JP3971577B2 - Speech synthesis apparatus and speech synthesis method, portable terminal, speech synthesis program, and program recording medium - Google Patents

Abstract

PROBLEM TO BE SOLVED: To automatically suppress an unnatural rhythm generated by the expression that is not following a grammatical rule. SOLUTION: A first accent phrase generating section 23 generates an accent phrase based on the text analysis result of a text analysis section 21. A second accent phrase generating section 24 generates an accent phrase without depending on the text analysis result. An accent phrase generating discriminating section 22 allows the generation of an accent phrase in the section 24 based on the analysis result, when an inputted text is a speech. When the inputted text is a speech not following a grammar, an accent phrase is generated without depending on the text analysis result. Thus, the generation of an unnatural pitch pattern based on an incorrect text analysis result by the section 21 is prevented.

Description

[0001]
BACKGROUND OF THE INVENTION
  The present invention relates to a speech synthesizer and a speech synthesis method for synthesizing speech from character information, a portable terminal device, a speech synthesis program, and a program recording medium.
[0002]
[Prior art]
  Conventionally, as text-to-speech synthesis for synthesizing speech from character information, a method of sequentially performing three processes of text analysis, prosody generation, and speech synthesis is known. FIG. 9 shows a block diagram of a conventional speech synthesizer.
[0003]
  The text analysis unit 1 performs the above text analysis processing, detects a word boundary from the input character information, and obtains a phoneme symbol string of each word. The prosody generation unit 2 performs the prosody generation process and assigns prosodic information such as the obtained phoneme duration, word accent, sentence intonation, and the like. Also, the speech synthesizer 3 performs the speech synthesis process, and generates a control signal for the speech synthesizer based on the synthesis units and rules stored in advance.
[0004]
  Hereinafter, the text-to-speech synthesis method will be described in detail by taking a Japanese text-to-speech synthesis device as an example. Japanese is a so-called agglutinating word in which the word boundary is not clear, unlike the language that separates the word boundary with a space like English, and therefore, the word boundary is detected by performing text analysis processing. This text analysis processing is executed by sequentially performing collation processing from the beginning of a sentence using a dictionary storing word notation and reading information and a grammar storing word connection relation information.
[0005]
  The words include independent words such as nouns and verbs, and adjuncts such as particles and auxiliary verbs. For example, the sentence “Today is the weather” is text-analyzed as follows.
    “Today (noun) / is (particle) / weather (noun) / is (auxiliary verb).”
[0006]
  A general text-to-speech synthesis method performs prosody generation processing and speech synthesis processing based on such text analysis results. The details of the prosody generation processing and speech synthesis processing are as described in, for example, “Digital Speech Processing” by Toru University (Tokai University Press).
[0007]
[Problems to be solved by the invention]
  However, the conventional text-to-speech synthesis method has the following problems. That is, in recent years, character information digitized by the spread of the Internet and the like has come to be used everyday in the general society. In particular, text written in words used in daily conversation, such as e-mails, is increasing. So-called spoken words used in daily conversation are difficult to regularize with grammar because of their diverse expressions.
[0008]
  Thus, when spoken words that cannot be defined by grammar are given as input text, text analysis is often not performed correctly. In this case, since the prosody generation process is performed on the premise that the text analysis result is correct, an unnatural prosody is generated.
[0009]
  For example, it is assumed that the spoken word sentence “I have never seen it” is analyzed as follows by the text analysis process.
    `` Look (verb) / Tako (noun) / Nato (noun) / I (noun) ''
When prosody generation processing is performed based on this text analysis result, the voice component starts at the position of the syllable “ta” due to the misanalysis of the verb “see” and the noun “tako” It becomes an unnatural accent.
[0010]
  In order to solve such a problem, Japanese Patent Laid-Open No. 11-259094 proposes a speech synthesizer whose block diagram is shown in FIG. 10, the text analysis unit 11, the prosody generation unit 12 and the speech synthesis unit 13 are the same as the text analysis unit 1, the prosody generation unit 2 and the speech synthesis unit 3 in FIG. The speech synthesizer includes a prosody editing unit 14 that corrects prosodic information given to a character string selected by a user in accordance with a user instruction. Therefore, when an unnatural prosody is generated by the prosody generation unit 12 due to an erroneous analysis of the text analysis unit 11 or the like, the user uses the correction function of the prosody editing unit 14 to correct an unnatural part of the prosody. By correcting it, it can be corrected to a natural voice.
[0011]
  However, the speech synthesizer described in the above-mentioned Japanese Patent Application Laid-Open No. 11-259094 has a problem that it is necessary for the user to make corrections manually, which places a burden on the user.
[0012]
  Accordingly, an object of the present invention is to incorporate a speech synthesizer and speech synthesis method that can automatically suppress unnatural prosody generated due to expressions outside the grammatical rules that appear in spoken language, etc., and this speech synthesizer. Another object of the present invention is to provide a portable terminal, a speech synthesis program, and a program recording medium.
[0013]
[Means for Solving the Problems]
  In order to achieve the above object, the first invention provides:
  In a speech synthesizer having text analysis means for analyzing input text, prosody generation means for generating prosody information based on the text analysis result, and speech synthesis means for synthesizing speech based on the text analysis result and prosodic information ,
  First accent phrase generation means for generating an accent phrase based on a part of speech word in the text analysis result and sending it to the prosody generation means;
  Second accent phrase generation means for generating an accent phrase based on the text analysis result and without being bound by the part-of-speech word and sending the accent phrase to the prosody generation means;
  Based on the text analysis result, it is determined which of the first accent phrase generation means and the second accent phrase generation means generates the accent phrase.Using at least one of kana chain branch probability and text analysis likelihood branch probabilityAccent phrase generation determination means for determining
With,
  The kana chain branch probability is the probability that the kana character chain belongs to the spoken text corpus, and represents the branch probability to the accent phrase generation by the second accent phrase generation means,
  The text analysis likelihood branch probability is preset according to the part-of-speech condition and represents the branch probability to the accent phrase generation by the second accent phrase generation unit.
