JP4841339B2 - Prosody correction device, speech synthesis device, prosody correction method, speech synthesis method, prosody correction program, and speech synthesis program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To eliminate the necessity of changing processing content and pregenerating various pattern data in accordance with an accent type of input text data and to eliminate the necessity of storing much data such as pattern data for rhythm correction in the generation of intonation for expressing problems in a rhythm correction device or a speech synthesis device. <P>SOLUTION: Out of pitches of respective phonetic codes preimparted to text data, the pitch (PitL) of a terminating phoneme is corrected to a first pitch (PitL1) and a second pitch (PitL2). Wherein the values of the first and second pitches are calculated by using the maximum value (PitP) of pitches in a preimparted pitch series. <P>COPYRIGHT: (C)2008,JPO&amp;INPIT

Description

  The present invention relates to prosody correction, and more particularly to a prosody correction device, a speech synthesizer, a prosody correction method, a speech synthesis method, a prosody correction program, and a speech synthesis for correcting a prosody obtained by analyzing input text data. Regarding the program.

  Techniques for outputting text data expressed as electronic data by voice have been studied for a long time, and in recent years, mobile phones equipped with a function of reading incoming e-mails with synthesized voice have been commercialized.

  As an example of a conventional speech synthesizer, word boundary detection, word segmentation, conversion to word phoneme symbol strings, word accent / sentence intonation (prosodic generation) are performed from an input character / symbol sequence. Non-Patent Document 1 discloses a speech synthesizer that generates a control signal necessary for driving a speech synthesizer based on phoneme symbol strings and prosodic information and based on previously stored speech data and rules. ing. The speech synthesis method disclosed in Non-Patent Document 1 is a known regular speech synthesis method itself.

As another example of a conventional speech synthesizer, a basic pitch pattern for generating a synthesized speech signal is generated by inputting language attribute and phoneme duration information generated by analyzing an input sentence. Select one correction pattern for correcting the basic pitch pattern from a correction pattern dictionary including a plurality of correction patterns generated in advance, and add the selected correction pattern with the end position corresponding to the basic pitch pattern. A pitch pattern generation method for generating a corrected pitch pattern by combining them is disclosed in Patent Document 1.
JP 2004-226505 A Co-authored by Shuzo Saito and Kazuo Nakata, “Basics of Speech Information Processing”, 1st Edition, Ohm Co., Ltd., November 30, 1981, p. 166-171

  The method disclosed in Non-Patent Document 1 for generating a prosody from an input character / symbol sequence to generate a synthesized speech generally generates a prosody based on a plain text. That is, with the technique disclosed in Non-Patent Document 1, it is difficult to effectively express the sentence end feature that expresses the intention of the speaker such as a question sentence with synthesized speech.

  A method of expressing a question sentence can be easily considered by correcting a pitch by adding a predetermined value (or multiplying by a predetermined magnification) to the terminal pitch (basic frequency) of the generated prosody. .

  However, in the method of adding (or multiplying) a predetermined value to the end pitch, for example, if the predetermined value is set so that a natural interrogative synthesized speech is obtained with a type 1 accent (accent type with an accent kernel in the first mora), for example. When the predetermined value is applied to the 0 type accent (the accent nucleus does not exist), there arises a problem that the ending pitch becomes abnormally high. Conversely, for example, when a predetermined value is set so that a natural interrogative synthesized speech is obtained with a 0 type accent, the predetermined value is set for a 1 type accent (or an accent type other than 0 type such as 2 type, 3 type). When applied, the ending pitch is not sufficiently high, and there is a problem that intonation that expresses a question cannot be obtained.

  In addition, the technique disclosed in the above-mentioned Patent Document 1 is selected from a plurality of correction patterns generated in advance and added to the basic pitch pattern by generating a pitch correction pattern corresponding to a question sentence or the like in advance. It is possible to obtain intonation that expresses a question, and it is also possible to cope with the difference in the accent type by generating a correction pattern in advance according to the accent type.

  However, it is necessary to generate a correction pattern according to various conditions in advance, and furthermore, since the generated correction pattern needs to be stored and held, a complicated operation is required, and a correction pattern is generated. There is a problem that the cost of the speech synthesizer is increased due to the mounting of the held memory.

  The present invention has been made in view of the above problems, and does not require processing contents to be changed according to the accent type of input text data, and it is also necessary to create various pattern data in advance. Prosody correction device, speech synthesizer, prosody correction method, speech synthesis method, prosody can generate intonation expressing a question without holding a large amount of data such as pattern data for prosody correction. An object of the present invention is to provide a correction program and a speech synthesis program.

