JP2004246085A - Method, device, and program for speech synthesis - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To synthesize a speech which would not incorrectly be heard by extracting a characteristic expression from a text and changing a way of pronouncing the characteristic expression part. <P>SOLUTION: (S1) Reading printed in KANA (Japanese syllabary) of KANJI (Chinese character) is given to a text 1 of KANA-KANJI mixed sentences by using a reading dictionary and an accent dictionary 2 and an accent type is set. (S2) The proper expression (e.g. a telephone number) is extracted by using morpheme information and reading information outputted through text analysis. (S3) Rhythm parameters (fundamental frequency pattern of a speech, continuance of a phoneme, power of the speech, etc.) are generated. (S4) General speech (continuously generated speech) synthesis unit 6 are used for selection in speech synthesis of an ordinary text and the proper expression (a part sandwiched between a label and a tag) is selected by using speech synthesis units of monosyllabic pronunciation (independent pronunciation of a monosyllable). (S5) Signal processing of a selected speech synthesis unit is performed to match rhythm parameters, thereby outputting a synthesized speech. <P>COPYRIGHT: (C)2004,JPO&NCIPI

Description

[0001]
TECHNICAL FIELD OF THE INVENTION
The present invention relates to a speech synthesis method and apparatus for converting a text as an input into speech, and a speech synthesis program.
Information using text as a medium has been transmitted and accumulated in large volumes through books, newspapers, magazines, and the Internet, and will continue to be transmitted in large quantities in the future. If such information can be provided not only as text to the user but also converted into speech, not only can the means of accessing the user's information be diversified, but also the text information can be effectively used. It becomes. For example, event information provided as text on the Web can be accessed by telephone and heard by voice if there is a technology capable of converting text to voice.
[0002]
[Prior art]
As a technique for converting text to speech, speech synthesis from text has been developed.
This technology converts the input text into speech by performing the following processing.
(1) Analyzes text, assigns kanji reading kana, and sets accent type.
(2) Generation of prosodic parameters such as the fundamental frequency pattern of speech, the duration of phonemes, and the power of speech.
(3) Select a speech segment suitable for the reading kana from the speech database.
{Circle around (4)} Signal processing of the selected voice segment to match the prosody parameters.
At present, the content of the text is somehow audible, but the intonation is unnatural or incorrect if you do not listen carefully.
[0003]
(Conventional example 1)
The speech synthesizer of the first conventional example will be described with reference to FIG.
For the input text from the text input unit 20, the Japanese processing unit 21 performs morphological analysis to convert the reading kana with reference to the pause position, word / segment delimiter, or dictionary, and to add accents.
The phrase pattern calculation unit 22 calculates a phrase component (the pitch of a sound when speaking in a pause between pauses) obtained from the number of moras (the minimum unit of a sound represented by a consonant + a vowel) included in the text ( Inflection control pattern of phrase component).
The accent pattern generation processing unit 23 includes a basic accent pattern generation processing unit 24, a basic accent pattern table 25, and a correction processing unit 26. The basic accent pattern table 25 is a table in which pattern data obtained by pitch analysis of real speech is stored in advance, and the accent amounts of a plurality of points are tabulated for each combination of the mora accent and the accents before and after the mora. . The basic accent pattern generation processing unit 24 is provided with the number of mora in the delimiter and the accent pattern of the mora as a Japanese processing result of the input text, and reads the basic accent pattern corresponding to the input accent environment from the basic accent pattern table 25. Perform processing. The correction processing unit 26 uses the basic accent pattern generated by the basic accent pattern generation processing unit 24, obtains a correction amount based on the number of mora, the mora position, and the like within the delimiter of the accent pattern obtained by processing this pattern in Japanese. The basic accent pattern is corrected by the amount and output as accent pattern data (accent component inflection control pattern).