It is characterized by that.
[0014]
  According to the above configuration, the accent phrase generation determination unit generates the accent phrase based on the input text based on the part-of-speech word in the text analysis result and the text analysis result. In addition, it is determined in advance which of the second accent phrase generation means generates without being trapped by the part-of-speech word. Therefore, for example, an accent phrase relating to an input text that is easily misanalyzed by a text analysis unit such as a spoken word can be generated by the second accent phrase generation unit without being trapped by a part-of-speech word in the text analysis result. Become.
[0015]
  furtherBy the accent phrase generation determination means,Kana character chainSpokenThis is the probability of belonging to the text corpus and represents the probability of branching to accent phrase generation by the second accent phrase generation means.Kana chain information,and, Which is preset according to the part-of-speech condition, and represents the probability of branching to the accent phrase generation by the second accent phrase generation unitA determination is made as to whether the first accent phrase generation means or the second accent phrase generation means is based on at least one of the text analysis likelihood information. Therefore, it is accurately determined that the accent phrase generation based on the input text in the spoken language should be performed by the second accent phrase generation means.
[0016]
  Also,FirstThe embodiment of the first embodimentMysteriousIn a speech synthesizer,
  The second accent phrase generation means determines at least one of the kana component information, the text analysis likelihood information, the accent phrase candidate mora number, and the position in the accent phrase candidate in the accent phrase to be generated. Use to set
It is characterized by that.
[0017]
  According to this embodiment, the second accent phrase generating means generates an accent using at least one of kana chain information, text analysis likelihood information, accent phrase candidate mora number, and position in the accent phrase candidate. The starting position of the voice component in the phrase is set. In this way, an accent phrase is correctly generated without being trapped by the word with part of speech. That is, even if input text that cannot be defined by grammar, such as spoken language, is given, generation of an unnatural pitch pattern is suppressed and a natural prosody is generated.
[0018]
  The second embodiment is the speech synthesizer of the first invention,
  The kana chain information is a probability that a voice component is started between two consecutive kana characters obtained in advance based on text data,
  The text analysis likelihood information is a probability that a voice component given by a reciprocal value of the text analysis likelihood branch probability is started,
  The number of mora of the accent phrase candidate is the probability that a voice component given to the first character of the accent phrase candidate according to the number of accent phrase candidates is started,
  The position in the accent phrase candidate is the probability that a voice component given based on the position occupied by the character in the accent phrase candidate is started.
It is characterized by that.
[0019]
  According to this embodiment, the kana chain information which is the probability that the second accent phrase generating means will start a voice component between two consecutive kana characters obtained in advance based on text data, the text Text analysis likelihood information, which is the probability that the voice component given by the reciprocal value of the analysis likelihood branch probability is started, and the voice component given according to the accent phrase candidate mora number to the first character of the accent phrase candidate At least one of the number of mora of the accent phrase candidate that is the probability of starting and the position in the accent phrase candidate that is the probability of starting the voice component given based on the position occupied by the character in the accent phrase candidate Are used to set the start position of the voice component in the generated accent phrase.
[0020]
  In addition, the second invention,
  In a speech synthesis method for analyzing input text, generating prosodic information based on the text analysis result, and synthesizing speech based on the text analysis result and the prosodic information,
  A first accent phrase generating step for generating a first accent phrase to be used when generating the prosodic information based on a part-of-speech word in the text analysis result;
  A second accent phrase generation step for generating a second accent phrase to be used when generating the prosodic information based on the text analysis result and without being bound by the part-of-speech word;
  Based on the text analysis result, it is determined which of the first accent phrase and the second accent phrase is to be generated.Using at least one of kana chain branch probability and text analysis likelihood branch probabilityAccent phrase generation determination step for determination
With,
  The kana chain branch probability is the probability that the kana character chain belongs to the spoken text corpus, and represents the branch probability to the accent phrase generation by the second accent phrase generation means,
  The text analysis likelihood branch probability is preset according to the part-of-speech condition and represents the branch probability to the accent phrase generation by the second accent phrase generation unit.
It is characterized by that.
[0021]
  According to the above configuration, the generation of the accent phrase based on the input text is generated based on the part-of-speech word in the text analysis result, or is generated based on the text analysis result without being bound by the part-of-speech word. ButThe kana character chain is a probability belonging to the text corpus of spoken words, and is set in advance according to the kana chain branching probability representing the branching probability to the accent phrase generation by the second accent phrase generating means, and the part of speech condition, Using at least one of text analysis likelihood branch probabilities representing branch probabilities for accent phrase generation by the second accent phrase generatorIt is determined in advance. Therefore, for example, an accent phrase related to input text that is likely to be erroneously analyzed during text analysis, such as spoken language, can be generated without being trapped by the word with part of speech in the text analysis result.
[0022]
  According to a third aspect of the present invention, there is provided a portable terminal equipped with the speech synthesizer of the first aspect.
[0023]
  According to the above configuration, the mobile terminal is equipped with a speech synthesizer that can give natural accent phrases to input text that cannot be defined by grammar, such as spoken language. Therefore, even when an e-mail sentence written in words used in daily conversation is received, it is possible to output it accurately with synthesized speech, and the operability of the portable terminal is improved.
[0024]
  The speech synthesis program according to the fourth aspect of the invention is a computer that converts the text analysis means, prosody generation means, speech synthesis means, accent phrase generation determination means, first accent phrase generation means, and second accent phrase in the first invention. It is characterized by functioning as generation means.