The prosody correcting device according to the present invention uses the pitch peak extracting means for extracting the maximum value from the end of the pitch sequence given in advance to the phoneme symbol sequence, and using the maximum value of the extracted pitch, the phonological symbol A terminal phoneme pitch generating means for generating a terminal phoneme pitch which is a pitch to be given to a terminal phoneme in a sequence, and adding the generated terminal phoneme pitch to a terminal phoneme pitch in the phoneme symbol sequence, A terminal phoneme pitch correcting unit that corrects a pitch sequence previously assigned to the phoneme symbol sequence, and the terminal phoneme pitch generating unit preliminarily calculates the maximum value of the extracted pitch and the terminal phoneme in the phoneme symbol sequence. A reference pitch corresponding to a value between the assigned pitches is calculated, and a first pitch and a second pitch are generated as the terminal phoneme pitch. The first pitch is generated using a predetermined value representing a change amount of the pitch in the terminal phoneme of the phoneme symbol sequence, and is a value smaller than the reference pitch, and the second pitch is the predetermined value. The amount of change between the first pitch and the second pitch, which is generated using the reference pitch and is greater than the reference pitch, is calculated using the predetermined value, and the change between the first pitch and the second pitch. The quantity is a difference or ratio between the first pitch and the second pitch .

In the prosody correction device according to the present invention, it is preferable that the pitch peak extraction unit extracts a maximum value from an accent phrase at the end of a pitch sequence or an exhalation paragraph at the end of a sequence given to a phoneme symbol sequence. .

  In the prosody correction device according to the present invention, the terminal phoneme pitch generation means sets the first pitch to a value smaller than the maximum value of the extracted pitch, and sets the second pitch to the extracted pitch. A value larger than the maximum value is preferable.

Further, in the prosody correction device according to the present invention, the terminal phoneme pitch generation means sets the extracted maximum value as PitP, sets the pitch previously given to the terminal phoneme in the phoneme symbol sequence as PitL, and changes the pitch. When the predetermined value representing the quantity is PTDIFF and the values regarding the degree of correction of the pitch of the terminal phoneme in the phoneme symbol sequence are m, M, n, and N, these values are used to calculate the reference pitch. A PitS
PitS = PitL + (PitP−PitL) × m / M
And PDif which is the amount of change between the first pitch and the second pitch,
PDif = PTDIFF × n / N
It is preferable to calculate according to:

  In the prosody correction device according to the present invention, it is preferable that the values m, M, n, and N relating to the degree of correcting the pitch of the terminal phoneme are m = n and M = N.

  Further, the prosody correction device according to the present invention is a power correction means for correcting power given in advance to the terminal phoneme in the phoneme symbol sequence, or provided in advance to the terminal phoneme in the phoneme symbol sequence. It is preferable to further include a phoneme duration correction unit that corrects the duration.

  In the prosody correction device according to the present invention, it is preferable that the power correction unit corrects the power of the terminal phoneme by using power preliminarily given to the terminal phoneme.

  In the prosody correction device according to the present invention, the power correction unit corrects the power of the terminal phoneme, thereby providing a first power to be given to the terminal phoneme and a second power to be given to the terminal phoneme. And generating the second power based on a predetermined value representing an attenuation amount with respect to the power given in advance to the terminal phoneme, and smaller than the power given in advance to the terminal phoneme It is preferable to produce.

  In the prosody correction device according to the present invention, the phoneme duration correction unit may correct the phoneme duration of the terminal phoneme by using the phoneme duration previously given to the terminal phoneme. preferable.

  In the prosody correction device according to the present invention, the phoneme duration correction means corrects the phoneme duration of the terminal phoneme, thereby obtaining a ratio to the phoneme duration previously given to the terminal phoneme. It is preferable to generate a first phoneme duration and a second phoneme duration that are smaller than the phoneme duration given in advance to the terminal phoneme, based on the predetermined value that is represented.