Next, the inflection control patterns of the phrase component and the accent component are superimposed to obtain the intonation inflection control pattern of the input text, which is input to the speech synthesis unit 27. The speech synthesis unit 27 adjusts the pitch of each syllable according to the intonation control pattern, and obtains a synthesized speech as a response output of the articulation filter from the sound source wave pattern associated with each syllable and the parameters of the articulation filter (see Patent Document 1).
[0004]
(Conventional example 2)
Referring to FIG. 3, a description will be given of an embedded type combined speech synthesizer according to a second conventional example.
In the input character data 30, a character data normalization unit 31 converts characters mixed with kanji and kana into normalized character data, and outputs the normalized character data to a language processing analysis unit 33. The conversion table 32 is a conversion table for converting a sentence mixed with kanji and kana into normalized character data, and is referred to by the character data normalization unit 31.
The language processing analysis unit 33 converts the normalized character data into a series of words, identifies parts of speech and inflected forms, analyzes the attributes of specific words in addition to prosodic information such as pause setting, reading addition, and accent setting, Outputs characters with attributes and prosodic symbols. The word dictionary unit 34 has tens of thousands to hundreds of thousands of words, and describes grammatical information, readings, accent forms, and the like. Specific words include, for example, A, B, C, D,. A type code (to be described later) such as “*” is given, and is referred to by the language processing analysis unit 33.
The control unit 35 determines from the sentence example pattern table 36 whether the attribute and the (kana) character with the prosody symbol output from the language processing analysis unit 33 are a phonogram character string for rule synthesis or a fitted phonogram character string. (The determination will be described later). That is, it is determined whether the word identification attribute sequence obtained from the language processing analysis unit 33 matches the sentence example pattern in the sentence example pattern table 36. If it is determined that the phonetic character string for rule synthesis is used, ordinary text-to-speech synthesis control is performed as follows.
[0005]
(Rule synthesis type prosody generation processing)
The rule synthesis type prosody generation unit 51 generates intonation from the position of the pause in the phonogram string for rule synthesis, and generates a pitch pattern by overlapping the intonation with the accent. The prosody table 50 describes rules for setting poses, accents, and the like, and is referred to by the rule synthesis type prosody generation unit 51.
The phoneme duration generation unit 61 sets the duration of the phonemes constituting the phonogram to realize natural timing (rhythm) of the utterance. The phoneme length table 60 describes a unique time length rule of the phoneme in consideration of the influence of the adjacent phoneme, and is referred to by the phoneme duration generation unit 61.
When connecting the phonemes constituting the rule-synthesizing phonogram string and synthesizing a continuous voice, the waveform connection unit 62 performs an interpolation process in which the feature parameters such as the phoneme duration and the pitch pattern are expressed between phonemes. The continuous sound that has been subjected to the interpolation processing is output as PCM data 64. The phoneme dictionary unit 63 stores the phoneme waveform data as a synthesis unit, and is referred to by the waveform connection unit 62.
[0006]
(Inset type prosody generation processing)
The determination and the case where it is determined that the character string is a fit-in-format phonetic character string will be described.
For example, when the input character data (traffic information data) is “Congestion 2 km on the west of the second Shinmei road near Tazu,” the normalized character data to the language processing analysis unit 33 indicates “second Shinmei. "Road" is converted to "Danishin Mae Road" and Kana, meaning the road name, "Westbound" is converted to "Nishiyuki" and Kana, and the direction is indicated, and "Tamazu" is converted to "Tamamatsu" and Kana. It is converted and means a place name, "traffic jam" is converted to "jutai" and kana, and means a degree, and "2km" is converted to "ni-kilo" and kana and means a distance. In this case, the word dictionary unit has words in “Danishin mailroad”, “Nishiyuki”, “Tama”, “jutai”, “Nikuro”, etc., and these words include, for example, road names, directions, place names, A word having an attribute of A, B, C, D, or E is assigned to a word meaning degree or distance.
In the sentence example pattern table 36, word strings of limited sentence examples, for example, A + B + C + D + E, A + C + B + D + E,.
