JPH0634175B2 - Text-to-speech device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Outline of the Invention] The present invention identifies a word or the like while matching a written sentence with a word or the like in a dictionary, converts the sentence into a reading, and reads the sentence aloud in a speech synthesis unit. Regarding reading device.

Conventionally, in this type of text-to-speech device, a katakana or hiragana in a text cannot be matched with a word or the like in a dictionary, so that there is a case where word identification fails and correct text cannot be read aloud. In the sentence reading device of the present invention, when a katakana or hiragana character string that does not exist in the dictionary exists in the sentence, the analysis is temporarily stopped there and the sentence is searched backward to find the beginning of the word existing in the dictionary. Detecting, restarting the analysis from there, and converting the ignored katakana and hiragana unknown words into readings one by one, inputting them into the speech synthesizer, reading a sentence and reading a correct sentence. providing.

[Industrial application field]

The present invention relates to the configuration of a Japanese language processing device, and in particular, identifies each word or the like while matching each word or the like of a sentence composed of words such as words, clauses, phrases, etc. with a word or the like in a dictionary unit,
The present invention relates to the configuration of a text-to-speech device that reads a text by a voice synthesis unit.

[Prior art]

With the development of integration technology, the performance of computer systems has improved along with the high performance and low price of processors, memory devices, printers, disk devices, etc., and not only individual usage but also the diversification of systems via networks. I have advanced. In terms of the input / output device, the interface with humans such as kanji input / output or voice input / output has been improved with the progress of pattern recognition and voice recognition technology, and a dedicated processing device for performing Japanese language processing is also an OA device. It has been emphasized. In particular, a word processor has a function of processing a sentence structure and has been popularized as a device for automatically editing or storing a sentence or creating a sentence, and Japanese sentences can be easily stored in an external storage device of a computer system. In addition, recently, a so-called discrete word device that recognizes individual syllables and words by uttering human voices by dividing them into units such as syllables and words has come into practical use. You can enter into the system. When checking Japanese sentences input using such a word processor or voice input device for correct input, or simply listening to newspaper articles already stored in the system, the machine was stored. The importance of text-to-speech devices that read texts has increased. This text-to-speech device requires a function of identifying a word of a text or a word such as a phrase or a phrase while matching with a word or the like of a dictionary and reading each text through a voice synthesizer. However, as in Japanese newspaper articles, there are a lot of sentences in which proper nouns or new words of katakana character strings, katakana writing for emphasis, or hiragana writing instead of kanji are used. When analyzing such a sentence with a reading device, it is possible to identify a word or the like as a character string (that is, basically, a word, but a phrase or phrase may be generically called a word). Very often it fails because the word does not exist in the dictionary. Conventionally, this type of text-to-speech device has a drawback in that if the identification of a word or the like fails, the correct text cannot be read.

[Problems to be solved by the invention]

The present invention eliminates the drawbacks of the conventional text-to-speech device described above, and when a word or the like in the dictionary unit and an input word or the like cannot be matched, that is, an unknown word part is ignored once and a part that can be identified is analyzed. However, the present invention provides a reading device that can read the entire sentence correctly by converting the ignored unknown words into reading one by one.

[Means for solving problems]

In order to achieve the above object, the present invention provides a sentence file section for storing a sentence, a sentence input / output section for controlling input or output of a sentence from the sentence file section, and the sentence input / output from the sentence file section. The sentence read through the unit is divided into units of small character strings, and the sentence analysis unit that matches and identifies the contents of the dictionary unit for each unit, and the reading output from the sentence analysis unit is converted into speech. A voice synthesis unit, an unknown word detection unit that is connected to the sentence analysis unit and finds an unknown word that did not exist in the dictionary unit in the sentence, and a position detection unit that detects the start and end positions of the unknown word, It consists of an unknown word reading conversion unit that creates the reading of the unknown word from the notation, and if an unknown word that does not exist in the dictionary exists in the sentence, pause the sentence analysis at the beginning of the unknown word and move the sentence backward. Service Blood was to resume the sentence analysis reading of the unknown word from the portion that comes to the next unknown word is characterized in that it created from the notation.