The speech synthesizer according to the present invention analyzes the input text data, and obtains the phoneme symbol related to the text data and the sentence end prosody control information indicating the sentence end prosody control method, and the acquired Prosody generation means for generating prosody information including at least the pitch sequence of the text data based on phonemic symbols; pitch peak extraction means for extracting the maximum pitch value from the end of the pitch sequence generated by the prosody generation means; A terminal phoneme pitch generating means for generating a terminal phoneme pitch that is a pitch to be given to the terminal phoneme of the text data using the extracted maximum value of the text data, and the terminal phoneme pitch of the text data Prosody information of the text data generated by the prosody generation means is corrected by adding to the pitch of the terminal phoneme That by using the prosody correcting means and the acquired and phoneme symbol the corrected prosodic information, a synthesizing means for synthesizing a speech signal relating to the text data, the termination phoneme pitch generating means, pitch and the extracted A reference pitch corresponding to a value between a maximum value of the first phoneme and a pitch previously given to a terminal phoneme in the phoneme symbol sequence, and generating a first pitch and a second pitch as the terminal phoneme pitch, The first pitch is generated using a predetermined value representing a change amount of the pitch in the terminal phoneme of the phoneme symbol sequence, and is a value smaller than the reference pitch, and the second pitch is the predetermined value. The change amount of the first pitch and the second pitch, which is generated using the reference pitch and is larger than the reference pitch, is calculated using the predetermined value, and the first pitch and the second pitch are calculated. The amount of change in pitch is characterized in that the a difference or ratio of the first pitch and the second pitch.
In the speech synthesizer according to the present invention, it is preferable that the pitch peak extracting unit extracts a maximum value from an accent phrase at the end of a pitch sequence or a breath paragraph at the end of a pitch sequence previously assigned to a phoneme symbol sequence. .

  The speech synthesizer according to the present invention preferably further comprises speech unit storage means for storing a plurality of speech units representing speech information corresponding to one or more phoneme symbols.

Prosody correction method according to the present invention uses a step of extracting the maximum value of the pitch from the end of the sentence of previously assigned pitch sequence relative to the phoneme symbol sequence, the maximum value of the pitch and the extraction, the phoneme symbol sequence Generating a terminal phoneme pitch, which is a pitch to be given to the terminal phoneme of the phone, and a step of giving the terminal phoneme pitch to the terminal phoneme of the phoneme symbol sequence , and generating the terminal phoneme pitch The step of calculating a reference pitch corresponding to a value between the maximum value of the extracted pitch and a pitch given in advance to the terminal phoneme in the phoneme symbol sequence, and further, as the terminal phoneme pitch, 1 pitch and 2nd pitch are generated, and the 1st pitch is generated using a predetermined value indicating the amount of change in pitch in the terminal phoneme of the phoneme symbol sequence, and the reference The second pitch is generated using the predetermined value, and the amount of change between the first pitch and the second pitch is a value larger than the reference pitch. It is calculated using a predetermined value, and the amount of change between the first pitch and the second pitch is a difference or ratio between the first pitch and the second pitch.
Further, in the prosody correction method according to the present invention, the step of extracting the maximum value of the pitch extracts the maximum value from the accent phrase at the end of the pitch sequence or the exhalation paragraph at the end of the sequence given to the phoneme symbol sequence in advance. It is preferable to do.

  The prosody correction program according to the present invention is a prosody correction program for correcting the prosody of the terminal phoneme of the phoneme symbol sequence, and is a prosody correction program for realizing the above-mentioned prosody correction method on a computer. Features.

The speech synthesis method according to the present invention comprises analyzing the input text data to obtain a phoneme symbol related to the text data and a sentence end prosody control information representing a sentence end prosody control method, and the acquired phoneme symbol. On the basis of the step of generating prosodic information including at least the pitch sequence of the text data, extracting the maximum pitch value from the end of the generated pitch sequence , and using the extracted maximum pitch value, Correcting the generated prosodic information by generating a terminal phoneme pitch, which is a pitch to be added to the terminal phoneme of the text data, and adding the terminal phoneme pitch to the terminal phoneme of the text data; Using the acquired phoneme symbol and the corrected prosodic information, Synthesizing a speech signal, and the step of generating the terminal phoneme pitch corresponds to a value between a maximum value of the extracted pitch and a pitch previously given to the terminal phoneme in the phoneme symbol sequence A first pitch and a second pitch are generated as the terminal phoneme pitch, and the first pitch is a predetermined value representing a pitch change amount in the terminal phoneme of the phoneme symbol sequence. The second pitch is generated using the predetermined value and is greater than the reference pitch, and the second pitch is a value smaller than the reference pitch. The change amount of the pitch is calculated using the predetermined value, and the change amount of the first pitch and the second pitch is a difference or a ratio between the first pitch and the second pitch. That.
Further, in the speech synthesis method according to the present invention, the step of extracting the maximum value of the pitch extracts the maximum value from the accent phrase at the end of the pitch sequence or the exhalation paragraph at the end of the sequence given to the phoneme symbol sequence. It is preferable to do.

  A speech synthesis program according to the present invention is a speech synthesis program that synthesizes a speech signal based on input text data, and is a speech synthesis program for realizing the speech synthesis method as described above on a computer. It is characterized by that.

