JPH03237498A - Device for reading sentence aloud - Google Patents

Abstract

PURPOSE:To reduce the capacity of a memory to be previously prepared by sequentially executing dictionary reference and word identification processing and deleting unnecessary word information obtained from a dictionary in each reference. CONSTITUTION:A word identification (ID) part 47 connected to a word dictionary 45 is connected to a receiving part 41 and a phoneme/rhythm analytical part 49 is connected to the part 47. A synthesized parameter forming part 55 connected to a sound file 53, a synthesized sound forming part 57 and a speaker 59 are connected to the part 49. Thus, dictionary reference is applied to all candidate words in a character string is each prescribed rang of an inputted sentence, a word string is determined by using an evaluation value indicating which word positioned before each word is optimumly selected and the prescribed range is shifted from the head of a sentence to repeat this processing so the dictionary reference and word identification are sequentially executed to delete unnecessary information. Thereby, an area necessary for storing information obtained from the dictionary can be minimized and the necessary memory capacity can be reduced.

Description

[Detailed Description of the Invention] [Summary] To reduce the memory capacity required in word identification processing using dynamic programming, regarding a text reading device that reads Japanese sentences containing kanji and kana in a natural voice. With the aim of We look up the words in a dictionary, determine the word string using the evaluation value that indicates which of the words immediately before each word is best to select, and delete candidate words that are not selected each time. The apparatus is configured to include word identification means for identifying words by sequentially moving a predetermined range from the beginning of a sentence.

[Industrial Application Field] The present invention relates to a text reading device that reads out Japanese sentences containing kanji and kana (hereinafter referred to as "r text") in a natural voice.

A text reading device can be used as an information transmission means when visual information transmission is not possible. Examples include reading out data over a telephone line, terminals for the blind, proofreading text input into a word processor, and voice guidance.

[Conventional technology]

To read aloud a sentence containing kanji and kana, it is necessary to correctly identify the words in the sentence (hereinafter referred to as ``identifying''), automatically add the correct sound of the word (gfi mi) and accent, and add a natural accent to the sentence. Language analysis technology is required to automatically add pauses, intonation, etc.

Therefore, the structure of a text-to-speech device that reads aloud sentences containing kanji and kana that have been manually inputted from a human-powered device includes a language processing section that performs this language analysis, and a corresponding voice processing section based on the melody information and prosody information obtained here. It can be broadly divided into a speech synthesis section and a speech synthesis section.

In other words, the language processing unit breaks down the input kanji-kana mixed sentence into predetermined units (words, word strings), refers to the word dictionary, and performs word identification processing that adds pronunciation, accent information, and grammatical information to each of them. After doing so, use compound words, clauses,
Processing is performed to reconstruct the relationships between clauses (accent phrases, exhalation paragraphs) into higher-level linguistic units, modify phonological information as compound words or sentences, and determine prosodic information such as accent and intonation. The configuration is such that the information is passed to the speech synthesis section.

By the way, as a method for the word identification processing described above, the longest match method is a common method. The longest-length method is a method that sequentially selects the word with the longest word length included in the dictionary from the beginning of the sentence. However, this method does not include "big business".
"It's important,""karma," and "former prime minister" is interpreted as "head of state."
There was a problem in that it was interpreted as "phase".

As a method to solve this problem, there is a matching method using dynamic programming. A matching method using dynamic programming calculates an evaluation value for all candidate words in a string of human text to determine which of the words immediately preceding each word is best to select. This method is determined by using This method performs a dictionary lookup for all candidate words, examines all relationships between adjacent words, and then determines words in reverse order starting from the end of the sentence based on the evaluation value set between each word and adjacent words. This method produces the final identification result.

[Problem to be solved by the invention]

However, in the matching method using dynamic programming described above,
Since all candidate words had to be looked up in the dictionary and all relationships between adjacent words had to be investigated, the execution time was extremely long.

Also, since the method is to first look up the dictionary all at once,
It is necessary to store all of the word information that has been looked up in the dictionary, so the memory capacity must be increased in advance: f. It was not possible to deal with long sentences without punctuation marks.

The present invention provides word identification processing using dynamic programming.
It is an object of the present invention to provide a text reading device that can reduce the required memory capacity.

