JPH06337691A - Sound rule synthesizer - Google Patents

Abstract

PURPOSE:To automatically revise an accent type of a relevant word in a dictionary only by correctly vocalizing the word speech synthesized by the eroneous accent type for a user. CONSTITUTION:This synthesizer is constituted so that a text becoming an object of reading out by speech synthesis is preserved in a storage part 1, and phoneme information used for the speech synthesis is preserved in a buffer 10, and when the user vocalizes the relevant word correctly in the case of erring the accent type of the word at the time of reading out the text, the sound is fetched by a sound input part 12, and is analyzed by an input sound analysis part 13 to provide the phoneme information, and a part similar to the phoneme information is detected from the phoneme information in the buffer 10 by a matching part 14, and the word suitable for the similar part is extracted from the text in the storage part 1 by a word extraction part 15, and accent information dealing with the extracted word in a Japanese dictionary 3 is revised to the accent type estimated by an accent type estimation part 16 based on a input sound by a dictionary revision part 17.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech rule synthesizing device for reading input text as a synthesized voice by speech rule synthesizing, and in particular, speech rule synthesizing suitable for correcting an accenting error during speech synthesizing. Regarding the device.

[0002]

2. Description of the Related Art For example, Japanese rule synthesis in this type of speech rule synthesizing apparatus is first performed on a language analysis dictionary (Japanese dictionary) for an input sentence (text) containing kanji and kana.
While performing linguistic processing such as morphological analysis, the reading and accent type of each morpheme is obtained from the above dictionary. Regarding the accent, the accent is determined for the entire sentence by using the accent movement rule in the case of a compound word or the like.

From the readings obtained as described above, for example, phonological parameters consisting of cepstrum parameters are generated, and from accent information, prosodic parameters consisting of pitch patterns are generated.

Then, a pulse having a period obtained from the prosodic parameter in the voiced section and a random noise in the unvoiced section are used as sound source signals, and filter coefficients obtained from the phonological parameter are constituted by an LMA (logarithmic amplitude approximation) filter or the like. It synthesizes the voice by giving it to the synthesizer filter.

[0005]

In the conventional speech rule synthesizing apparatus described above, when the accent type of a word is erroneous when reading a text by speech synthesis, the synthesis is temporarily suspended,
It is necessary to find the corresponding word from the entry words in the language analysis dictionary and change the accent type of that item. However, this method has a problem in that it takes a lot of time to change the accent type of one word and is extremely inefficient.

Therefore, according to the present invention, when the accent type of a word is erroneous at the time of reading a text by voice synthesis, the user can simply say the word with the correct accent type and the accent type of the corresponding word in the language analysis dictionary is automatically generated. A first object of the present invention is to provide a speech rule synthesizing device that can be changed dynamically.

Further, the present invention provides a speech rule synthesizing device which can automatically register a word uttered by a user in the linguistic analysis dictionary when the word is not present in the linguistic analysis dictionary. The second purpose.

[0008]

In order to solve the above problems, the present invention comprises a dictionary (language analysis dictionary) containing word accent information, and performs a language analysis on an input text using the dictionary. In the phonetic rule synthesizer that generates readings and accents according to the analysis result, generates phonological information from readings, prosody information from accents, and synthesizes arbitrary speech according to these informations.
A text storage means for storing the input text, a phonological information storage means for storing phonological information used for speech synthesis, a voice input means for inputting a voice, and an analysis of the input voice, An input speech analysis unit for obtaining phonological information, and a phonological information matching unit for detecting a part of the phonological information stored in the phonological information storage unit that is similar to the phonological information of the input speech found by the input speech analysis unit; , A word extraction means for extracting a word corresponding to the detected similar phonological information portion from the input text stored in the text storage means, and an accent type estimation means for estimating the accent type of the input speech. Dictionary changing means for changing the accent information in the dictionary for the word extracted by the word extracting means according to the accent type. Is shall.

[0009]

In the above structure, when the accent type of a word is incorrect due to a dictionary registration error when reading the input text by voice synthesis, when the user correctly utters the corresponding word, the voice is input by the voice input means. The phoneme information of the input voice is obtained by the input voice analysis means. The accent type of the input voice is estimated by the accent type estimating means.

On the other hand, the input text to be read aloud by speech synthesis is stored (saved) in the text storage means, and the phoneme information used for speech synthesis is stored (saved) in the phoneme information storage means.

The phoneme information matching means, when the phoneme information of the input voice is obtained by the input phonetic analysis means, detects a portion similar to the phoneme information from the phoneme information stored in the phoneme information storage means. When this similar phoneme information is detected, the word corresponding to the similar phoneme information part is
Extracted from the input text stored in the text storage means by the extraction means.

