JP4736524B2 - Speech synthesis apparatus and speech synthesis program - Google Patents

Description

  The present invention relates to a speech synthesizer and a speech synthesizer program, and more particularly to a speech synthesizer and a speech synthesizer program capable of outputting speech without a sense of discomfort.

Conventionally, in speech synthesis, in the case of synthesizing speech of a sentence with an increasing pitch unlike a plain sentence, the tone of a sentence such as a question sentence, a sentence requesting consent, a sentence prompting an action, etc., the invention described in Patent Document 1 is used. Like the pitch pattern generation method of the above, a plurality of correction patterns that increase the pitch are prepared in advance, and the pitch pattern of a sentence with an increased pitch is generated by adding the end position to the basic pattern. As in the speech synthesizer according to the invention described in 2, the pitch correction amount is set for each mora position according to the number of mora in the exhalation paragraph, and the original pitch is corrected.
JP 2004-226505 A Japanese Patent Laid-Open No. 2002-196800

  However, the pitch pattern generation method and the speech synthesizer according to the inventions disclosed in Patent Documents 1 and 2 only have a correction amount added to the original pitch. is there.

  The present invention has been made to solve the above-described problems, and an object of the present invention is to provide a speech synthesizer and a speech synthesizer program that can output speech without a sense of incongruity.

  In order to solve the above-described problem, in the speech synthesizer according to the first aspect of the present invention, a phoneme model created from phoneme data obtained by analyzing speech into an acoustic parameter sequence and a prosody model created from fundamental frequency data obtained by analyzing speech Acoustic dictionary storage means for storing an acoustic dictionary that is a set of acoustic models including at least the above, and the voice of a voice that utters a sentence whose tone at the end of the sentence is different from a plain sentence, such as a question sentence, a sentence requesting consent, and a sentence prompting an action An interrogative sentence acoustic model including at least an interrogative sentence phonological model created from phonological data and an interrogative sentence prosodic model created from the fundamental frequency data of speech uttered by a sentence whose tone is different from that of a plain sentence A question sentence acoustic dictionary storage means for storing a question sentence acoustic dictionary different from the acoustic dictionary; Language analysis means for determining an accent type indicating the accent position for each cent phrase and determining reading of a sentence that generates the speech; and from the acoustic dictionary based on an analysis result analyzed by the language analysis means. An acoustic model selecting means for selecting an acoustic model; and a speech generating means for generating speech based on the phonological model and the prosodic model constituting the acoustic model selected by the acoustic model selecting means, The model selection means may determine whether the sentence that generates the speech has a predetermined sentence ending pattern or at least one of the cases where the sentence that generates the speech includes an interrogative. A phoneme at the end of a sentence, a predetermined number of mora at the end of the sentence, an accent phrase at the end of the sentence, or an acoustic model of the whole sentence instead of the acoustic dictionary And selecting from the question statements with the acoustic model of the serial question statements with acoustic dictionary.

  Further, in the speech synthesizer of the invention according to claim 2, in addition to the configuration of the invention of claim 1, the predetermined sentence ending pattern has a question ending character as a question mark.

  Moreover, in the speech synthesizer of the invention according to claim 3, in addition to the configuration of the invention according to claim 1 or 2, the pattern at the end of the predetermined sentence prompts a word that the sentence ending asks, a word or an action for seeking consent. Characterized by words.

In the speech synthesis program of the invention according to claim 4, a sentence for generating speech in a computer is decomposed into words to determine parts of speech, an accent type indicating the accent position is determined for each accent phrase, and the speech Acoustic model including at least a linguistic analysis step that determines the reading of a sentence that generates speech, a phonological model created from phoneme data obtained by analyzing speech into an acoustic parameter sequence, and a prosodic model created from fundamental frequency data analyzed from speech An acoustic model selection step for selecting the acoustic model from an acoustic dictionary that is an analysis of the language analysis step, and the phoneme constituting the acoustic model selected by the acoustic model selection step fruit on the basis of the model and the prosody model a voice generating step of generating a voice to the computer A speech synthesis program for case, the acoustic model selection step, endnotes statements for generating the sound when a predetermined end of the sentence pattern, or that contain interrogative sentences to produce the speech In the case of at least one of the above, the phoneme at the end of the sentence that generates the speech, the predetermined number of mora at the end of the sentence, the accent phrase at the end of the sentence, or the acoustic model of the whole sentence is not the acoustic dictionary, A phonological model for interrogative sentences made from the phonological data of a voice that utters a sentence whose ending tone is different from a plain sentence, such as a question sentence, a sentence requesting consent, a sentence that prompts an action, etc. Is a set of interrogative sentence acoustic models including at least interrogative sentence prosodic models created from the fundamental frequency data of speech uttered by different sentences. And selecting from.

  Further, in the speech synthesis program of the invention according to claim 5, in addition to the configuration of the invention of claim 4, the predetermined sentence ending pattern handled by the computer is characterized in that the ending character is a question mark.

  Further, in the speech synthesis program of the invention according to claim 6, in addition to the configuration of the invention of claim 4 or 5, the predetermined sentence ending pattern handled by the computer is a word asked by the sentence ending, a word seeking consent, It is characterized by words that encourage action.