[Means to solve the problem] FIG. 1 is a block diagram of the principle of the present invention.

In the figure, the text reading device uses the language analysis unit to
The human-generated kanji-kana mixed sentences are analyzed, and the speech synthesis section uses the reading and other phonological information obtained through language analysis.
and speech is synthesized based on pauses between words, accent combinations, and other prosodic information.

In the present invention, all candidate words in a character string within a predetermined range of an input sentence are looked up in a dictionary, and an evaluation value indicating which of the words located immediately before each word is optimal to select is used to search the word string. The present invention includes a word identification means 11 which performs word identification by sequentially moving the predetermined range from the beginning of the sentence while deleting unselected candidate words each time.

[Operation] In the present invention, a dictionary lookup is performed for all candidate words in a character string for each predetermined range of an input sentence, and it is shown which of the words immediately before each word is best to select. A word string is determined using the evaluation value, and candidate words that are not selected are deleted.

By moving the predetermined range from the beginning of the sentence and repeating the above-described process, dictionary lookup and word identification can be performed sequentially, and information that is no longer needed can be deleted.

Therefore, the area for storing the dictionary information can be minimized, and the memory capacity required in the word identification process using dynamic programming can be reduced.

(Example) Hereinafter, the operation of the example will be described in detail based on the drawings.

FIG. 2 is a block diagram showing an embodiment of the structure of the text reading device.

In the figure, the receiving section 41 is connected to a word identifying section 47 to which a word dictionary 45 is connected. In the word identification section 47,
A phoneme/prosody analysis section 49 is connected. The music/prosody analysis section 49 is connected to a synthesis parameter creation section 55 to which the audio file 53 is connected. A synthesized speech generator 57 is connected to the synthesis parameter generator. Synthetic speech creation section 5
A speaker 59 is connected to 7.

The word identification unit 47 identifies words or word strings while referring to the word dictionary 45, and extracts the pronunciation, accent information, and grammatical information of each word. In addition, the word dictionary 45 contains notation for word identification, grammatical attributes representing connection relationships between words, and other word attributes, as well as sounds fi (reading), accents, and accents used in the next stage of phonemic and prosody processing. Combined attributes are registered respectively.

The sound n-prosody analysis unit 49 performs a process of reconstructing into higher-level linguistic units in the order of compound words, clauses, and relationships between clauses. That is, based on the pronunciation of the word obtained by the word identification section 47,
Speech synthesis is performed according to phonological rules that obtain rendaku or phoneticization information generated by word concatenation, and prosodic rules that determine accent phrases, exhalation paragraphs, and other tonal control information based on grammatical information of word and clause boundaries. Create various control information (phonological information, prosody information) necessary for

The synthesis parameter creation unit 55 refers to the audio file 53 in which basic Japanese syllables are stored in parameter format, reads out necessary parameters based on the phonetic symbols, synthesizes a continuous parameter string, and synthesizes this. Audio creation section 5
Output to 7.

The synthesized speech creation unit 57 synthesizes a speech waveform from the synthesized parameters and causes the speaker 59 to output continuous speech.

Here, the correspondence between FIG. 1 and FIG. 2 will be shown.

The word identification means 11 includes a word dictionary 45 and a word identification section 117.
corresponds to

The 5-word analysis section corresponds to the phoneme/prosody analysis section 49.

The voice synthesis section corresponds to the voice file 53, the synthesis parameter generation section 55, the synthesized voice generation section 57, and the speaker 59.

The receiving unit 41 is a standard component of a text reading device.

The operation of the word identification process will be described below with reference to FIG.

The human text "Today's weather is sunny" is received by the receiving unit 41.

The human text "Today's weather is sunny" is input to the word identification unit 47, and after being decomposed into candidate words as shown in Figure 3, candidate words starting at the same position from the beginning of the sentence are sequentially processed. By doing so, the identification is performed sequentially.

FIG. 4 is a diagram explaining the flow of word identification processing. Moreover, FIG. 5 is a diagram explaining word identification processing.

Hereinafter, with reference to FIGS. 4 and 5, the word identification processing will be described assuming that word identification up to the word "no" has been completed and the word identification processing position has been set to the word "wa".