[0025]
  A program recording medium according to a fifth aspect is characterized in that the speech synthesis program according to the fourth aspect is recorded.
[0026]
  According to the configurations of the fourth and fifth inventions, as in the case of the first invention, an accent phrase relating to an input text that is easily misanalyzed by text analysis means, such as spoken language, is obtained. The generation means can generate the text analysis result without being trapped by the words with parts of speech.
[0027]
DETAILED DESCRIPTION OF THE INVENTION
  Hereinafter, the present invention will be described in detail with reference to the illustrated embodiments. FIG. 1 is a block diagram of the speech synthesizer according to the present embodiment. The text analysis unit 21 analyzes the input text to detect a word boundary, and obtains a phoneme symbol string of each word. The accent phrase generation determination unit 22 determines whether the accent phrase generation is performed by the first accent phrase generation unit 23 or the second accent phrase generation unit 24 based on the text analysis result. Then, the first accent phrase generator 23 generates an accent phrase based on the text analysis result. On the other hand, the second accent phrase generator 24 generates an accent phrase without depending on the text analysis result.
[0028]
  The prosody generation unit 25, for each accent phrase generated by the first accent phrase generation unit 23 or the second accent phrase generation unit 24, prosodic information such as phoneme duration, accent nucleus position and sentence intonation. Is granted. The speech synthesizer 26 generates a control signal for the speech synthesizer based on the synthesis units and rules stored in advance based on the assigned prosody generation information.
[0029]
  Details of the text analysis unit 21, the first accent phrase generation unit 23, the prosody generation unit 25, and the speech synthesis unit 26 are as described in, for example, “Digital Speech Processing” by Toru University (Tokai University Press). Yes, here is a brief explanation of the terminology.
[0030]
  A group of accents with one accent core is called an accent phrase. Here, the accent nucleus means a position where the pitch of the voice shifts from high to low in each word. The voice is high at the beginning, but gradually decreases due to a decrease in subglottic pressure. Such a characteristic in which the pitch (fundamental frequency) decreases with time is called a voice component, and the pitch of the whole sentence is superimposed on this characteristic by adding a word and phrase-specific pitch pattern (accent component) determined by the accent. The pattern is determined. FIG. 6 shows a process for obtaining the pitch pattern.
[0031]
  In the following, for ease of explanation, the accent phrase generation determination unit 22 looks at the text analysis result and sends the analysis result data to the first accent phrase generation unit 23 if it is a written word, while it is a spoken word. It is assumed that determination result data is sent to the second accent phrase generator 24. However, the present invention is not limited to this. For convenience of explanation, first, problems related to spoken language processing by the first accent phrase generator 23 will be described. The function of the first accent phrase generator 23 is performed by either or both of the text analyzers 1 and 11 and the prosody generators 2 and 12 in the conventional speech synthesizer shown in FIGS. Is. Next, the processing of the accent phrase generation determination unit 22 and finally the processing of the second accent phrase generation unit 24 will be described.
[0032]
  The first accent phrase generation unit 23 generates an accent phrase based on the result of text analysis by the text analysis unit 21, and is a technique that is generally performed conventionally as described above. As an example, the pitch pattern generation in the first accent phrase generation unit 23 for the sentence “Today is the weather” in accordance with the grammar storing the word connection relation information is performed according to the procedure shown in FIG. 6 as described above. It is. Thus, if the sentence conforming to the grammar is correctly analyzed, the first accent phrase generator 23 generates the pitch pattern without any problem.
[0033]
  Here, if the pitch pattern of the sentence “Nanchattee” not conforming to the grammar is generated by the first accent phrase generator 23, the result is as shown in FIG. That is, the text analysis result by the text analysis unit 21 is “na (auxiliary: final particle) / n (particle: case particle) / cha-t (verb: 5-stage wa line) / te (particle: connection particle) / ee (unknown) Word) /-(unknown word) '', and it is judged that there is an accent phrase separated between `` n '' and `` Cha '', so the next voice component at `` Cha '' is Be started. This is indicated by the fact that the voice component is divided into two in FIG. 7, which causes an unnatural pitch pattern.
[0034]
  Therefore, in the speech synthesizer in the present embodiment, the second accent phrase generation unit 24 is provided to suppress the start of the next voice component generated at the accent phrase delimiter as shown in FIG. Further, by suppressing the accent component, the variation of the pitch pattern is suppressed so as not to be largely mistaken.
[0035]
  If the text analysis by the text analysis unit 21 is performed correctly, the first accent phrase generation unit 23 is sufficient. However, in the text analysis process at the present time, there are still incomplete parts such as an error in the delimiter position, an error in judgment of the part of speech, or a process of an unknown word that is not registered in the dictionary. In particular, it is difficult to obtain accurate information for adding prosodic information from input text that is out of grammatical rules such as spoken language. In other words, even if you try to express a spoken word such as “Nanchattee” in a dictionary or grammar, there are many variations that make it difficult to regularize compared to written language.
[0036]
  The features of the spoken language appear in the kana character string. In this embodiment, the characteristic of this kana character string is captured to suppress an unnatural prosody. For example, in the case of “Nanchattee”, the pitch pattern for one accent phrase “Nanchattee” is not used because the text analysis result that “Chat” is a verb is not used. Can be generated.
[0037]
  Next, the accent phrase generation determination unit 22 that determines that “Nanchattee” should be processed by the second accent phrase generation unit 24 as one accent phrase will be described. In general, the result of text analysis of written words is a series of independent words and adjunct words. On the other hand, when a spoken language is analyzed by text, a phenomenon in which a phrase without an independent word is created or an unknown word that is not registered in the dictionary is determined due to a mistaken analysis. Therefore, if this phenomenon is captured and it is determined that the text analysis result is reliable, the first accent phrase generation unit 23 performs accent phrase generation processing; otherwise, the second accent phrase generation unit 24 generates accent phrase generation. Processing is performed.