  In the speech synthesizer of the invention according to claim 1, the acoustic dictionary storage means includes a phonological model created from phonological data obtained by analyzing speech into an acoustic parameter sequence, and a prosodic model created from fundamental frequency data analyzed by speech. An acoustic dictionary, which is a set of acoustic models including at least, is stored, and the acoustic dictionary storage means for question sentences utters sentences in which the tone at the end of the sentence is different from the plain text, such as question sentences, sentences for asking consent, sentences for prompting actions, etc. A phonological model for interrogative sentences including at least a phonological model for interrogative sentences made from phonetic data of speech and a prosodic model for interrogative sentences made from fundamental frequency data of speech uttered sentences whose tone at the end of the sentence is different from the plain text The speech analysis dictionary for question sentences, which is different from the acoustic dictionary, is stored, and the language analysis means decomposes the sentence for generating the speech into words to determine the part of speech, and for each accent phrase The accent type indicating the current position is determined, and the reading of the sentence that generates speech is determined. The acoustic model selection means selects an acoustic model from the acoustic dictionary based on the analysis result analyzed by the language analysis means, and The generation unit can generate speech based on the phoneme model and the prosody model constituting the acoustic model selected by the acoustic model selection unit. In addition, the acoustic model selection unit may generate a voice when the sentence ending sentence has a predetermined sentence ending pattern or when the sentence producing the voice includes an interrogative word. Select either the phoneme at the end of the sentence that generates the sentence, the specified number of mora at the end of the sentence, the accent phrase at the end of the sentence, or the acoustic model of the whole sentence from the acoustic model for question sentence in the question dictionary instead of the sound dictionary be able to. Therefore, since speech is generated using an acoustic model that is originally made from phonetic data of a voice that utters a sentence whose tone is different from a plain sentence, such as a question sentence, a sentence requesting consent, a sentence that prompts action, It is possible to synthesize speech without a natural feeling of strangeness.

  In addition, in the speech synthesizer of the invention according to claim 2, in addition to the effect of the invention according to claim 1, a predetermined sentence end pattern can be a sentence end character being a question mark. Therefore, when there is a question mark, it can be determined that the sentence is different in tone from the plain sentence, so it is possible to easily determine whether or not to use the question sentence acoustic dictionary.

  In addition, in the speech synthesizer of the invention according to claim 3, in addition to the effect of the invention according to claim 1 or 2, the predetermined pattern at the end of the sentence is a word for asking the end of the sentence, a word for seeking consent, or a word for prompting an act It can be. Therefore, if there are words that ask questions, words that require consent, or words that prompt action, at the end of the sentence, it can be determined that the sentence is different in tone from the plain text, so it is easy to create an acoustic dictionary for question sentences. It can be determined whether or not to use.

  In the speech synthesis program of the invention according to claim 4, a sentence for generating speech to a computer is decomposed into words to determine parts of speech, an accent type indicating the accent position is determined for each accent phrase, and speech is An acoustic model including at least a linguistic analysis step for determining the reading of a sentence to be generated, a phonological model created from phonological data obtained by analyzing speech into an acoustic parameter sequence, and a prosodic model created from fundamental frequency data analyzed from speech. An acoustic model selection step for selecting an acoustic model from the acoustic dictionary as a set based on the analysis result analyzed in the language analysis step, and a phonological model and a prosodic model constituting the acoustic model selected in the acoustic model selection step are also provided. And a sound generation step for generating sound. Then, in the acoustic model selection step, when the sentence end of the sentence that generates the sound is a predetermined sentence end pattern, or when the sentence that generates the sound includes an interrogative word, the sound is Sentence sentence phoneme, sentence end number of mora, sentence sentence accent phrase, or full sentence acoustic model is not an acoustic dictionary, but a question sentence, a sentence requesting consent, a sentence promoting action, etc. Phonetic model for interrogative sentences made from speech phonetic data uttered with a sentence whose tone is different from that of plain text and fundamental frequency data of speech uttered with a sentence whose tone is different from that of plain text It is possible to select from the question sentence acoustic model of the question sentence acoustic dictionary which is a set of question sentence acoustic models including at least the question sentence prosodic model. Therefore, since speech is generated using an acoustic model that is originally made from phonetic data of a voice that utters a sentence whose tone is different from a plain sentence, such as a question sentence, a sentence requesting consent, a sentence that prompts action, It is possible to synthesize speech without a natural feeling of strangeness.

  In addition, in the speech synthesis program of the invention according to claim 5, in addition to the effect of the invention of claim 4, a predetermined sentence end pattern may be a question end character. Therefore, when there is a question mark, it can be determined that the sentence is different in tone from the plain sentence, so it is possible to easily determine whether or not to use the question sentence acoustic dictionary.

  In addition, in the speech synthesis program of the invention according to claim 6, in addition to the effect of the invention of claim 4 or 5, the predetermined pattern at the end of the sentence is a word for asking the end of the sentence, a word for seeking consent, or a word for prompting an act It can be. Therefore, if there are words that ask questions, words that require consent, or words that prompt action, at the end of the sentence, it can be determined that the sentence is different in tone from the plain text, so it is easy to create an acoustic dictionary for question sentences. It can be determined whether or not to use.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings. First, the speech synthesizer 1 of the present embodiment will be described with reference to FIG. FIG. 1 is a block diagram showing an electrical configuration of the speech synthesizer 1. As shown in FIG. 1, the speech synthesizer 1 is provided with a CPU 2 that controls the speech synthesizer 1. The CPU 2 includes a keyboard 3, a RAM 4 that temporarily stores various data, an acoustic dictionary 50, A ROM 5 storing a sentence end pattern dictionary 55, an interrogative dictionary 56, and the like, a digital analog converter (DAC) 6, and a timing device 9 are connected. An amplifier (AMP) 7 is further connected to the DAC 6, and a speaker 8 is connected to the AMP 7.