  According to the present invention, the terminal phoneme pitch (pitch to be given to the terminal phoneme of the phoneme symbol sequence) is generated using the maximum pitch value extracted from the pitch sequence previously given to the phoneme symbol series. The terminal phoneme pitch is added to the terminal phoneme of the phoneme symbol sequence, so that the pitch corresponding to the terminal phoneme in the previously assigned pitch sequence is corrected. That is, according to the present invention, the terminal phoneme pitch is generated using the maximum pitch value previously assigned to the phoneme symbol sequence, so that the processing content is changed according to the accent type of the input text data. It is not necessary to prepare various pattern data in advance, and it is not necessary to store a large amount of data such as pattern data for prosody correction.

Hereinafter, embodiments of the present invention will be described with reference to the drawings.
[Configuration of speech synthesizer]
First, the configuration of a speech synthesizer according to an embodiment of the present invention (a speech synthesizer including a prosody correcting device according to an embodiment of the present invention) will be described. FIG. 1 is a diagram illustrating a hardware configuration of the speech synthesizer.

  Referring to FIG. 1, speech synthesizer 100 includes a central processing unit 110 that controls the overall operation thereof. The central processing unit 110 performs correction processing for correcting prosody information related to text data, and processing for synthesizing a speech signal related to the text data using the phoneme symbols of the text data and the corrected prosodic information. A synthesis unit 111 is included.

  The speech synthesizer 100 further includes an antenna 120 that is used to transmit / receive information to / from an external device, a communication control unit 121 that controls transmission / reception of information using the antenna 120, various setting values and external A flash memory 130 for storing information such as an e-mail document received from the apparatus, an input unit 140 operated when information is input from the outside, a prosody information table storage unit 150 for storing a prosody information table to be described later, A ROM (Read Only Memory) 160 that stores various data such as a program executed in the synthesis unit 111 and a speech unit dictionary, a display unit 170 including a liquid crystal display, a speaker 181, and a speaker 181 An amplifier 180 for outputting is included. For example, the input unit 140 is configured by buttons provided as appropriate, or by a touch button displayed on the display unit 170 when the display unit 170 configures a touch panel.

FIG. 2 is a functional block diagram for explaining the configuration of the correction / combination unit 111 in FIG.
Referring to FIG. 2, correction / combination unit 111 receives text data received via antenna 120, text data stored in flash memory 130, or text data stored in ROM 160, and analyzes the text data. The prosody corresponding to the end portion of the text data is corrected based on the analysis result, and the prosody information including the corrected prosody related to the text data and the phonological symbol of the text data are combined. An audio signal is generated and output as audio data. The configuration of the correction / combination unit 111 will be described below.

  In the correction / synthesis unit 111, the text data is first input to the text analysis unit 1. The text data input here is, for example, composed of a character code series expressed as ASCII (American Standard Code for Information Interchange) code or SJIS (Shift Japan Industrial Standard) code that can be handled as digital data. Electronic text data. The text analysis unit 1 performs analysis including morphological analysis and syntactic analysis on the input text data, and includes phonemic symbol strings representing a series of phonemic symbols representing how to read, accent type, part of speech, number of mora, dependency, and the like. Language attribute information representing a language attribute and sentence end prosody control information representing a method for controlling a sentence end prosody such as a question are generated.

  The speech segment dictionary (stored in the ROM 160 in FIG. 1) stores speech waveform information necessary for synthesis for each phoneme. The language attribute information described above also includes information for specifying a phoneme corresponding to each phoneme symbol among the phonemes stored in the phoneme dictionary. The dictionary used for speech synthesis generation does not necessarily store information such as speech waveforms necessary for synthesis for each phoneme, but stores such information for each speech unit defined in other units. May be. For example, syllable (CV), VC (vowel-consonant), VCV (vowel center-consonant-vowel center) unit, etc. may be used. In addition, the dictionary used for generating the synthesized speech may store not only speech waveform information necessary for synthesis but also speech unit information defined in another manner. For example, it may be a parameter (for example, sound source and spectrum) for generating a speech waveform necessary for synthesis.

  The text analysis unit 1 is connected to the prosody generation unit 2 and the synthesis unit 4, and outputs the generated phoneme symbol string and language attribute information to the prosody generation unit 2 and the synthesis unit 4. The text analysis unit 1 is connected to the prosody correction unit 3 and outputs the generated sentence end prosody control information to the prosody correction unit 3. The text analysis unit 1 is configured by a dedicated LSI (Large Scale Integration), for example.

  The prosody generation unit 2 includes a power generation unit 21, a phoneme duration generation unit 22, and a pitch generation unit 23.