[0038]
  Therefore, when processing is performed by the second accent phrase generator 24, it is necessary to determine in advance which unit is the accent phrase. In that case, since the result of the text analysis by the text analysis unit 21 is low in reliability, the break position and the part of speech information are not used. Then, since there is a high possibility that a word determined as an unknown word and a portion including a small letter “e” or a long sound symbol “-” are written words, a wide range is collected as an accent phrase without chopping up the accent phrase.
[0039]
  In this way, by using the information that the text analysis result includes an unknown word or that there is a character unique to spoken language, the accent phrase generation determination unit 22 determines whether the input character string is a written word or not. That is, it is possible to determine whether the first accent phrase generator 23 or the second accent phrase generator 24 processes.
[0040]
  FIG. 2 shows a flowchart of an accent phrase generation processing operation performed by the text analysis unit 21, the accent phrase generation determination unit 22, the first accent phrase generation unit 23, and the second accent phrase generation unit 24. Hereinafter, the normal text “Today is the weather” processed by the first accent phrase generator 23 and the spoken text “Nanchattee” processed by the second accent phrase generator 24 are taken as an example. A specific method of the accent phrase generation processing operation will be described.
[0041]
  In step S1, the text analysis unit 21 performs text analysis processing on the input text. In step S2, an initial value “1” is set to the word number i. In step S3, it is determined whether or not the word number i is larger than the word number N1 of the input text based on the text analysis processing result. As a result, if it is larger than N1, the accent phrase generation processing operation is terminated. On the other hand, if N1 or less, the process proceeds to step S4. In step S4, the i-th word is read and assigned to the variable Ti. In step S5, it is determined whether or not there is a continuous kana string in the word Ti. As a result, if present, the process proceeds to step S6. On the other hand, if not, the process proceeds to step S9. In step S6, the kana chain branch probability table is referenced.
[0042]
  Here, the kana chain branching probability is determined to branch to the process in the second accent phrase generation unit 24 when the first character Wi and the second character Wj of two kana characters appear in succession. (That is, the probability of being a spoken language), which is obtained in advance and stored in the kana chain branching probability table. The kana chain branching probability table is obtained as follows.
[0043]
  Based on a large amount of text data in advance, the probabilities P (Wi, Wj, L1) and P (Wi, Wj) that arbitrary hiragana character chains Wi, Wj appear in the written text corpus L1 and the spoken text corpus L2, respectively. , L2). Then, when the hiragana character chain Wi, Wj appears, the probability R (Wi, Wj) belonging to the text corpus L2 of the spoken language is expressed by the following equation:
    R (Wi, Wj) = P (Wi, Wj, L2) / {(P (Wi, Wj, L1) + P (Wi, Wj, L2)}
Ask for. The probability R (Wi, Wj) belonging to the text corpus L2 of the spoken language thus obtained is stored as a branch probability in the table in association with the first character Wi and the second character Wj. A table is obtained.
[0044]
  FIG. 3 shows an example of the kana chain branching probability table. For example, when the first character “de”, the second character “su”, and both kana character chains appear, the probability value R ( Are stored in association with each other. In this case, since the chain between the kana characters “de” and “su” is not unique to spoken language, the branch probability R (de, su) is small. On the other hand, the chain between the kana characters “na” and “−” is unique to spoken language, and the value of the branching probability R (na, −) is large.
[0045]
  In step S7, the analysis likelihood branch probability table is referenced. Here, the analysis likelihood branching probability is a probability (that is, a probability of being a spoken word) that it is determined to branch to the processing in the second accent phrase generation unit 24 because the reliability of the text analysis result is low. . For example, if the part of speech is an “unknown word”, the analysis likelihood branch probability is high, and if it is any other part of speech, it is small. Further, when the sentence head begins with an attached word, the reliability of the text analysis is considered to be low, so that the analysis likelihood branch probability is high. This analysis likelihood branch probability is stored in association with the part-of-speech condition and the branch probability determined to branch to the processing in the second accent phrase generation unit 24 when the part-of-speech condition is satisfied. It is obtained by referring to the degree branch probability table. FIG. 4 shows an example of the analysis likelihood branch probability table. For example, “is” in “Today is the weather” is an auxiliary verb and an adjunct, but because it follows the noun “weather”, it is not an adjunct to the beginning of the sentence, and the analysis likelihood branch probability value is small Value.
[0046]
  In step S8, a branch probability is calculated based on the kana chain branch probability value obtained in step S6 and the analysis likelihood branch probability value obtained in step SS7. In step S9, it is determined whether or not an accent phrase is formed. As a result, if an accent phrase is formed, the process proceeds to step S10, whereas if not formed, the process proceeds to step S13. In step S10, it is determined whether the branch probability is greater than a predetermined value α. As a result, if larger than the predetermined value α, the process proceeds to step S11, and if it is equal to or smaller than the predetermined value α, the process proceeds to step S12. In step S11, the second accent phrase generator 24 generates an accent phrase. After that, the process proceeds to step S13. In step S12, an accent phrase is generated by the first accent phrase generator 23 based on the text analysis result. In step S13, the word number i is incremented. After that, the process returns to step S3, and the process proceeds to the next word number i. When it is determined in step S3 that the word number i is larger than the number N1 of words in the input text, the accent phrase generation processing operation is terminated.
[0047]
  Hereinafter, the above-described accent phrase generation processing operation will be specifically described by taking as an example the case where the normal text “Today is the weather” is input. First, text analysis is performed on the text “Today is the weather”, and the processing result “Today (noun) / is (particle) / weather (noun) / is (auxiliary verb)” is obtained. In this case, the input text “Today is the weather” is divided into four words (N1 = 4) by the text analysis process.