  In addition, the RAM 4 is provided with various storage areas for storing variables and generation data used when performing speech synthesis processing. For example, the text storage area 41 stores text that is input from the keyboard 3 and is subjected to speech synthesis. The analysis result storage area 42 stores the result of analyzing the text stored in the text storage area 41 by the language analysis 11 (see FIG. 2).

  The msep sequence storage area 43 stores the msem sequence generated by the msep sequence generation 15 (see FIG. 2), and the pitch sequence storage area 44 generates the pitch sequence generation 16 (see FIG. 2). A pitch string is stored. The sound source signal storage area 45 stores the sound source signal generated by the sound source signal generation 17 (see FIG. 2), and the output sound waveform storage area 46 is generated by the MLSA filter 23 (see FIG. 2). The waveform of the output sound is stored.

  The acoustic model information storage area 61 stores an acoustic model for the text stored in the text storage area 41 for each phoneme, and the interrogative flag storage area 62 stores it in the text storage area 41. An interrogative flag for the entered text is stored for each phoneme.

  Next, with reference to FIG. 2, the functional configuration of the speech synthesis apparatus and speech synthesis program of the present embodiment will be described. FIG. 2 is a functional configuration diagram of the present embodiment. As shown in FIG. 2, first, language analysis 11 is performed on the text to be synthesized. In this language analysis 11, the input text is analyzed and its reading and accent type are output.

  For example, as shown in FIG. 3, it is assumed that a text “Is that you went to Kyoto?” Is stored in the text storage area 41. FIG. 3 is a schematic diagram of a text storage area 411 that is an example of the text storage area 41. First, a linguistic dictionary (not shown) with part-of-speech information, reading information, connection information, accent information, etc. is referred to, and morphological analysis is performed using the well-known longest match method. “Kyoto”, “ni”, “go”, “ta”, “?” Are analyzed, and the part of speech is further determined. Then, the connection information of the language dictionary is referred to and the compound words are put together, and “Speaking of which”, “in Kyoto”, “Did you go?”. In the morphological analysis, the accent position is also determined from the accent information in the language dictionary. When words are moved to a compound word, accent position change processing is also performed for words with accent position movement. Finally, the reading information of the language information is referred to, the character string is replaced with a katakana character string, and an analysis result “Soeva (4) | Kotoni (1) / Itta (0)?” Is output. As shown in FIG. 4, it is stored in the analysis result storage area 42. FIG. 4 is a schematic diagram of an analysis result storage area 421 that is an example of an analysis result obtained by performing the language analysis 11 on the information in the text storage area 411. The part-of-speech information for each word is also stored in the analysis result storage area 42. Here, “|” indicates an exhalation paragraph delimiter, () indicates an accent section delimiter, and the numbers in () indicate the accent positions of the accent section.

  Then, an acoustic model selection 12 is performed based on the accent type and the reading analyzed by the language analysis 11. Here, the reading is further broken down into phonemes. Then, an acoustic model is selected from the acoustic dictionary 50 for each phoneme. As shown in FIG. 2, the acoustic dictionary 50 is formed of a phoneme model 51 and a prosody model 52, and a combination of the phoneme model and the prosody model is referred to as an acoustic model. The phoneme model 51 includes a normal phoneme model 511 and a question sentence phoneme model 512, and the prosody model 52 includes a normal phoneme model 521 and a question sentence prosody model 522.

  The normal phoneme model 511 and the normal phoneme model 521 are a set of phoneme models and a set of prosody models made from recorded data obtained by recording a voice when a plain text is uttered. In addition, the interrogative sentence phoneme model 512 and interrogative sentence prosodic model 522 are different from plain sentences such as interrogative sentences, sentences that require consent, and sentences that prompt action, and voices when uttering sentences that increase the pitch at the end of the sentence. A set of phonological models and a set of prosodic models created from recorded recording data.

  In this embodiment, when “?” Is added at the end of a sentence, when the end of the sentence corresponds to a predetermined end-of-sentence pattern, or when a predetermined question word is present in the sentence, the phoneme model of the last accent phrase is changed Selects the phoneme model of the last accent phrase from the question sentence phoneme model 512 and selects the prosody model from the question sentence prosody model 522 to create a question sentence acoustic model.

  The normal phoneme model 511 includes “a, b, by, ch, cl, d, dy, e, f, fy, g, gy, h, hy, i, j, k, ky, m, my, n”. , N, ny, o, p, pau, py, r, ry, s, sh, t, ts, ty, u, w, y, z ”phoneme models for 38 phonemes are stored. For example, this phonological model is obtained by performing mel cepstrum analysis on natural speech. Each phoneme model is divided into frames (one frame is 10 ms), and a mel cepstrum coefficient and information about whether the frame is voiced or unvoiced are stored for each frame. Note that “pau” indicates a pause (expired paragraph break).

  The normal prosodic model 521 stores a prosodic model that becomes information for generating a pitch sequence in correspondence with information on the accent type and the mora position in the accent phrase for each phoneme. For example, this prosodic model is obtained by performing fundamental frequency analysis of natural speech, and pitch data for each frame is stored.

  In the present embodiment, information on the phoneme model and the acoustic model selected for each phoneme is integrated and called an acoustic model. For example, the acoustic model of the phoneme “a” is “(a)”. In this way, the phonemes are displayed with “()” added.

