corresponding to the constant part is determined and encoded.

In the embodiment, the output voice instruction signal input from the outside is formed by arranging a character string signal for specifying a variable word and a selection signal for selecting a fixed word in the order of occurrence, and from this output voice instruction signal, a character string signal for specifying a variable word and a selection signal for selecting a fixed word are arranged. The fixed word/variable word separator 22 separates the fixed word/variable word separator 22 into a column signal and a selection signal, and converts the audio data, that is, the code, of the fixed word specified by the selection signal distributed by the fixed word/variable word separator 22 into the fixed word audio. It is the audio data extraction unit 23 that extracts from the data file 18.

When performing speech synthesis of a predetermined fixed sentence, an output speech instruction signal in which a number specifying a predetermined fixed word and a character string specifying a variable word are arranged in the order in which they are output is used for speech synthesis. The fixed word/variable word separation section 22 divides the given signal into a number signal as a selection signal and a character string signal, and sends the signal to the audio data extraction section 23 or the text synthesis section 6. For fixed words, the code specified and extracted from the fixed word audio data file 1 by the selection signal is output, and for variable words, the code generated by the matrix quantization LPC bogorder code generation unit 21 is output. The signals are sent to the matrix quantized LPG bogorder 8 via the multiplexer 7, decoded and synthesized, and the synthesized speech is emitted from the speaker 4. The synthesis section 8b and codebook 8a of the matrix quantized LPG bogorder 8 correspond to the linear prediction synthesis means 16 and decoding means 15 shown in FIG. 2(b).

In this way, in the device of this embodiment, both fixed words and variable words can be synthesized in the synthesizing section 8b of the matrix quantization LPC bogorder 8 using the same method. The audio parameter file for the standard pattern and the audio parameter file for the synthesis unit used in the text synthesis unit 6 can be shared, making it possible to further reduce the scale. The difference between the quality and quality becomes smaller. Therefore, the only difference between fixed words and variable words in this embodiment is the incompleteness of the prosodic information generation by the text synthesis section 6 and the incompleteness of the synthesis unit creation method.

Note that the text synthesis section 6 is not built into the speech synthesis device, but is provided separately as a variable word speech data file creation device A as shown in FIG. 3(a). Word speech data, that is, codes, are created in advance and stored as a variable word speech data file 25 using an appropriate storage means, and this variable word speech data file 25 is subjected to speech synthesis as shown in FIG. 3(b). It may be provided in the device. In this case, only pre-registered voices can be output, but the common voice data extraction unit 23 extracts the voice data, that is, the code, specified by the selection signal from the fixed word voice data file 18 or the variable word voice data file 24. As a result, the demultiplexer 7 of the embodiment shown in FIG. 1 is no longer necessary, and the circuit configuration can be further simplified to reduce cost, size, and weight.

It is also possible for the user to easily register additional variable words.

Although matrix quantization is used in each of the above embodiments,
Of course, other encoding methods such as a waveform segment vocoder method may be used as long as the encoding method uses segmentation and vector quantization or matrix quantization.

[Effects of the Invention] The invention as claimed in claim 1 is provided with a storage means for storing a code formed by vector quantization or matrix quantization of fixed word output speech based on a standard pattern, and also for text synthesis. A Bogorder is provided for decoding and synthesizing the code of the variable word generated from the correspondence table between units and standard patterns and the phonological information and prosody information obtained by text synthesis of the character string of the variable word and the code extracted from the storage means. Because of this, the synthesis of fixed words and variable words can be performed using a common Bogorder, which reduces the circuit scale, and also reduces the difference in quality between the synthesized speech of fixed words and the quality of synthesized speech of variable words. can be made smaller, resulting in higher quality,
This has the effect of realizing a low-cost speech synthesis device.

The invention according to claim 2 is characterized in that a text synthesis section is provided in a speech synthesis device which is provided with a storage means for storing the codes of variable words generated in advance, and extracts the codes of the variable words from the storage means and provides them to Bogoda. Since this is not necessary, the circuit scale can be further reduced, and as a result, the speech synthesis device can be made smaller and lighter, and the cost can be reduced.

[Brief explanation of the drawing]

FIG. 1 is a circuit configuration diagram of an embodiment of the present invention, FIG. 2 (a) is a circuit configuration diagram for explaining the encoding of matrix quantization LPC Bogoda as above, and Section 2 (b) is a circuit configuration diagram for explaining matrix quantization of same as above. A circuit configuration diagram for explaining decoding of LPC hogoder, Fig. 3 (
a) is a circuit configuration diagram of a variable language audio data file creation device used in another embodiment of the present invention, FIG. 3(b) is a circuit configuration diagram of the same as the above, FIG. FIG. 5 is a circuit configuration diagram of another conventional example. 6 is a text synthesis section, 7 is a demultiplexer, 8 is a matrix quantization LPG bogorder, 18 is a fixed word data file, 1 is a sentence analysis section, 20 is a prosodic information generation section, 21
is a code generation unit for matrix quantized LPG bogorder, 22
23 is a post-fixed/variable word separation unit, and a voice data extraction unit.

Claims (2)

[Claims] (1) In a speech synthesis device that outputs the speech of a fixed sentence by combining fixed words whose words do not change and variable words consisting of words that can be changed, the output speech of fixed words is converted into vector quantum based on a standard pattern. quantization or matrix quantization, a correspondence table between synthesis units of text synthesis and standard patterns, and phonological information and prosody information obtained by text synthesis of character strings of variable words. 1. A speech synthesis device comprising a bogorder that decodes and synthesizes a code of a variable word generated from the above and a code extracted from the storage means. (2) The speech synthesis apparatus according to claim 1, further comprising a storage means for storing codes of variable words generated in advance, and extracting codes of variable words from the storage means and providing them to the bogoda.