[0007]
When one sentence composed of the attribute and the Kana character with the prosody symbol is input, the control unit 35 compares the sentence with the sentence example pattern table 36. If the word attribute string does not match the sentence example of the sentence example pattern table at all, the pitch is adjusted. The rule synthesis type prosody generation processing is selected for the generation processing. On the other hand, if they match, the embedding type prosody generation processing is selected, and a phonetic character string in the inset format is generated.
The inlaid format phonetic character string is formed from the sentence number and the reading of the inlaid word. The inlaid sentence example table 40 is created as a phonetic character string of a connecting word such as a particle, an auxiliary verb, or a verb that connects the words constituting the sentence example, corresponding to each of the sentence examples in the sentence example pattern table 36. It consists of "ha,""fukinde," and "death." In this case, the inset-type prosody generation unit 42 inserts each word of the attribute string of the word forming the inset format phonetic character string between the phonogram character strings of connective words such as particles, auxiliary verbs, and verbs in the sentence example parameter table. Fit in. That is, in the above example, "A", "B", "C", "C", "DE", "Death".
The inset type prosody generation unit 42 refers to the prosody table 41 having the same configuration as the prosody table 51 for the accent and pause of the inset word, and refers to the inset sentence example table 40 for the intonation of the entire sentence to superimpose the intonation and accent. To form a pitch pattern. Similarly to the above, the waveform connection unit 62 sets the phoneme duration for the phoneme of the inlaid phonetic character string, sets the pitch pattern from the inlaid prosody generation unit 42, and connects the phoneme waveforms. It is output as data 64.
[0008]
[Patent Document 1]
Patent No. 3070127 (FIG. 1, paragraph (0016))
[Patent Document 2]
Patent No. 3192981 (FIG. 2, FIG. 3, paragraphs (0008) to (0014))
[0009]
[Problems to be solved by the invention]
In conventional speech synthesis from text, the manner of pronunciation was not controlled according to the content of the text. However, humans change the way of pronunciation according to the content of the text. A simple example is reading a phone number. When giving a telephone number to a person, speak slowly and with a single sound, one note at a time, so as not to give a wrong number. By doing so, the clarity of the phoneme is raised. A similar pronunciation is also used to convey the name of the restaurant at the meeting place. In conventional speech synthesis from text, in such a case, the pronunciation method was not changed, so that there was a problem that the telephone number and the restaurant name were not correctly transmitted to the user.
According to the present invention, by changing the manner of pronunciation in accordance with the contents of such texts, it is possible to realize speech synthesis without mistaken listening, and to enable smooth information transmission. Further, as more advanced examples, it is conceivable to change the pronunciation method according to the content of the text, such as reading out news or reading out an emergency.
[0010]
[Means for Solving the Problems]
In order to solve the above-mentioned problem, the present invention estimates, from a text, a portion where the pronunciation method should be changed according to the content of the text and, when synthesizing this portion, switches between prosodic parameters and a voice database. Get proper speech synthesis.
[0011]
BEST MODE FOR CARRYING OUT THE INVENTION
An embodiment of the present invention will be described with reference to the telephone number reading described above with reference to the speech synthesizer shown in FIG.
The input is text 1 of a sentence containing kana-kanji characters created by a word processor or the like.
S1: The reading dictionary and the accent dictionary 2 are dictionaries used for text analysis of the text analysis unit 3. The text analysis unit 3 uses the reading dictionary and the accent dictionary 2 to assign the reading kana of the kanji to the text 1 of the sentence mixed with the kana-kanji and sets the accent type. These configurations need only be equivalent to speech synthesis processing from ordinary text, and are not limited to the configurations described here.
S2: The named entity extraction unit 4 is a module that extracts a phrase considered as a telephone number (extracts a named entity) by using the morphological information and the reading information output by the text analysis. Labels or tags are added before and after the extracted named expression.