[0048]
  Next, the first word “today” is read out. Since there is no continuous kana string for the word “today”, it is determined whether or not an accent phrase is to be formed. Then, since the particle continues behind, it is determined that an accent phrase is not formed, and the second word “ha” is read out. Then, even if the concatenation with the previous word “today” is taken into consideration, there is no continuous kana string, so it is determined whether or not an accent phrase is to be formed. Since the phrase “Today is” is combined with the previous word “Today”, it is determined that an accent phrase is to be formed. Here, since there is no continuous kana string and branch probability calculation processing is not performed, the branch probability is “0”, and the first accent phrase generation unit 23 generates an accent phrase based on the text analysis result.
[0049]
  Next, the processing for the third word “weather” is performed in the same manner as in the case of the first word “today”. Next, the fourth word “is” is read out. Since this word “is” has a continuous kana string (“de” and “su”), the kana chain branch probability of “de” and “su” and the analysis likelihood branch probability are obtained. It is done. Further, the branch probability is calculated based on the obtained kana chain branch probability value and the analysis likelihood branch probability value. In this case, since both the kana chain branch probability value and the analysis likelihood branch probability value are small, the branch probability value of the word “I” is small. Furthermore, it is determined that an accent phrase “we are the weather” is formed. Since the branch probability value is small, it is determined that the branch probability is smaller than α, and the first accent phrase generation unit 23 generates an accent phrase based on the text analysis result. When the content of the word number i becomes larger than the number of words “4”, the accent phrase generation unit determination processing operation with the text “Today is the weather” is terminated. In the above example, a two-chain kana string has been described as an example, but the same applies to three or more chains.
[0050]
  Next, the above-described accent phrase generation processing operation will be specifically described by taking as an example the case where the text “Nanchattee” is input in spoken language. First, text analysis is performed on the text “Nanchattee”, and the processing result “Na (particle: final particle) / n (particle: case particle) / chat (verb: 5-stage wa line) / Te ( Particle: conjunctive particle) / é (unknown word) / ー (unknown word) ”. In this case, the input text “Nanchatte” is divided into six words by the text analysis process.
[0051]
  Next, the first word “na” is read out. Since this word “Na-” has consecutive kana strings (“na” and “-”), the kana chain branch probability and the analysis likelihood branch probability of “na” and “-” are obtained. It is done. In this case, since the chain between “NA” and “-” is unique to spoken language, the value of the kana chain branching probability R (NA,-) is large. In addition, when the sentence head starts with an attached word, it is considered that the reliability of the text analysis is low, so the analysis likelihood branch probability is large. Then, the branch probability is calculated based on the obtained kana chain branch probability value and the analysis likelihood branch probability value. In this case, since both the kana chain branch probability value and the analysis likelihood branch probability value are large, the value of the branch probability of the word “na” is large.
[0052]
  Furthermore, since the accent phrase is formed together with the subsequent word “n”, it is determined that the accent phrase is not formed only by the word “na”. Next, the processing for the second word “n” is performed in the same manner as in the case of the first word “na”. When determining whether or not to form an accent phrase, it is determined that there is no break in the accent phrase with the subsequent verb “Cha”, and it is determined that an accent phrase will not be formed with “Nan” alone. Is done. This is judged from the fact that the branching probabilities of “nan” and “cha-cha” are high to some extent. Thereafter, the same processing is performed for the third word “Cha” to the sixth word “—”, and it is determined that no accent phrase is formed because any word has a high branching probability. Eventually, words separated by text analysis for the input text “Nanchatte” have a high probability of branching, so one accent phrase “Nanchatte” is formed.
[0053]
  One accent phrase formed in this way is determined to be larger than α due to the large branch probability, and the second accent generation unit 24 generates an accent phrase without using the result of text analysis. It is. Therefore, the first accent generation unit 23 can avoid generation of an unnatural accent by generating an accent phrase using a misanalysis result of text analysis.
[0054]
  Next, an accent phrase generation process using the text analysis result executed by the second accent phrase generation unit 24 will be described in detail. FIG. 5 shows a flowchart of the accent phrase generation processing operation by the second accent phrase generation unit 24. In step S11 of the accent phrase generation processing operation shown in FIG. 2, when the accent phrase candidate “Nanchatte” is sent to the second accent generation unit 24, the accent phrase generation processing operation starts.
[0055]
  In step S21, an initial value “1” is set to the mora number j of the input accent phrase candidate. In step S22, the character corresponding to the jth mora is read from the input accent phrase candidate “Nanchattee” and substituted into the variable Mj. In step S23, based on the kana chain M (j-1), Mj, the probability that the voice component is started at the character Mj portion (hereinafter referred to as the voice probability) is obtained using the kana chain information table. Is assigned to the variable a1. Here, the kana chain information table is obtained in advance based on a large amount of text data, the probability that a voice component is started between two consecutive kana characters. In the kana chain branch probability table used in the accent phrase generation determination unit 22, the probability value (branch probability value) is a probability value that is spoken language. On the other hand, the only difference is that the probability value of the kana chain information table is the above-mentioned voice probability value. Therefore, if the probability value of the kana chain information table is large, there is a high possibility that a voice component is started at the second character Mj. For example, in the case of “n” and “cha” in the input accent phrase candidate “Nanchattee”, a voice component is started between “n” and “cha” in a large amount of text data. Since there are few cases, the voice probability value is low.