[Action]

The present invention, when katakana that does not exist in the dictionary, or hiragana character string exists in the sentence, temporarily stops the identification analysis there, searches the sentence backward, and analyzes from the word etc. that can be recognized as the beginning of a newly identifiable word. Resume, the katakana and hiragana in which the analysis is ignored are converted into reading one by one and transferred to the voice synthesizing unit so that the sentence is read aloud.

〔Example〕

Next, an embodiment of the text-to-speech device of the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram showing the structure of the text-to-speech device of the present invention.

The main control unit 1 is a unit that, like a general-purpose computer, sequentially retrieves a program stored in advance in a storage device and performs sequential control according to the program, and includes a central processing unit, a main storage device, an input / output control device, and the like. It is the part that is. The display device 10 and the keyboard input unit 11 are parts that are connected to the input / output control device in the main control unit 1 and exchange information between the human being who is the user and the text-to-speech device. The document file 2 is a file in which a document to be read out is stored, and is generally stored in an external storage device such as a floppy disk device or a disk device. Each document is read from the text file 2 by the document input / output unit 3. The text input / output unit 3 is a unit that controls the transfer of each document from the document file 2 to the text analysis unit 4 under the control of the main control unit 1. Each sentence is formed as a notation of a character string, and the character string is basically a word, but may be a clause or a phrase, and will be referred to as a “word or the like” hereinafter. Each sentence read from the document file 2 by the sentence input / output unit 3 is separated by the sentence analysis unit 4 by a symbol such as “,”, “.” Identification is performed. The dictionary unit 5 is, for example, a kanji-kana conversion dictionary that inputs and outputs kanji and converts the kanji into kana. The sentence analysis unit 4 can immediately read and convert a word such as a Chinese character stored in the dictionary unit 5 in each sentence output from the sentence input / output unit 3, and the rule synthesis unit 6 The converted reading is given. The rule synthesizing unit 6 is a unit that processes the information content corresponding to the reading so that it can be output from the voice output unit 7 in the form of voice. The voice synthesizing unit composed of the rule synthesizing unit 6 and the voice output unit 7 is a unit for preliminarily analyzing and storing data for voice output and reading a character string such as a word by voice corresponding to reading. .

Various analysis techniques such as waveform coding and linear predictive analysis have already been used for data analysis for speech synthesis, and processing of a waveform processing system that performs coding processing on the speech waveform itself may be used. It is also possible to use the processing of the generation source processing system that uses the control signal as data based on the generation model. The former retains the characteristics of long sound coupling between phonemes and the characteristics of accents or intonations that are included in natural speech such as words and phrases, so the sound quality is good, but speech output other than the combination of words and phrases that has been analyzed in advance. Has the drawback that it cannot output. On the other hand, according to the latter method, a word or a sentence to be output is input in the form of a phoneme or a syllable symbol sequence, so that the sentence of an arbitrary sentence can be synthesized. However, there is a problem that various characteristics of voice must be added artificially. The rule synthesizing unit 6 is a method of generating source processing by phonemes or syllables in particular, and follows a format in which phoneme or syllable-based speech analysis data is used as a control signal of a speech generation model. Therefore, the rule synthesizing unit 6 has no limitation on the number of vocabulary words of the synthesized speech, and regularizes the modification of the acoustic characteristics according to the long sound combination and the prosody characteristic found in natural speech. The modification of this acoustic property is regularized, and the control signal is modified based on this, so that the sound quality of the synthesized sound is enhanced to be very similar to natural speech. In the rule synthesizing unit 6, if only a phoneme, a syllable symbol, or a series of characters is input as a reading from the sentence analyzing unit 4, an arbitrary synthesized sound corresponding to the phoneme can be output from the voice output unit 7. When each sentence of the document file 2 is synthesized using the rule synthesizing unit 6 under the analysis of the sentence analyzing unit 4, it is ideal that a human reading process can be performed. Boundaries such as words must be detected and divided from character strings, and phoneme symbol strings such as detected words must be converted, and finally word accents or sentence intonations must be added. Further, the rule synthesizing unit 6 generally has a smaller storage capacity than that of the waveform processing type, but the circuit configuration becomes complicated. Therefore, the complexity of the circuit configuration can be reduced by digitally executing it with a one-chip LSI. is necessary. Furthermore, it is necessary to generate the control signal necessary for the operation of the voice synthesizer according to the predetermined rule with the voice data stored in advance. As LSI for that purpose, formant type and PARCOR type 2
There are types. For example, a PARCOR-based voice synthesis LSI synthesizes voice by connecting about 10 stages of lattice type digital filters each composed of two dividers and two adders in cascade. In the rule synthesizing unit 6 configured by using such an LSI, an input character string is converted into a phoneme sequence according to a rule, and is divided according to a so-called CV-V syllable of consonant + vowel + vowel system, for example.
With regard to the prosodic characteristics, the fundamental frequency, amplitude, phoneme duration, etc. for the synthesized speech are determined using the stored data of the acoustic data stored in advance. It is known that the sound quality of synthetic speech according to such rules is 90% or more clear for 100 Japanese syllables.