[0012]
The named entity extraction will be described.
The named entity extraction rule generation system morphologically analyzes a training document (text) prepared in advance, divides the word into words, and gives information such as a part of speech name and a constituent character type to each word. From the word strings obtained in this way, the word strings that make up the named entity are extracted, and a large number of eigenwords are obtained by generalization means such as empirical rules and minimum generalization by referring to the correct answer list prepared in advance corresponding to the training document. Generate rules for expression extraction. Then, these rules are applied to the training document independently of each other, and a record of where the rule matches in the training document is stored. What is included in this record is a candidate for a named entity output by the system for the training document. Then, when combining a plurality of rules, from among all the candidates included in the record corresponding to those rules, an example of a candidate to be finally output is determined in consideration of a competitive relationship and a priority order. Choose out based on clear criteria. As a result, if there is a rule in which the frequency or ratio of incorrect answers in the training document is extremely high, the rule is deleted. However, it is possible to know at which position in the training document the rule is correct and at which position the rule is incorrect. Therefore, by comparing the word string before and after the correct answer part with the word string before and after the incorrect answer part and adding constraints, it can be determined whether or not a rule that improves the score in the training document can be created. If so, add rules with restrictions.
Using the named entity extraction rule generation system created in this way, the named entities in any document are extracted based on the generated rules, and if the extracted named entities partially overlap, the document In which the description start position is earlier is preferentially extracted, and if the description start position is the same, the later description end position is preferentially extracted, and furthermore, the same expressions but different types of named entities If so, the rule with the higher priority given in advance to each rule used for extracting each named expression is preferentially extracted. (See paragraphs (0009) and (0010) of JP-A-2001-318792)
[0013]
From now on, the phone number and the number sequence recognized are subjected to a process for clearly producing one note or one note. Note that named entity extraction is a technology for extracting and classifying important elements (named entities) such as a person's name, place name, amount of money, date and time from text, and is not limited to the telephone number of the present embodiment, but may be a person's name or a restaurant's name. Can be extracted, and a pronunciation method described below is applied to these to synthesize a speech with little misunderstanding. Further, in the case where a pattern is formed in a notation such as a telephone number, a process of certifying a telephone number by pattern matching of the notation instead of extracting the unique expression may be considered.
S3: In the prosody generation of the prosody generation unit 5, processing equivalent to speech synthesis from ordinary text is performed (that is, generation of prosody parameters (basic frequency pattern of speech, duration of phoneme, power of speech, etc.)). However, in particular, for the telephone number portion, the phoneme is made to sound slowly by setting the duration of the phoneme longer. This improves clarity.
S4: The speech database includes a general-purpose speech (continuous utterance speech) synthesis unit 6 storing a waveform corresponding to a continuously generated voice and a single syllable utterance speech synthesis unit 7 storing a speech waveform corresponding to a single syllable. The general-purpose speech synthesis unit 6 stores a speech synthesis unit prepared for synthesizing an arbitrary speech.
[0014]
The speech synthesis unit selection unit 8 selects a speech synthesis unit suitable for a phoneme sequence to be synthesized from the general speech synthesis unit 6. In speech synthesis from ordinary text, in order to realize a smooth phoneme sequence such that an announcer reads out news using a general-purpose speech synthesis unit 6 in a speech database, a speech synthesis unit that takes phoneme sequence into consideration is used. Is Such a speech synthesis unit has the disadvantage that the clarity of one sound or one sound is reduced, but in the case of news, etc., humans listen while appropriately estimating words from the surrounding context, etc. Sex decline is not a problem. However, in the case of a telephone number or the like, there is no context information, so that the user may be mistakenly heard. Therefore, the named entity extracted by the named entity extraction unit 4 (the portion sandwiched between the label and the tag) is converted into a single syllable utterance (single syllable alone) without considering the effect of the phonological environment before and after the single syllable uttered speech synthesis unit 7. Is prepared, and a telephone number is synthesized using this unit. The monosyllable utterance is one sound, one sound, and a stuttering utterance. However, since it has a feature of high clarity, it is possible to obtain a speech synthesis that does not easily cause a mistake in listening.