  In addition, the interrogative sentence phoneme model 512 and interrogative sentence prosodic model 522 are different from plain sentences such as interrogative sentences, sentences that require consent, and sentences that prompt action, and voices when uttering sentences that increase the pitch at the end of the sentence. A set of phonological models and prosodic models made from recorded data. The recorded data used to create various types of phonemes of pitch-up characters, accent types, and mora positions in accent phrases It is a utterance of the pattern text. Accordingly, the interrogative sentence acoustic model composed of the interrogative sentence phoneme model and the interrogative sentence prosodic model selected from the interrogative sentence phoneme model 512 and the interrogative sentence prosody model 522 is one acoustic model for one phoneme. Rather, there exist combinations of phoneme types, accent types of accent phrases to which the phonemes belong, and mora positions in the accent phrases. In this embodiment, in order to distinguish the acoustic model for question sentences from the acoustic model for normal use, “(aq)”, “(aq2)”, “(aq3)”, and the like are placed behind the phoneme type. Suppose that q, q1, q2, etc. are assigned and displayed.

  For example, in the example of “I went to Kyoto, so to speak”, it is decomposed into phonemes “s_o_o_i_e_b_a_pau_ky_o_o_t_o_n_i_i_cl_t_a” as shown in FIG. FIG. 5 is a schematic diagram of an acoustic model storage area 611 that stores an acoustic model of “I went to Kyoto. In the phoneme model selection 13, the mel cepstrum coefficient for each frame and information on whether the frame is voiced or unvoiced are selected from the normal phoneme model 511 for each phoneme. In the prosody model selection 14, the number of mora of the accent type and the accent phrase is arranged as “(5, 4), pau, (4, 1), (3, 0)”. For each phoneme, a prosodic model that is information for generating a pitch sequence is selected corresponding to the information on the mora position in the accent type or accent phrase. “(5,4), pau, (4,1), (3,0)” has a pose after an accent type 4 prosody model of 5 mora (beats), and then 4 mora This indicates that the accent type is 1 and the accent type is 0.

  The acoustic model is stored in the acoustic model information storage area 61 for each phoneme. FIG. 5 is a schematic diagram of an acoustic model information storage area 611 as an example. In the schematic diagram shown in FIG. 5, the accent type and the mora position are also described, but these pieces of information are information included in the acoustic model. The acoustic model of the phoneme “s” is “(s)”, the accent type is “4”, the mora position is the first, the acoustic model of the next phoneme “o” is “(o)”, and the accent type is The acoustic model of “4”, the mora position is the first, the next phoneme “o” is “(o)”, the accent type is “4”, the mora position is the second, the acoustic model of the next phoneme “i” Is “(i)”, the accent type is “4”, the mora position is third, the acoustic model of the next phoneme “e” is “(e)”, the accent type is “4”, and the mora position is The acoustic model of the fourth and next phoneme “b” is “(b)”, the accent type is “4”, the mora position is the fifth, and the acoustic model of the next phoneme “a” is “(a)”. Yes, the accent type is “4”, and the mora position is the fifth.

  The acoustic model of the next phoneme “ky” is “(ky)”, the accent type is “1”, the mora position is the first, and the acoustic model of the next phoneme “o” is “(o)”. , The accent type is “1”, the mora position is the first, the acoustic model of the next phoneme “o” is “(o)”, the accent type is “1”, the mora position is the second, the next phoneme “t” ”Is the acoustic model“ (t) ”, the accent type is“ 1 ”, the mora position is the third, the acoustic model of the next phoneme“ o ”is“ (o) ”, and the accent type is“ 1 ”. , The mora position is the third, the acoustic model of the next phoneme “n” is “(n)”, the accent type is “1”, the mora position is the fourth, the acoustic model of the next phoneme “i” is “( i) ", the accent type is" 1 ", and the mora position is the fourth. The acoustic model of the next phoneme “i” is “(i)”, the accent type is “0”, the mora position is the first, and the acoustic model of the next phoneme “cl” is “(cl)”. The accent type is “0”, the mora position is second, the acoustic model of the next phoneme “t” is “(t)”, the accent type is “0”, the mora position is third, and the next phoneme “a” The acoustic model is “(a)”, the accent type is “0”, and the mora position is the third.

  Also, for example, consider a sentence with “?” At the end of the sentence such as “Speaking of which, did you go to Kyoto?”. Also in this case, it is decomposed into phonemes of “s_o_o_i_e_b_a_pau_ky_o_o_t_o_n_i_i_cl_t_a”. FIG. 6 is a schematic diagram of an acoustic model storage area 612 in which an acoustic model of “Is that you went to Kyoto?” Is stored as an example of the acoustic model information storage area 61. Since “?” Is added at the end of the sentence, the phoneme model “i_cl_t_a” of the last accent phrase is selected not from the normal phoneme model 511 but from the interrogative sentence phoneme model 512, and the normal prosody model 521. The prosody model is selected from the prosody model for question sentence 522 instead.

  FIG. 6 is a schematic diagram of the acoustic model information storage area 612 in this case. The acoustic model of the phoneme “s” is “(s)”, the accent type is “4”, the mora position is the first, the acoustic model of the next phoneme “o” is “(o)”, and the accent type is The acoustic model of “4”, the mora position is the first, the next phoneme “o” is “(o)”, the accent type is “4”, the mora position is the second, the acoustic model of the next phoneme “i” Is “(i)”, the accent type is “4”, the mora position is third, the acoustic model of the next phoneme “e” is “(e)”, the accent type is “4”, and the mora position is The acoustic model of the fourth and next phoneme “b” is “(b)”, the accent type is “4”, the mora position is the fifth, and the acoustic model of the next phoneme “a” is “(a)”. Yes, the accent type is “4”, and the mora position is the fifth.