Each sentence read from the document file 2 via the sentence input / output unit 3 is divided into word units or phrase units and matched with the words in the dictionary unit 5 to be identified. In that case, unknown words such as hiragana or katakana that do not exist in the dictionary unit 5 may exist in the sentence. Especially in newspaper articles, there are many sentences that use hiragana instead of proper nouns of katakana character strings, new words or katakana writing for emphasis or kanji, and analysis of such sentences makes it quite high in word identification. The identification may fail. Therefore, the correct sentence cannot be read aloud as it is. Therefore, in the present invention, the unknown word detection unit 40, the restart position detection unit 41, and the unknown word reading conversion unit 42 are connected to the sentence analysis unit 4 and these circuits are operated to operate the words in the dictionary unit 5 and the sentences. If the word or the like is not matched, the unknown word is ignored and first analyzed, and the reading of that part is supplemented from the notation to enable correct reading. That is, if an unknown word such as hiragana or katakana that does not exist in the dictionary unit 5 exists in the sentence, the word cannot be identified and the unknown word detection unit 4
At 0, the position P ₁ of the unknown word is determined. Then, the sentence is searched backward as it is, and the resumption position detection unit 41 finds the position P ₂ of the notation of kanji and katakana that is known as the start of a new word. Then, first, the analysis up to the position P ₁ of the unknown word is performed, and then the analysis is performed again from P ₂ . Then, the unknown word reading conversion unit 42 finds the reading of the unknown word portion from the position P _{1 of the} unknown word to the position before the word position P ₂ of the word existing in the dictionary unit 5 and converts the reading into the reading, and the rule synthesis unit 6 The text is transmitted, synthesized by voice, and read by the voice output unit 7.

For example, in the case of reading out a sentence "Putting up there" as one sentence by the sentence reading device of the present invention, the analysis is started from the first character string "There". In this case, the hiragana word "There" is the dictionary part 5.
Exist and the identification of this term shall be carried out. However, it is assumed that the next katakana "Puturi" does not exist as a word in the dictionary. Therefore, since this word cannot be identified, the position P ₁ of the unknown word, that is, the position of the character “p” is stored as a pointer in the unknown word detection unit 40 and the sentence is searched backward as it is. Then, the restart position detecting unit 41 finds the position P _{2 of the} kanji that is known as the start of a new word, that is, the position P _{2 of} the word “interruption”, that is, the position of the character “middle”. Therefore, after performing the analysis up to P ₁ , analysis is not performed for the unknown word “Puturito”, and analysis is restarted for the next “interrupted” portion. After that, the unknown word portion "Putto" from P ₁ to P ₂ is read one by one in katakana, and the unknown word reading conversion unit 42 adds the reading.
In other words, the reading of "Puttori" is written in katakana and becomes "Puttorito". According to the above analysis results, all the sentences are converted into katakana, converted into katakana called "socodepturitochudanshitairu" with katakana, given to the rule synthesizing section 6 as reading, synthesized with speech, and outputted from the speech output section 7 as speech. It