S5: In the audio signal processing unit 9, the basic frequency, the duration length, and the like set by the prosody generation unit 5 by the audio signal processing for the audio synthesis unit selected by the audio synthesis unit selection unit 8 ( To the prosody parameter). It is to be noted that a case may be considered in which only the voice is connected without setting the value set in the prosody generation unit 5. This is because conversion by audio signal processing may cause deterioration in sound quality.
[0015]
The speech synthesizer of the present invention includes a computer having a CPU and a memory, a terminal used by a user, and a machine-readable recording medium such as a CD-ROM, a magnetic disk device, and a semiconductor memory.
The speech synthesis program recorded on the recording medium or the speech synthesis program transmitted via the line is read by the computer, and the above-described components are realized on the computer, and each process is executed.
[0016]
【The invention's effect】
As described above, according to the present invention, speech can be synthesized in a pronunciation manner suitable for a text to be read. As a result, the overall performance of speech synthesis from text can be improved.
[Brief description of the drawings]
FIG. 1 is a diagram showing a configuration example of a speech synthesizer of the present invention.
FIG. 2 is a diagram showing a configuration of a speech synthesizer of Conventional Example 1.
FIG. 3 is a diagram showing a configuration of a speech synthesizer of Conventional Example 2.
[Explanation of symbols]
1 ... Text, 2 ... Reading dictionary, accent dictionary, 3 ... Text analysis unit, 4 ... Named expression extraction unit, 5 ... Prosody generation unit, 6 ... General-purpose speech synthesis unit, 7 ... single syllable uttered speech synthesis unit, 8 ... speech synthesis unit selection unit, 9 ... speech signal processing unit, 10 ... synthesized speech

Claims (9)

In a text-to-speech method that inputs any text, converts it into speech, and reads it out,
A speech synthesis method characterized by extracting a named entity from a text and reading out the named entity portion in a different manner of pronunciation from a portion other than the named entity. The speech synthesis method according to claim 1,
A speech synthesis method characterized in that the way of changing the pronunciation is performed by controlling prosodic parameters. The speech synthesis method according to claim 1 or 2,
A speech synthesis method characterized by specially providing speech data suitable for reading as a way of changing pronunciation, and synthesizing speech using the speech data. In a text-to-speech device that inputs any text, converts it into voice, and reads it out,
A named entity extracting means for extracting a named entity from the text,
Means for reading out the named entity part by changing the way of pronunciation with a part other than the named entity. The speech synthesizer according to claim 4,
A voice synthesizing apparatus characterized in that the means for reading out the named entity part by changing the pronunciation method from the part other than the named entity is performed by controlling prosodic parameters. The speech synthesizer according to claim 4 or 5,
The voice synthesizing device is characterized in that the means for reading out the named entity part by changing the way of pronunciation is different from the part other than the named entity, specifically holding voice data suitable for reading and synthesizing voice using the voice data. A text-to-speech program that allows a computer to execute a text-to-speech method that inputs an arbitrary text, converts it into voice, and reads it out,
The process of extracting named entities from the text;
A speech synthesis program that causes a computer to execute a process of reading out the part other than the named entity from the part other than the named entity in a different manner of pronunciation. The speech synthesis program according to claim 7,
A speech synthesis program in which the process of reading out the named entity portion by changing the pronunciation method with the portion other than the named entity is performed by controlling the prosodic parameters. The speech synthesis program according to claim 7 or 8,
In the process of reading out the named entity part by changing the way of pronunciation with the part other than the named entity, speech data suitable for reading is held specially, and using it, the named entity part can be used as the part other than the named entity and the pronunciation method. A speech synthesis program that synthesizes speech by changing.

Images