The dictionary changing means changes the accent information in the dictionary for the extracted word according to the accent type estimated by the accent type estimating means. If the extracted word is not registered in the dictionary,
The dictionary changing means registers the word and the estimated accent type in the dictionary.

As described above, according to the above configuration, when the accent type of a word is erroneous at the time of reading a text, the user correctly utters the word, and the corresponding word in the dictionary (language analysis dictionary) containing the accent information is read. The accent type of a word can be changed automatically, and if the corresponding word is not in the dictionary, the word and the accent type are automatically registered in the dictionary. All you have to do is to let the synthesizer read it and point out (speak) the word at the point where you read the accent by mistake with the correct voice. Even users without specialized knowledge can easily and efficiently change and register the accent dictionary. Can be done.

[0014]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing an embodiment of a Japanese speech rule synthesizing device to which the present invention is applied.

The Japanese speech rule synthesizing apparatus of FIG. 1 stores an input text storage unit 1 for sequentially storing input Kanji / Kana mixed sentences (text), linguistically analyzes the input sentence,
A language analysis unit 2 for generating reading and accent information, and a Japanese dictionary (language analysis dictionary) 3 used for language analysis of the language analysis unit 2 are provided. In this Japanese dictionary 3, in addition to grammatical information such as part-of-speech information, reading and accent type are registered.

The Japanese speech rule synthesizing device of FIG. 1 also includes a phoneme length determining unit 4 and a language analyzing unit 2 for determining the duration (phoneme length) of each phoneme based on the reading information from the language analyzing unit 2.
It is provided with a phoneme parameter generation unit 5 for generating a phoneme parameter consisting of a cepstrum parameter based on the phoneme length information determined by the phoneme length determination unit 4 and a phoneme unit memory 6. The voice unit memory 6 holds voice units corresponding to all phonemes.

The Japanese speech rule synthesizer shown in FIG. 1 also includes accent type information from the language analysis unit 2 and a phoneme length determination unit 4.
A prosody parameter generation unit 7 that generates a prosody parameter composed of a pitch pattern and voiced / unvoiced information using the phoneme length determined in step S6. A sound source signal is generated based on the prosody parameter generated by the prosody parameter generation unit 7. A synthesizing filter 8 for synthesizing voice using the phonological parameter generated by the phonological parameter generating unit 5 as a filter coefficient, a speaker 9 for outputting the voice synthesized by the synthesizing filter 8, and a phonological parameter buffer 10. The phoneme parameter generator 10 includes a phoneme parameter generator 5
The phoneme parameters generated in step 1 are sequentially saved.

The Japanese voice rule synthesizing device of FIG. 1 also includes a voice input unit 12 for taking in a voice uttered by a user via a microphone 11, a phonological parameter generated during voice synthesis by analyzing the input voice ( Synthetic phoneme parameter)
An input speech analysis unit 13 for converting into a phoneme parameter (input speech phoneme parameter) of the same kind of parameter (here, a cepstrum parameter), and a synthetic phoneme parameter stored in the phoneme parameter buffer 10 and the input phoneme parameter are matched. The matching unit 14 for detecting the most similar part in the phoneme parameter and the word extracting unit 15 are provided. The word extracting unit 15 extracts a word corresponding to the phoneme parameter portion detected by the matching unit 14 from the kanji / kana mixed sentence stored in the input text storage unit 1.

The Japanese speech rule synthesizing apparatus of FIG. 1 further includes an accent type estimating unit 1 for estimating the accent type of the input speech.
6, and the word extraction unit 15 registered in the Japanese dictionary 3
Accent type estimation unit 1 for accent type of words extracted in
Dictionary changing unit 17 for changing to the accent type detected in 6
Is equipped with. When the corresponding word is not registered in the Japanese dictionary 3, the dictionary changing unit 17 registers the word and the accent type detected by the accent type estimating unit 16 in the Japanese dictionary 3.

Next, the operation of the Japanese voice rule synthesizing device having the configuration of FIG. 1 will be described. First, it is assumed that a sentence (text) mixed with kanji and kana to be read is input to this device.

This input sentence (input text) is sequentially stored in the input text storage unit 1 and is also passed to the language analysis unit 2. The language analysis unit 2 divides the input sentence (input text) into morpheme units by performing well-known language analysis processing such as morpheme analysis with reference to the Japanese dictionary 3. Then, the language analysis unit 2 refers to the Japanese dictionary 3 and
The pronunciation and accent type are given to each morpheme, and the accent movement rule for compound words is used to determine the accent for the entire sentence.