[0056]
  In step S24, the value of the analysis likelihood branch probability referred to in step S7 in the accent phrase generation processing operation shown in FIG. 2 is searched based on the character string following the kana Mj, and the reciprocal value is stored in the variable a2. Assigned. Here, the fact that the analysis likelihood branch probability is high means that the reliability of the text analysis result is low. Therefore, if the value of the analysis likelihood branch probability is large, the character Mj is the start position of the voice component. Is less likely. For example, taking part-of-speech information as an example of a measure for measuring the analysis likelihood, a kana character string analyzed as an unknown word has a low probability of being a correct text analysis result, so it is less likely to be the starting position of a voice component. On the other hand, hiragana analyzed as pronouns, adverbs, etc. has a high probability that the text analysis result is correct, so it is highly likely that it will be the starting position of the voice component.
[0057]
  In the case of the kana character chain “nan” as the input accent phrase candidate, it is analyzed as a particle + particle (that is, an adjunct to the beginning of the sentence) even though it is the beginning of the sentence, so the value of the analysis likelihood branch probability is high. Become. Therefore, the value of a2 which is the reciprocal is small.
[0058]
  In step S25, the voice component start probability based on the number of mora of the input accent phrase candidate is substituted into the variable a3. If the number of mora of the input accent phrase candidate is large, the necessity of starting a voice component at the head of the accent phrase candidate becomes higher. Therefore, the voice probability of the first character is a function that increases monotonously with respect to the number of mora. Therefore, when the character Mj is the first character of the input accent phrase candidate, the voice probabilities are obtained based on the function. For example, in the case of the input accent phrase candidate “Nanchattee”, since there are 7 mora, there is a high possibility that the voice component starts at “Na”. If the character Mj is not the first character of the input accent phrase candidate, “0” is assigned to the variable a3.
[0059]
  In step S26, the start probability of the voice component based on the position occupied by the character Mj in the input accent phrase candidate is substituted into the variable a4. If the noticed character Mj is the head of the input accent phrase candidate, the voice component is more likely to be started and becomes lower as it approaches the tail. Therefore, the voice probability for the position from the head is a monotonically decreasing function. . Therefore, based on this function, the voice probabilities for the target character Mj are obtained. That is, in the case of the input accent phrase candidate “Nanchattee”, the probability that a voice component starts at “NA” is high, but the probability that a voice component starts at “Cha” is low. .
[0060]
  In step S27, the variables a1 to a4 obtained in steps S23 to S26 as described above are multiplied by the weighting factors b1 to b4 and added, and assigned to the variable A. In step S28, it is determined whether or not the value of the variable A is larger than a predetermined value β. As a result, if A> β, the process proceeds to step S29, and if A ≦ β, the process proceeds to step S30. In step S29, voice components are given to the character strings M1 to M (j-1). After that, the process proceeds to step S31. In step S30, no voice component is given to the character strings M1 to M (j-1).
[0061]
  In step S31, it is determined whether or not the mora number j is smaller than the total mora number N2 of the input accent phrase candidates. If the result is smaller than the total number of mora N2, the process proceeds to step S32. If the total number of mora is N2 or more, the accent phrase generation processing operation is terminated. In step S32, the mora number j is incremented. After that, the process returns to step S22, and the process proceeds to the process for the character corresponding to the next mora. When it is determined in step S31 that the mora number j is equal to or greater than the total number of mora N2, the accent phrase generation processing operation is terminated.
[0062]
  As described above, the second accent phrase generation unit 24 includes the voice probabilities based on the kana chain of the input accent phrase candidates, 1 / analysis likelihood branching probability, the voice establishment based on the number of mora, and the accent phrase candidates. Based on the above voice establishment based on the occupied position, it is determined whether or not a new voice component start position is set for the input accent phrase candidate. Therefore, when the accent phrase candidate “Nanchattee” based on the spoken language text is input, the above voiced probability based on the kana chain for the character string “Cha”, 1 / analysis likelihood branching probability, number of mora The values for the voice establishment based on the above and the voice establishment establishment based on the position occupied in the accent phrase candidate are both small, and the voice component is not started with the character string “Cha”. In this way, it is suppressed that the voice component is divided into two and causes an unnatural pitch pattern.
[0063]
  As described above, in the present embodiment, in addition to the first accent phrase generation unit 23 that generates an accent phrase based on the text analysis result by the text analysis unit 21, the accent phrase does not depend on the text analysis result. A second accent phrase generator 24 is provided. Then, the accent phrase generation determination unit 22 determines that the first accent phrase generation unit 23 generates the accent phrase based on the text analysis result when the input text is a written word. On the other hand, if it is a spoken language, the second accent phrase generator 24 determines that it is performed.
[0064]
  Therefore, when the input text is a spoken word “Nanchattee” that does not conform to the grammar, the second accent phrase generation unit 24 may generate an accent phrase without depending on the text analysis result. it can. As a result, the next voice component is not started at “Cha” as in the case where an accent phrase is generated based on an erroneous text analysis result by the text analysis unit 21, and an unnatural pitch pattern is generated. It can be prevented from being generated.
[0065]
  At this time, the accent phrase generation determination unit 22 includes a kana chain branching probability table that associates a chain of two kana characters with a probability to branch to the processing in the second accent phrase generation unit 24, a part of speech condition, Whether the first accent phrase generation unit 23 performs processing with reference to the analysis likelihood branch probability table that associates the probability to be branched to the processing in the second accent phrase generation unit 24 when the part of speech condition is satisfied The second accent phrase generator 24 determines whether to process. Therefore, based on the kana character string information peculiar to the spoken language and the part-of-speech condition, it is possible to accurately determine whether or not the second accent phrase generation unit 24 performs processing.
[0066]
  Further, the second accent phrase generation unit 24 is configured to generate the voice based on the kana chain of the accent phrase candidates input from the accent phrase generation determination unit 22, 1 / analysis likelihood branching probability, and the voice based on the number of mora. Whether or not to newly set a voice component start position for the input accent phrase candidate is determined based on the voice establishment and the voice establishment based on the position occupied in the accent phrase candidate. Therefore, even when text that cannot be specified by grammar, such as spoken language, is input, it is suppressed that an unnatural voice is given based on an erroneous text analysis result, and a natural prosody is generated. .