  The acoustic model of the next phoneme “ky” is “(ky)”, the accent type is “1”, the mora position is the first, and the acoustic model of the next phoneme “o” is “(o)”. , The accent type is “1”, the mora position is the first, the acoustic model of the next phoneme “o” is “(o)”, the accent type is “1”, the mora position is the second, the next phoneme “t” ”Is the acoustic model“ (t) ”, the accent type is“ 1 ”, the mora position is the third, the acoustic model of the next phoneme“ o ”is“ (o) ”, and the accent type is“ 1 ”. , The mora position is the third, the acoustic model of the next phoneme “n” is “(n)”, the accent type is “1”, the mora position is the fourth, the acoustic model of the next phoneme “i” is “( i) ", the accent type is" 1 ", and the mora position is the fourth. The acoustic model of the next phoneme “i” is “(iq)”, the accent type is “0”, the mora position is the first, and the acoustic model of the next phoneme “cl” is “(clq)”. , The accent type is “0”, the mora position is second, the acoustic model of the next phoneme “t” is “(tq)”, the accent type is “0”, the mora position is third, and the next phoneme “a” The acoustic model is “(aq)”, the accent type is “0”, and the mora position is the third.

  Here, referring to FIG. 7 to FIG. 10, the question sentence acoustic models “(iq)”, “(clq)”, “(tq) selected from the question sentence phoneme model 512 and the question sentence prosody model 522 are used. ) ”And“ (aq) ”as examples, the creation and selection of the question sentence acoustic model will be described. FIG. 7 is a schematic diagram 711 of phonemes, accent types, and mora positions in the accent phrase of the example sentence “I already bought?” When creating the acoustic model for question sentence “(aq)”. FIG. 9 is a schematic diagram 712 of phonemes, accent types, and mora positions in an accent phrase when creating an acoustic model for a question sentence of “(tq)”, FIG. 9 shows “(clq) FIG. 10 is a schematic diagram 713 of a phoneme, an accent type, and a mora position in an accent phrase of an example sentence “What did you say?” When creating an acoustic model for an interrogative sentence, and FIG. 10 is an interrogative sentence of “(iq)” FIG. 714 is a schematic diagram 714 of phonemes, accent types, and mora positions in an accent phrase of an example sentence “She's not there?

  For example, “(aq)” has “?” At the end of the sentence, the end-of-sentence phoneme is “a”, the accent type of the accent phrase to which the phoneme “a” belongs is type 0, and the accent phrase It is an acoustic model for question sentences whose inner mora position is “3”. This is formed by a phonological model or a prosodic model generated from data obtained by recording an utterance of an example sentence “I already bought?” That satisfies the same condition. In addition to this “(aq)”, the acoustic model for interrogative sentences of the phoneme “a” has “?” At the end of the sentence, the end-of-sentence phoneme is “a”, and the accent to which the phoneme “a” belongs. The accent type of the phrase is type 1 and the mora position in the accent phrase is “2”, or there is a question sentence “why” in the sentence, and the phoneme “a” is not a sentence end phoneme, Accent type of accent phrase to which phoneme “a” belongs is 0 type, and mora position in accent phrase is “1”, etc. Accent type of various patterns, mora position in accent phrase, sentence An acoustic model for a question sentence is created from the recorded data by a combination of types (when “?” Is attached, when the sentence end is a predetermined sentence end pattern, or when there is a question word in the sentence). When selecting the question sentence acoustic model, an acoustic model that satisfies the accent type of the accent phrase, the mora position in the accent phrase, and the like is selected.

  Similarly, “(tq)” has “?” At the end of the sentence, the phoneme of the first mora from the end of the sentence is “t”, and the accent type of the accent phrase to which the phoneme “t” belongs is 0 type. There is an acoustic model for question sentences in which the mora position in the accent phrase is “3”. This is formed by a phonological model or a prosodic model generated from data obtained by recording an utterance of an example sentence “Lend me?” That satisfies the same conditions. “(Clq)” has “?” At the end of the sentence, the phoneme of the second mora from the end of the sentence is “cl”, and the accent type of the accent phrase to which the phoneme “cl” belongs is 0 type. And an acoustic model for a question sentence in which the mora position in the accent phrase is “2”. This is formed by a phonological model or a prosodic model generated from data obtained by recording an utterance of an example sentence “What did you say?” That satisfies the same condition. “(Iq)” has “?” At the end of the sentence, the phoneme of the third mora from the end of the sentence is “i”, and the accent type of the accent phrase to which the phoneme “i” belongs is 0 type. And the acoustic model for question sentences in which the mora position in the accent phrase is “1”. This is formed by a phonological model or a prosodic model generated from data obtained by recording a utterance of an example sentence “she is not?” That satisfies the same condition.

  When the acoustic model is selected by the acoustic model selection 12 in this way, as shown in FIG. 2, the generated prosodic model sequence is connected by the pitch sequence generation 16 to generate a pitch sequence. However, the mora length is expanded and contracted to synchronize with the phoneme model according to the length of each phoneme in the phoneme model sequence at the time of connection. Next, based on the phoneme model selected by the phoneme model selection 13, the phoneme models of each phoneme are combined to generate a mel cepstrum sequence and a voiced / unvoiced information sequence (hereinafter referred to as a "mcep sequence"). 15).

  Then, the sound source signal generation 17 is performed based on the voiced / unvoiced information of the msep sequence generated by the msep sequence generation 15 and the pitch sequence generated by the pitch sequence generation 16. As for the sound source signal, a pulse train signal is generated in the voiced portion and a noise signal is generated in the unvoiced portion based on the pitch sequence. Then, the sound source signal is output as sound through the MLSA filter 23.