Then, the phoneme length determining unit 4 calculates the duration, that is, the phoneme length, of each phoneme included in the reading given by the language analyzing unit 2. The phoneme parameter generation unit 5 reads a phoneme unit corresponding to each phoneme included in the reading given by the language analysis unit 2 from the phoneme unit memory 6 and matches the phoneme length calculated by the phoneme length determination unit 4. As described above, the phoneme parameters are generated by expanding and contracting the segments and connecting them, and then applying smoothing so as not to cause distortion at the connection parts. The phoneme parameters generated by the phoneme parameter generator 5 are sequentially stored in the phoneme parameter buffer 10.

On the other hand, the prosody parameter generation unit 7 is composed of pitch patterns and voiced / unvoiced information using the accent type information given by the language analysis unit 2 and the phoneme length information calculated by the phoneme length determination unit 4. Generate prosody parameters.

The phonological parameters generated by the phonological parameter generating section 5 and the prosody parameters generated by the prosody parameter generating section 7 are passed to the synthesis filter 8. The synthesis filter 8 is, for example, an LMA (logarithmic amplitude approximation) filter, and based on the pitch pattern and the voiced / unvoiced information, which are the prosody parameters passed from the prosody parameter generation unit 7, the impulses of the pitch period in the voiced section, White noise is generated in the unvoiced section, and this is used as a sound source (sound source signal), and at the same time, the speech is synthesized using the phonological parameter consisting of the cepstrum parameter passed from the phonological parameter generating unit 5 as a filter coefficient. The voice synthesized by the synthesis filter 8 is output from the speaker 9.

In this way, it is assumed that, for example, a sentence (text) "Tomorrow's weather will be raining" is read aloud by the synthesizer of FIG. In this sentence, "rain" is usually in the Tokyo dialect as shown in Figure 2 (b).
It is uttered as "Type 1" (Height type) as shown in. However, the accent type of "rain" in the Japanese dictionary 3 of the synthesizer of FIG. 1 is "0" (flat type) as shown in FIG. 3A due to a registration error. Assuming that
In the synthesizing device, when reading the above sentence with synthesized speech, "rain" is read as "type 0" as shown in FIG. 2 (a).

In such a case, in the synthesizer of FIG. 1, if the user hears the synthesized voice, determines that the accent is wrong, and correctly utters the word ("rain") at the wrong place, As described below, a registration error in the Japanese dictionary 3 can be automatically corrected.

First, the user hears the synthesized voice of the sentence "Tomorrow's weather will be rain." As a result, the user judges that the accent of "rain" is wrong, and the microphone connected to the voice input unit 12 is detected. Towards 11, the correct accent shown in FIG. 2 (b), that is, "Type 1" accent, "rain"
It is assumed that he uttered.

The voice "rain" uttered by this user is input to the voice input unit 12 via the microphone 11. The voice input unit 12 digitizes this input voice and transfers it to the input voice analysis unit 13 and the accent type estimation unit 16.

Then, the input voice analysis unit 13 performs a cepstrum analysis on the digitized input voice to obtain a cepstrum parameter, and passes the cepstrum parameter to the matching unit 14.

The matching unit 14 is an input voice analysis unit 13
When the cepstrum parameter (input speech phoneme parameter) of the input speech is received from the phoneme parameter buffer 1
Among the cepstrum parameters (synthesized phoneme parameters) stored in 0 (generated by the phoneme parameter generation unit 5 and used for voice synthesis), the received voice (input voice) “rain” of the user's voice is received. For example, DP (Dynamic Prog
ramming) Detect by matching method.

Then, the word extraction unit 15 determines which of the words in the kanji / kana mixed sentence (input text) stored in the input text storage unit 1 is the voice (input voice) of the user detected by the matching unit 14. ) The word "rain" is extracted by judging whether it corresponds to a similar part to "rain". Then, the word extracting unit 15 passes the extracted character code of the word “rain” to the dictionary changing unit 17.

On the other hand, the accent type estimation unit 16 performs pitch analysis on the digitized input voice "rain" from the voice input unit 12 to obtain the locus of the pitch frequency. Then, the accent type estimating unit 16 estimates the accent type of the input speech “rain” from the locus of the pitch frequency, and passes the accent type (here, “Type 1”) to the dictionary changing unit 17.

The dictionary changing unit 17 searches the Japanese dictionary 3 for the word "rain" passed from the word extracting unit 15, and when the word "rain" is registered as shown in FIG. 3 (a). Is the accent type item "0" for the word "rain",
The accent type "1 type" estimated by the accent type estimating unit 16 is changed as shown in FIG.

As described above, the word "rain" in the Japanese dictionary 3
As a result of changing the accent type for "0 type" from "1 type", the next time the synthesizer of FIG. 1 reads the input sentence, the word "rain" is the correct accent type "1 type". Will be synthesized.