[0067]
  In the above embodiment, whether the accent phrase is generated by the first accent phrase generator 23 or the second accent phrase generator 24 is determined by the accent phrase generation determination unit 22 as to whether it is a written word or a spoken word. However, the present invention is not limited to this. In short, it is only necessary to determine that the second accent phrase generation unit 24 processes a sentence that cannot be defined by a grammar that causes an erroneous analysis by text analysis.
[0068]
  Text input by spoken language as described above is often performed when a mail sentence is input by a portable terminal. And in the said portable terminal device, since there is a limit in the number of display characters on a screen, it is desirable to output the received mail sentence by a synthetic voice. Therefore, by mounting the speech synthesizer as described in the above embodiment on the portable terminal, the function of the portable terminal can be greatly improved.
[0069]
  By the way, the text analysis unit 21, the accent phrase generation determination unit 22, the first accent phrase generation unit 23, and the second accent phrase generation unit 24, the text analysis unit, the accent phrase generation determination unit, and the first accent phrase in the above embodiment. The functions as the generating means and the second accent phrase generating means are realized by a speech synthesis processing program recorded on a program recording medium. The program recording medium in the above embodiment is a program medium composed of a ROM (Read Only Memory). Alternatively, it may be a program medium that is loaded into an external auxiliary storage device and read out. In any case, the program reading means for reading the voice synthesis processing program from the program medium may have a configuration in which the program medium is directly accessed and read, or the random access memory (RAM). You may have the structure which downloads to the program storage area (not shown) provided, and accesses and reads the said program storage area. It is assumed that a download program for downloading from the program medium to the program storage area of the RAM is stored in the main unit in advance.
[0070]
  Here, the program medium is configured to be separable from the main body side, and is a tape system such as a magnetic tape or a cassette tape, a magnetic disk such as a floppy disk or a hard disk, a CD (compact disk) -ROM, or MO (magneto-optical). Optical discs such as discs, MDs (mini discs), DVDs (digital video discs), card systems such as IC (integrated circuit) cards and optical cards, mask ROMs, EPROMs (ultraviolet erasable ROMs), EEPROMs (electrical This is a medium that carries a fixed program including a semiconductor memory system such as an erasable ROM) and a flash ROM.
[0071]
  Further, when the speech synthesizer in the above embodiment has a configuration that includes a modem and can be connected to a communication network including the Internet, the program medium is fluidly downloaded by downloading from the communication network. It can be a medium that carries the program. In this case, it is assumed that a download program for downloading from the communication network is stored in the main device in advance. Or it shall be installed from another recording medium.
[0072]
  It should be noted that what is recorded on the recording medium is not limited to a program, and data can also be recorded.
[0073]
【The invention's effect】
  As is clear from the above, the speech synthesizer of the first invention is based on the first accent phrase generating means for generating an accent phrase based on the part-of-speech word in the text analysis result, the above text analysis result, and the above Second accent phrase generating means for generating an accent phrase without being constrained by a word with part of speech, and generating the accent phrase by the first accent phrase generating means by the accent phrase generation determining means or generating the second accent phrase Therefore, for example, an accent phrase relating to input text that is easily misanalyzed during text analysis, such as a spoken word, is captured by a part-of-speech word in the text analysis result by the second accent phrase generating means. It becomes possible to generate without.
[0074]
  Therefore, according to the present invention, it is possible to give a natural pitch pattern to text that cannot be defined by grammar, such as spoken language, and to suppress unnatural prosody.
[0075]
  furtherThe accent phrase generation determination meansIsAs a criterion for the above judgment,Kana character chainSpokenThis is the probability of belonging to the text corpus and represents the probability of branching to accent phrase generation by the second accent phrase generation means.Kana chain information,and, Which is preset according to the part-of-speech condition, and represents the probability of branching to the accent phrase generation by the second accent phrase generation unitSince at least one of the text analysis likelihood information is used, it is possible to accurately determine that the accent phrase generation based on the text that cannot be defined by the grammar such as spoken language should be performed by the second accent phrase generation means. Can do.
[0076]
  Also,FirstIn the embodiment of the present invention, the second accent phrase generating means determines the start position of the voice component in the accent phrase to be generated, the kana chain information, the text analysis likelihood information, the number of mora of the accent phrase candidate, and the position in the accent phrase candidate. Therefore, the accent phrase can be generated correctly without being trapped by the part-of-speech word in the text analysis result. Therefore, even if an input text that cannot be defined by grammar such as spoken language is given, generation of an unnatural pitch pattern can be suppressed and a natural prosody can be generated.
[0077]
  In the second embodiment, kana chain information which is a probability that a voice component is started between two consecutive kana characters obtained in advance based on text data by the second accent phrase generating means, Text analysis likelihood information, which is the probability of starting a voice component given by the reciprocal value of the text analysis likelihood branching probability, and voices given to the first character of accent phrase candidates according to the number of accent phrase candidate mora The number of mora of the accent phrase candidate that is the probability that the component will start, and the position in the accent phrase candidate that is the probability that the voice component given based on the position occupied by the character in the accent phrase candidate Using at least one, the start position of the voice component in the generated accent phrase is set. Therefore, even if input text that cannot be defined by grammar such as spoken language is given, generation of an unnatural pitch pattern can be suppressed and a more natural prosody can be generated.