Next, the operation of the text-to-speech apparatus of the present invention shown in FIG. 1 will be described with reference to the flowchart shown in FIG. According to the flowchart of FIG. 2, each sentence is input to the sentence analysis unit 4, and each word or the like of the sentence is identified by being matched with the word or the like of the dictionary unit 5,
The word lattice is created. Then, it is determined whether or not an unknown word of katakana or hiragana exists in each sentence. If there is no unknown word, WLIST that analyzes the sentence from SP, which is the pointer at the beginning of the sentence, to EP, which is the pointer at the end of the sentence, and stores the analysis result.
The analysis result is added to and transferred to the rule synthesis unit 6. However, if there is an unknown word of katakana or hiragana, the unknown word detection unit 40 finds the pointer P ₁ of the _first character of the unknown word. Then search the sentence backwards. As a result of the search, the pointer P ₂ of the first character of the word existing in the dictionary is found again. That is, the restart position detecting unit 41 finds a position in the sentence at which it can be recognized as the start of a new word after the unknown word. Then SP, which is the pointer at the beginning of the sentence
To P ₁ -1 immediately before P ₁ which is the pointer of the character at the beginning of the unknown word are analyzed and added to WLIST which is an area for storing the analysis result. Next, P ₁ which is the _first character of the unknown word to P ₂ which is the pointer of the end of the unknown word
-Make readings up to -1 from each letter,
Add to T. That is, according to the unknown word reading conversion unit 42, the reading of the unknown portion is added to WLIST. Then, P ₂ is set as SP as the beginning of the sentence, and the analysis is restarted from P ₂ .
After that, it checks again whether there is an unknown word. According to such repetition, analysis or reading of an unknown word can be added to WLIST until the end of one sentence, and all the readings of the sentence are transferred to the rule synthesizing unit 6 and converted into speech according to speech synthesis, and reading of the sentence is performed. It is possible to execute.

〔The invention's effect〕

Thus, the sentence raising device of the present invention ignores sentence analysis for unknown words in which the words in the sentence cannot be matched with the words in the dictionary and analyzes the other parts, and the unknown words With respect to, there is an effect that correct reading can be performed by supplementing the reading from the notation, and can be synthesized with respect to an arbitrary sentence voice.

[Brief description of drawings]

FIG. 1 is a block diagram showing the structure of the text-to-speech device of the present invention, and FIG. 2 is a flowchart of its operation. 1 ... Main control unit, 2 ... Document file, 3 ... Sentence input / output unit, 4 ... Sentence analysis unit, 5 ... Dictionary unit, 6 ... Rule synthesis unit, 7 ... Speech output unit, 10 ... ... display section, 11 ... input section, 40 ... unknown word detection section, 41 ... restart position detection section, 42 ... unknown word reading conversion section.

Claims (1)

[Claims] 1. A sentence file unit for storing a sentence, a sentence input / output unit for controlling input or output of a sentence from the sentence file unit, and a read from the sentence file unit via the sentence input / output unit. The sentence is divided into units of small character strings and matched with the contents of the dictionary unit for each unit, a sentence analysis unit for identifying, and a voice synthesis unit for converting the reading output from the sentence analysis unit into speech, An unknown word detection unit that is connected to the sentence analysis unit to find an unknown word that did not exist in the dictionary unit in the sentence, a position detection unit that detects the start and end positions of the unknown word, and reading of the unknown word And an unknown word reading conversion means for creating from the notation, when an unknown word that does not exist in the dictionary part is present in the sentence, the sentence analysis is temporarily stopped at the beginning of the unknown word and the sentence is searched backward to unknown. Of the word A text-to-speech device, wherein text analysis is restarted from the next part and the reading of the unknown word is created from the notation.