  Next, the sentence ending pattern dictionary 55 and the interrogative dictionary 56 will be described with reference to FIGS. 11 and 12. FIG. 11 is a schematic diagram of the sentence ending pattern dictionary 55, and FIG. 12 is a schematic diagram of the question word dictionary 56. These dictionaries are used when determining whether or not to change the acoustic model from an ordinary acoustic model to an acoustic model for question sentences.

  As shown in FIG. 11, the sentence ending pattern dictionary 55 stores word phrases at the end of sentences used in sentence-in-tone sentences such as a question sentence, a sentence requesting consent, and a sentence prompting an action together with part-of-speech information. . In the example shown in FIG. 11, the auxiliary verb “de”, the adverb “how”, the verb “shi” and the particle “tara”, the particle “te” and the verb “i” and the particle “te”, the particle "Kana" (including the case where "a" is added at the end), particle "ne" (including the case where "e" is added at the end), particle "yo" ("ぉ" at the end) ) (Including the case where the letter is attached), “no” (including the case where “ぉ” is suffixed), and “sa” (including the case where the suffix is suffixed with “a”) Yes, other end-of-sentence patterns are omitted.

  In addition, as shown in FIG. 12, the interrogative dictionary 56 stores interrogative words used in interrogative sentences, sentences that require consent, sentences that prompt actions, and the like. In the example shown in FIG. 12, “what”, “when”, “who”, “where”, “which”, “how”, “how much”, “how many”, “why”, “why”, etc. Other interrogatives are omitted.

  Next, with reference to the schematic diagram of FIG. 13 and the flowcharts of FIG. 14 and FIG. FIG. 13 is a schematic diagram of the question flag storage area 621. FIGS. 14 and 15 are flowcharts of processing performed in the acoustic model selection 12. FIG. 15 is a continuation of the flowchart shown in FIG.

  First, an interrogative flag storage area 621 that is an example of the interrogative flag storage area 62 of the RAM 4 will be described with reference to the schematic diagram of FIG. This interrogative flag storage area 621 is an example of an interrogative flag for a sentence “Is that you went to Kyoto?”, And interrogative flags are stored corresponding to phonemes. In the example illustrated in FIG. 13, the phoneme “s”, the phoneme “o”, the phoneme “o”, the phoneme “i”, the phoneme “e”, the phoneme “b”, the phoneme “a”, the phoneme “ky”, and the phoneme “o”. , Phoneme “o”, phoneme “t”, phoneme “o”, phoneme “n”, and phoneme “i” have interrogative flags “0”, phoneme “i”, phoneme “cl”, phoneme “t” , And the interrogative flag of phoneme “a” is “1”.

  First, the phonemes stored in the phoneme column of the acoustic model information storage area are set in the phoneme column of the interrogative flag storage area 62, all initial values “0” are set in the interrogative flag column, and initialization is performed. Performed (S1). Then, it is checked whether or not the symbol at the end of the sentence stored in the text storage area 41 is “?” And it corresponds to the sentence end pattern (S2). If the symbol at the end of the sentence is “?” And it corresponds to the sentence end pattern (S3: YES), “1” is set to the interrogative flag of the phoneme belonging to the last accent phrase (S4). If it does not correspond to the sentence end pattern (S3: NO), nothing is set in the question flag.

  Then, in the analysis result such as part of speech information stored in the analysis result storage area 42, there is one corresponding to the sentence end pattern dictionary 55, and it is checked whether or not it corresponds to the sentence end pattern (S5). If it corresponds to the sentence end pattern (S6: YES), “1” is set to the interrogative flag of the phoneme belonging to the last accent phrase (S7). If it does not correspond to the sentence end pattern (S6: NO), nothing is set in the question form flag.

  Then, in the analysis result such as the part of speech information stored in the analysis result storage area 42, it is checked whether or not there is a question word registered in the question word dictionary 56 (S8). If there is an interrogative (S9: YES), “1” is set to the interrogative flag of the phoneme belonging to the last accent phrase (S10). If there is no question word (S9: NO), nothing is set in the question flag.

  And the acoustic model of each phoneme is selected from the acoustic dictionary 50, referring to the question form flag set in S1 to S10 (S11 to S17). First, a pointer is placed on the first pointer (S11). Then, it is determined whether or not the interrogative flag of the phoneme is “1” (S12). If the question form flag is not “1” (S12: NO), there is no need to select from the question sentence phoneme model 512 and the question sentence prosody model 522, so the sound from the normal phoneme model 511 and the normal prosody model 521 is used. A model (phoneme model and prosody model) is selected (S15). Then, the pointer is advanced to the next phoneme (S16).

  If the question flag is “1” (S12: YES), it is determined whether the accent phrase to which the phoneme belongs is the last accent phrase in the sentence (S13). If it is not the last accent phrase (S13: NO), it is not necessary to select from the interrogative sentence phoneme model 512 and the interrogative sentence prosody model 522, so the acoustic model (phoneme) is selected from the normal phoneme model 511 and the normal prosody model 521. Model and prosody model) are selected (S15). Then, the pointer is advanced to the next phoneme (S16).

  If it is the last accent phrase (S13: YES), it is necessary to increase the pitch of the ending by changing the tone for the question sentence, so the acoustic model (phoneme model and phonological model) is derived from the question sentence phoneme model 512 and the question sentence prosody model 522. Prosody model) is selected (S14). Then, the pointer is advanced to the next phoneme (S16). If the processing for all phonemes has been completed (S17: YES), the acoustic model selection processing is completed. If the processing for all phonemes has not been completed (S17: NO), the process proceeds to S12. Returning, the process of selecting the acoustic model for the phoneme pointed by the pointer is performed (S12 to S16).