[0078]
  In the speech synthesis method according to the second aspect of the present invention, generation of an accent phrase based on the input text is generated based on a word with a part of speech in a text analysis result or based on the text analysis result and is trapped by the word with a part of speech. To generate withoutThe kana character chain belongs to the spoken text corpus and is set in advance according to the kana chain branching probability representing the branching probability to the accent phrase generation by the second accent phrase generating means, and the part of speech condition, Using at least one of text analysis likelihood branching probabilities representing branching probability to accent phrase generation by the second accent phrase generating unitSince the accent phrase is generated in advance according to the determination result, the accent phrase related to the input text that is easily misanalyzed at the time of text analysis, such as spoken language, is captured by the word with part of speech in the text analysis result. It becomes possible to generate without.
[0079]
  In addition, since the portable terminal device of the third invention is equipped with the speech synthesizer of the first invention that can give a natural accent phrase to input text that cannot be defined by grammar like spoken language, Even when an e-mail message written in a language used in conversation is received, it is possible to output it accurately with synthesized speech, and the operability of the portable terminal can be improved.
[0080]
  The speech synthesis program according to the fourth aspect of the invention is a computer that converts the text analysis means, prosody generation means, speech synthesis means, accent phrase generation determination means, first accent phrase generation means, and second accent phrase in the first invention. It functions as a generation means. A program recording medium according to a fifth aspect records the speech synthesis program according to the fourth aspect. Therefore, as in the case of the first invention, an accent phrase related to input text that is easily misanalyzed by text analysis means such as spoken words is converted into a part-of-speech word in the text analysis result by the second accent phrase generation means. It becomes possible to generate without being caught.
[Brief description of the drawings]
FIG. 1 is a block diagram of a speech synthesizer according to the present invention.
FIG. 2 is a flowchart of an accent phrase generation processing operation performed by the speech synthesizer shown in FIG.
FIG. 3 is a diagram illustrating an example of a kana chain branch probability table;
FIG. 4 is a diagram illustrating an example of an analysis likelihood branch probability table.
FIG. 5 is a flowchart of an accent phrase generation processing operation performed by a second accent phrase generation unit in FIG. 1;
FIG. 6 is a diagram illustrating a process of obtaining a pitch pattern.
7 is a diagram showing a process of generating a pitch pattern based on spoken language by a first accent phrase generator in FIG. 1; FIG.
8 is a diagram showing a process of generating a pitch pattern based on spoken language by a second accent phrase generator in FIG. 1. FIG.
FIG. 9 is a block diagram of a conventional speech synthesizer.
FIG. 10 is a block diagram of a conventional speech synthesizer different from FIG.
[Explanation of symbols]
21 ... Text analysis part,
22 ... Accent phrase generation determination unit,
23. First accent phrase generator,
24 ... second accent phrase generator,
25 ... Prosody generation part,
26: Speech synthesis unit.

Claims (7)

In a speech synthesizer having text analysis means for analyzing input text, prosody generation means for generating prosody information based on the text analysis result, and speech synthesis means for synthesizing speech based on the text analysis result and prosodic information ,
First accent phrase generation means for generating an accent phrase based on a part of speech word in the text analysis result and sending it to the prosody generation means;
Second accent phrase generation means for generating an accent phrase based on the text analysis result and without being bound by the part-of-speech word and sending the accent phrase to the prosody generation means;
Based on the text analysis result, whether the accent phrase is generated by the first accent phrase generation means or the second accent phrase generation means is determined by using at least one of the kana chain branch probability and the text analysis likelihood branch probability. An accent phrase generation judging means for judging using ,
The kana chain branch probability is the probability that the kana character chain belongs to the spoken text corpus, and represents the branch probability to the accent phrase generation by the second accent phrase generation means,
The speech synthesis apparatus characterized in that the text analysis likelihood branch probability is preset according to a part of speech condition and represents a branch probability to the accent phrase generation by the second accent phrase generation unit. . The speech synthesis apparatus according to claim 1,
The second accent phrase generating means determines at least one of the kana chain information , the text analysis likelihood information , the number of mora of the accent phrase candidate, and the position in the accent phrase candidate in the accent phrase to be generated. speech synthesis apparatus characterized that you set using. The speech synthesis apparatus according to claim 2 ,
The kana chain information is a probability that a voice component is started between two consecutive kana characters obtained in advance based on text data,
The text analysis likelihood information is a probability that a voice component given by a reciprocal value of the text analysis likelihood branch probability is started,
The number of mora of the accent phrase candidate is the probability that a voice component given to the first character of the accent phrase candidate according to the number of accent phrase candidates is started,
Positions in the accent phrase candidates, the speech synthesis apparatus according to claim probability der Rukoto voices freshly component given based on the position occupied by the characters in the accent phrase candidates is started. In a speech synthesis method for analyzing input text, generating prosody information based on the text analysis result, and synthesizing speech based on the text analysis result and the prosody information,
A first accent phrase generating step for generating a first accent phrase to be used when generating the prosodic information based on a part-of-speech word in the text analysis result;
A second accent phrase generation step for generating a second accent phrase to be used when generating the prosodic information based on the text analysis result and without being bound by the part-of-speech word;
Based on the text analysis result, which one of the first accent phrase and the second accent phrase is generated is determined using at least one of the kana chain branch probability and the text analysis likelihood branch probability. An accent phrase generation determination step for determining ,
The pseudonym chain branching probability, kana characters chain is a probability of belonging to the text corpus of spoken language, Ri you represent the branch probability of the accent phrase generated by the second accent phrase generating means,
The speech synthesis method, wherein the text analysis likelihood branching probability is preset according to a part-of-speech condition and represents a branching probability to the accent phrase generation by the second accent phrase generation unit .   A portable terminal device comprising the speech synthesizer according to any one of claims 1 to 3. Computer
The speech synthesis program according to claim 1, wherein the speech synthesis program functions as text analysis means, prosody generation means, speech synthesis means, accent phrase generation determination means, first accent phrase generation means, and second accent phrase generation means.   A computer-readable program recording medium on which the speech synthesis program according to claim 6 is recorded.

Images