(a) All candidate words for the word "ha" at the processing position are looked up in the dictionary (FIG. 5(a)).

(b) For all candidate words "wa", the candidate words immediately before the word "wa" are "weather", ■ "ki", and ■ "ki" (the two words "ki" have different parts of speech and conjugations) It is determined by dynamic programming which word is most suitable as the immediately preceding word (in Fig. 5, dotted line in Fig. 3).

In dynamic programming, three evaluation values are introduced.

Each evaluation value will be explained with reference to FIG.

(1) Starting from the beginning, each word is “weather”, ■ “ki”, ■ “ki”
Cumulative evaluation value G up to (i=1.2.3).

(11) The word “wa” and each word immediately before it ■ “weather”, ■
``Ki J, ■ Adjacent word evaluation value C determined by the connection relationship when ``Ki'' is connected, (i = 1.2.3) (Mountain) Regardless of the context of the word ``wa'' in the character string Evaluation (I!!V.

Each evaluation value is set in advance and registered in the word dictionary 45.

After dictionary lookup is performed for all candidate words of the word "wa", the word immediately before the cumulative evaluation value G of the word "wa" is selected from the evaluation values mentioned above based on the following formula. be done.

(C) As a result of calculating the cumulative evaluation value G, words that end immediately before the word "wa" and are not indicated as having a high connection relationship from the word "ha" ■ "ki", ■ "ki J (cumulative evaluation value G (words that do not give the maximum value) are deleted.

In addition, along with this, the word ``Ki'' and the word ``Ten'' immediately before ``Ki'' are deleted (FIG. 5(C)).

(d) As a result of calculating the cumulative evaluation value G, the word "weather" (the word that gives the maximum value to the cumulative evaluation value G) that was not deleted as the immediately preceding word is arbitrarily determined and made the final word ( Figure 5(d)).

Since the word immediately before the word "ha" has been determined, the analysis processing position is moved to the next candidate word (word "hare", "hare") starting at the same position, and the processing from step (a) is repeated.

In this way, dictionary lookup and word identification processing can be performed sequentially from the beginning of the sentence. Further, the word identification results of the embodiment of the present invention are completely equivalent to the results obtained by the conventional method of performing word identification after all dictionary lookups, and it is possible to obtain a sufficiently high analysis rate.

〔Effect of the invention〕

As described above, according to the method of the present invention, dictionary search and word identification processing are performed sequentially, and unnecessary dictionary lookup word information is deleted each time, so that the area for storing dictionary lookup word information is The size of the data can be reduced to the minimum required for analysis, and the memory capacity to be prepared in advance can be reduced. For example, when analyzing a sentence of about 100 characters, the conventional method of searching all words in advance required preparing an area in advance to store word information of just under 1000 words.
In the method of the present invention, it is sufficient to prepare an area capable of storing word information of about 100 words.

[Brief explanation of drawings]

FIG. 1 is a block diagram of the principle of the present invention, FIG. 2 is a block diagram showing the configuration of an embodiment, and FIG. 3 is a diagram illustrating the decomposition of human text into words. FIG. 4 is a diagram for explaining the flow of word identification processing, and FIG. 5 is a diagram for explaining word identification processing. FIG. 6 is a diagram explaining dynamic programming. In the figure, 11 is a word identification means, 41 is a receiving section, 45 is a word dictionary, 47 is a word identification section, 49 is a phoneme/prosody analysis section, 53 is a voice file, 55 is a synthesis parameter creation section, 57 is a synthetic speech creation section 59 is a speaker. Fig. 1 is a block diagram of the principles of the present invention. Fig. 1 is a block diagram showing the configuration of an embodiment. Fig. 1 is a diagram explaining the flow of word identification processing.

Claims (1)

[Claims] (1) In a text-to-speech device that synthesizes speech based on the phonological and prosodic information obtained from the input of a sentence containing kanji and kana, all candidate words in a predetermined range of character strings of the input sentence are A dictionary is looked up, a word string is determined using an evaluation value indicating which of the words immediately before each word is best to select, and candidate words that are not selected are deleted each time. A text reading device comprising word identification means (11) for identifying words by sequentially moving a range from the beginning of a sentence.