  For example, in the case of “Speaking of which, did you go to Kyoto?”, A pointer is placed on the first phoneme “s” in S11. As shown in FIG. 13, since the question flag of the phoneme “s” is “0” (S12: NO), the acoustic model “511” selected from the normal phoneme model 511 and the normal phoneme model 521 is selected. (S) "is stored in the acoustic model information storage area 61 (S15). Then, the pointer is advanced to the next phoneme “o” (S16). Since all the phonemes have not been processed yet (S17: NO), the process for the phoneme “o” indicated by the pointer is performed. Since the interrogation flag of this phoneme “o” is also “0” (S12: NO), the acoustic model “(s)” selected from the normal phoneme model 511 and the normal phoneme model 521 is stored in the acoustic model information. It is stored in the area 61 (S15). Then, the pointer is advanced to the next phoneme “o” (S16).

  In this way, the following phoneme “o”, phoneme “i”, phoneme “e”, phoneme “b”, phoneme “a” phoneme “ky”, phoneme “o”, phoneme “o”, phoneme “t”, The phoneme “o”, phoneme “n”, and phoneme “i” are processed. Since all of these phonemes have the question flag “0”, the sound from the normal phoneme model 511 and the normal phoneme model 521 is used as the sound. A model is selected.

  In the next phoneme “i”, the question-type flag is “1” (S12: YES), and is the last accent phrase (S13: YES). Therefore, the question phoneme model 512 and the question sentence prosody model 522 are used. An acoustic model (phoneme model and prosody model) is selected from (S14). Then, the pointer is advanced to the next phoneme “cl” (S16). Since processing for all phonemes has not been completed (S17: NO), the process returns to S12. The phoneme “cl” has an interrogative flag “1” (S12: YES), and is the last accent phrase (S13: YES). Therefore, the phoneme model 512 for the question sentence and the prosody model 522 for the question sentence are used as the acoustic model ( (Phoneme model and prosody model) are selected (S14).

  Similarly, the subsequent phoneme “t” and phoneme “a” also have the question flag “1” (S12: YES) and are the last accent phrase (S13: YES). Then, an acoustic model (phoneme model and prosody model) is selected from the question sentence prosody model 522 (S14). Since all phonemes have been processed (S17: YES), the processing ends.

  As described above, the interrogative sentence phoneme model 512 and the interrogative sentence prosodic model 522 can be used to create a sentence that has a different tone from that of a plain sentence, such as a question sentence, a sentence requesting consent, and a sentence prompting an action. Created in advance from the phonological data, and when the sentence ends with a predetermined pattern (if "?" Is attached, a predetermined questioning word, a word asking for consent or a word prompting an action), or when there is a questionable word in the sentence Selects the acoustic model (phonological model and prosodic model) from the interrogative phonological model 512 and interrogative sentence prosodic model 522, thereby approaching the tone at the end of the sentence such as an interrogative sentence, a sentence requesting consent, and a sentence prompting an action. Can output natural sound.

  It should be noted that the normal phoneme model 511 and the normal prosody model 521 stored in the acoustic dictionary 50 of the ROM 5 of the present embodiment correspond to the “acoustic dictionary storage means” and the question stored in the acoustic dictionary 50 of the ROM 5 The sentence phoneme model 512 and the question sentence prosody model 522 correspond to the “question sentence sound dictionary storage unit”. The CPU 2 that performs the processing of the language analysis 11 corresponds to “language analysis means”, the CPU 2 that performs the processing of the acoustic model selection 12 corresponds to “acoustic model selection means”, and generates the msec sequence generation 15, the pitch sequence generation 16, and the sound source signal. The CPU 2 that performs processing of the generation 17 and the MLSA filter 23 corresponds to “sound generation means”.

  Note that the speech synthesizer and the speech synthesis program of the present invention are not limited to the above-described embodiments, and it is needless to say that various modifications can be made without departing from the gist of the present invention.

  In the above embodiment, data for question sentences is used for phonological models and prosodic models among acoustic models. However, only phonological models are used for question sentences, or only prosodic models are used for question sentences. You may do it. In creating the model for the question sentence, in the above embodiment, the recording data is created in consideration of the type of phoneme, the accent type of the accent phrase to which the phoneme belongs, and the mora position in the accent phrase. I created data for the question sentence, but in addition to these, the length of the sentence, the length of the exhalation paragraph, the distance from the accent position of the accent phrase (how many phonemes are apart), the presence or absence of the exhalation paragraph in the accent phrase The data may be created taking into account the degree of dependency between accent phrases, part of speech, part of speech utilization, part of speech utilization. In addition, the situation of previous and subsequent phonemes such as the previous phoneme, the previous phoneme, the next phoneme, and the second phoneme may be considered. In addition, it is not necessary to consider the type of phoneme.

  Further, in the above embodiment, when the last accent phrase has a predetermined pattern (if “?” Is attached, a predetermined question word, a word for asking for consent, or a word prompting an action), When there is an interrogative, an acoustic model (phonological model and prosodic model) is selected from the interrogative sentence phoneme model 512 and interrogative sentence prosodic model 522, but the phoneme corresponding to the phoneme selected from the interrogative sentence acoustic model is selected. Is not necessarily the phoneme corresponding to the last accent phrase, and may be only the last phoneme or the phoneme corresponding to the last mora.

  The speech synthesizer and speech synthesis program of the present invention can be applied to a speech synthesizer and a speech synthesis program for outputting speech of a sentence having a tone different from that of a plain text.