If the word extracted by the word extracting unit 15 is not registered in the Japanese dictionary 3 (unknown word or compound word), the dictionary changing unit 17 determines the word and the accent type estimating unit. The accent type estimated by 16 is newly registered in the Japanese dictionary 3.

Although one embodiment of the present invention has been described above, the present invention is not limited to the above embodiment. For example, in the above embodiment, the cepstrum parameter is used as the synthesis parameter, but another parameter such as LPC (linear prediction coefficient) may be used.

Further, although the matching section 14 applies the DP matching method to the matching method of the phoneme information (phoneme parameters), the invention is not limited to this. Further, the present invention does not limit the use of the synthesis parameter as the phoneme information at the time of matching as in the above embodiment. Therefore, the input voice is once recognized by the input voice analysis unit 13 and converted into a phoneme sequence. After that, the word may be estimated by matching with the reading information output from the language analysis unit 2.

Furthermore, the present invention is not limited to speech synthesis of Japanese sentences, but can be applied to a speech rule synthesis device for speech synthesis (speech) of sentences other than Japanese such as English sentences. In short, the present invention can be variously modified and implemented without departing from the gist thereof.

[0039]

As described above, according to the present invention, the input text to be read aloud by voice synthesis and the phonological information used for voice synthesis are respectively stored, and the user utters them. The phoneme is analyzed to obtain its phoneme information, the accent type is estimated, and a part similar to the phoneme information is detected from the stored phoneme information. Extract the words you want to use from the saved input text,
The accent information in the dictionary (language analysis dictionary) for the extracted word is changed to the estimated accent type,
Alternatively, when the corresponding word is not in the dictionary, the word and the estimated accent type are registered in the dictionary, so that the following effects can be obtained.

(1) If the accent type of a word is incorrect due to an error in registering the dictionary when reading the input text (sentence) by voice synthesis, if the user correctly utters the corresponding word, the accent in the dictionary is incorrect. The accent type of the read word can be automatically changed to the correct accent type estimated from the word uttered by the user.

(2) When the word uttered by the user does not exist in the dictionary, the word and the accent type can be automatically registered in the dictionary. (3) From the above (1) and (2), the user simply causes the phonetic rule synthesizing device to read the necessary sentence, and points out (speaks) the word at the position where the accent is read aloud with the correct voice. Even if the user has no specialized knowledge, the accent dictionary can be changed / registered very easily, which is efficient.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of a Japanese speech rule synthesizing device to which the present invention is applied.

FIG. 2 is a diagram showing the accent of a word of a synthetic voice in which an error has occurred and the accent of a word uttered by a user.

FIG. 3 is a diagram showing examples of contents of the Japanese dictionary 3 corresponding to FIG. 2 before and after the change.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 ... Input text storage unit (text storage unit), 2 ... Language analysis unit, 3 ... Japanese dictionary (language analysis dictionary), 4 ... Phoneme length determination unit, 5 ... Phoneme parameter generation unit, 6 ... Speech unit memory, 7 ... Prosody parameter generator, 8 ... Synthesis filter,
9 ... Speaker, 10 ... Phoneme parameter buffer (phoneme information storage means), 11 ... Microphone, 12 ... Voice input part, 13 ... Input voice analysis part, 14 ... Matching part, 15
... word extracting unit, 16 ... accent type estimating unit, 17 ... dictionary changing unit.

Claims (2)

[Claims] 1. A dictionary including word accent information is provided, linguistic analysis is performed on the input text using the dictionary, and reading and accent are generated according to the analysis result, and phonological information is obtained from this reading. , In the speech rule synthesizing device for respectively generating prosody information from accents and synthesizing arbitrary voices according to these information, a text storage means for storing the input text, and phonological information used for the speech synthesis. Phoneme information storage means for storing, voice input means for inputting voice, input voice analysis means for analyzing voice input by the voice input means to obtain phoneme information of the input voice, and the phoneme information Of the phoneme information stored in the storage means, a part similar to the phoneme information of the input voice obtained by the input voice analysis means is detected. Phonological information matching means, word extracting means for extracting a word corresponding to the similar phonological information portion detected by the phonological information matching means from the input text stored in the text storage means, and input by the voice input means. Accent type estimating means for estimating the accent type of the speech that has been produced, and a dictionary for changing accent information in the dictionary for the word extracted by the word extracting means according to the accent type estimated by the accent type estimating means. A voice rule synthesizing device comprising: a changing unit. 2. The dictionary changing means, when the word extracted by the word extracting means is not registered in the dictionary, the dictionary and the accent type estimated by the accent type estimating means. The voice rule synthesizing device according to claim 1, wherein