2 is a block diagram showing an electrical configuration of the speech synthesizer 1. FIG. It is a functional block diagram of this Embodiment. 3 is a schematic diagram of a text storage area 411. FIG. FIG. 10 is a schematic diagram of an analysis result storage area 421. 5 is a schematic diagram of an acoustic model information storage area 611. FIG. 5 is a schematic diagram of an acoustic model information storage area 612. FIG. FIG. 7B is a schematic diagram 711 of phonemes, accent types, and mora positions in an accent phrase of an example sentence “I already bought?” When creating an acoustic model for question sentences of “(aq)”. FIG. 712 is a schematic diagram 712 of a phoneme, an accent type, and a mora position in an accent phrase of an example sentence “Lend me?” When creating the question sentence acoustic model of “(tq)”. FIG. 713 is a schematic diagram 713 of a phoneme, an accent type, and a mora position in an accent phrase of an example sentence “What did you say?” When creating an acoustic model for a question sentence of “(clq)”. FIG. 714 is a schematic diagram 714 of a phoneme, an accent type, and a mora position in an accent phrase of an example sentence “She is not?” When creating an acoustic model for question sentences of “(iq)”. It is a schematic diagram of the sentence end pattern dictionary. 3 is a schematic diagram of an interrogative dictionary 56. FIG. It is a schematic diagram of an interrogative flag storage area 621. 10 is a flowchart of processing performed in acoustic model selection 12; It is a flowchart shown in FIG.

Explanation of symbols

1 speech synthesizer 2 CPU
4 RAM
5 ROM
11 Language analysis 12 Acoustic model selection 13 Phonological model selection 14 Prosodic model selection 41 Text storage area 42 Analysis result storage area 50 Acoustic dictionary 51 Phonological model 52 Prosodic model 55 End-of-sentence pattern dictionary 56 Interrogative dictionary 61 Acoustic model information storage area 62 Question form Flag storage area 511 Normal phoneme model 512 Question phoneme model 521 Normal phone model 522 Question phone model

Claims (6)

Acoustic dictionary storage means for storing an acoustic dictionary that is a set of acoustic models including at least a phonological model created from phonological data obtained by analyzing speech into an acoustic parameter sequence and a prosodic model created from fundamental frequency data obtained by analyzing speech; ,
A phonological model for interrogative sentences made from the phonological data of a voice that utters a sentence whose ending tone is different from a plain sentence, such as a question sentence, a sentence requesting consent, a sentence that prompts an action, etc. Is a set of interrogative sentence acoustic models including at least interrogative sentence prosodic models created from the fundamental frequency data of speech uttered by different sentences, and stores interrogative sentence acoustic dictionaries different from the acoustic dictionary An acoustic dictionary storage means for sentences;
Language analysis means for decomposing speech generating sentences into words to determine part of speech, determining an accent type indicating the accent position for each accent phrase, and determining reading of the sentence generating the speech;
Acoustic model selection means for selecting the acoustic model from the acoustic dictionary based on the analysis result analyzed by the language analysis means;
Voice generation means for generating voice based on the phonological model and the prosodic model constituting the acoustic model selected by the acoustic model selection means,
The acoustic model selection means, in the case where the sentence ending sentence is a predetermined sentence ending pattern, or in the case where at least one of the sentence containing the voice includes a question word, The phoneme at the end of the sentence for generating the speech, the predetermined number of mora at the end of the sentence, the accent phrase at the end of the sentence, or the acoustic model of the whole sentence is used for the question sentence of the question sentence acoustic dictionary instead of the sound dictionary A speech synthesizer characterized by selecting from an acoustic model.   The speech synthesizer according to claim 1, wherein the predetermined sentence end pattern is a question mark character at the end of the sentence.   The speech synthesizer according to claim 1, wherein the predetermined sentence ending pattern is a word that the sentence ending asks, a word that asks for consent, or a word that prompts an action. A language analysis step for determining a part of speech by decomposing a sentence that generates speech, determining an accent type indicating an accent position for each accent phrase, and determining a reading of the sentence that generates the speech;
Analyzed by the language analysis step from an acoustic dictionary that is a set of acoustic models including at least a phonological model created from phonological data obtained by analyzing speech into an acoustic parameter sequence and a prosodic model created from fundamental frequency data analyzed from speech. An acoustic model selection step of selecting the acoustic model based on the analyzed result,
A speech synthesis program for causing a computer to execute a speech generation step for generating speech based on the phonological model and the prosodic model constituting the acoustic model selected by the acoustic model selection step,
In the acoustic model selection step, when the sentence end of the sentence that generates the sound is a pattern of a predetermined sentence end, or when the sentence that generates the sound includes a question word, The phoneme at the end of the sentence that generates the speech, the predetermined number of mora at the end of the sentence, the accent phrase at the end of the sentence, or the acoustic model of the whole sentence is not an acoustic dictionary, but a question sentence, a sentence for seeking consent, or an act The phonological model for interrogative sentences created from the phonological data of the speech uttered by a sentence whose tone is different from that of the plain sentence and the basic of the speech uttered by a sentence whose tone is different from that of the plain sentence A speech synthesis program selected from the question sentence acoustic model of the question sentence acoustic dictionary, which is a set of question sentence acoustic models including at least a question sentence prosody model created from frequency data Lamb.   5. The speech synthesis program according to claim 4, wherein the predetermined sentence end pattern is a question mark character at the end of the sentence.   6. The speech synthesis program according to claim 4, wherein the pattern at the end of the predetermined sentence is a word that the sentence end asks, a word that asks for consent, or a word that prompts an action.

Images