  The power generation unit 21 receives the phoneme symbol string and the language attribute information from the text analysis unit 1. Then, based on the phoneme symbol string and the language attribute information, power information (information used when determining the amplitude of the waveform) corresponding to each phoneme symbol in the input phoneme symbol string is generated. The power generation unit 21 is connected to the prosody correction unit 3 and outputs the generated power information to the prosody correction unit 3. The power information generated by the power generation unit 21 and corresponding to each phoneme symbol is hereinafter referred to as “standard power information”.

  The phoneme duration generation unit 22 receives the phoneme symbol string and language attribute information from the text analysis unit 1, and based on the phoneme symbol string and language attribute information, the duration corresponding to each phoneme symbol in the phoneme symbol string The phoneme duration information indicating the length is generated. The phoneme duration generation unit 22 is connected to the pitch generation unit 23 and the prosody correction unit 3 and outputs the generated phoneme duration length information to the pitch generation unit 23 and the prosody correction unit 3. The duration information corresponding to each phoneme symbol generated by the language duration generator 22 is hereinafter referred to as “standard duration information”.

  The pitch generation unit 23 receives the phoneme symbol string and the language attribute information from the text analysis unit 1, and also receives the phoneme duration length information from the phoneme duration generation unit 22, and based on these, the phoneme symbol sequence Pitch information representing the pitch (fundamental frequency F0) corresponding to each phoneme symbol is generated. The pitch generation unit 23 is connected to the prosody correction unit 3 and outputs the generated pitch information to the prosody correction unit 3. The pitch information corresponding to each phoneme symbol generated by the pitch generator 23 is hereinafter referred to as “standard pitch information”.

The prosody generation unit 2 is configured by a dedicated LSI, for example.
The prosody correction unit 3 includes a power correction unit 31, a phoneme duration correction unit 32, and a pitch correction unit 33.

  The prosodic correction unit 3 receives the end-of-sentence prosody control information from the text analysis unit 1, and switches whether to correct the prosody based on the input end-of-sentence prosody control information.

  The power correction unit 31 receives the standard power information from the power generation unit 21 and corrects a part of the input standard power information. Details of the correction processing will be described later. The power correction unit 31 is connected to the synthesis unit 4 and outputs the corrected power information to the synthesis unit 4.

  The phoneme duration correction unit 32 receives the standard phoneme duration information from the phoneme duration generator 22 and corrects part of the input standard phoneme duration information. Details of the correction processing will be described later. The phoneme duration correction unit 32 is connected to the synthesis unit 4 and outputs the corrected phoneme duration information to the synthesis unit 4.

  The pitch correction unit 33 receives the standard pitch information from the pitch generation unit 23 and corrects a part of the input standard pitch information. Details of the correction processing will be described later. The pitch correction unit 33 is connected to the synthesis unit 4 and outputs corrected pitch information to the synthesis unit 4.

The prosody correction unit 3 is configured by a dedicated LSI, for example.
The synthesizing unit 4 receives the phoneme symbol string and the language attribute information from the text analysis unit 1, and receives the correction power information, the corrected phoneme duration information, and the correction pitch information from the prosody correction unit 3.

  Then, the synthesis unit 4 selects a speech segment based on information such as the input phonemic symbol string, and based on the input prosodic information (corrected power information, corrected phoneme duration length information, corrected pitch information). An audio signal is synthesized by connecting the selected speech segments. A method for synthesizing an audio signal by connecting audio segments is detailed in, for example, Japanese Patent Application Laid-Open No. 2001-109500. Note that the speech signal synthesized based on the corrected power information, the corrected phoneme duration information, and the corrected pitch information is hereinafter referred to as “corrected synthesized speech”.

  The synthesizer 4 outputs the synthesized voice signal to the digital amplifier (amplifier 180), whereby the voice synthesizer 100 outputs a voice corresponding to the input text data from the speaker 181. The synthesizing unit 4 may be configured to output the synthesized audio signal to a D / A (digital-analog) converter, a speaker, headphones, or the like (not shown). The combining unit 4 is configured by a dedicated LSI, for example.

  In the present embodiment, the case where the prosody correction unit 3 is one of the components of the speech synthesizer 100 has been described. However, the prosody correction unit 3 corrects input power information, phoneme duration time information, and pitch information. Alternatively, the power information, the phoneme duration information, and the pitch information after correction may be used as an independent device that outputs the information.

  For example, a voice waveform stored in a non-volatile storage medium such as a ROM or a communication path such as a LAN (Local Area Network) or a microphone and input to a storage medium such as a RAM (Random Access Memory). The temporarily stored speech waveform is decomposed into an acoustic feature parameter sequence and a sound source parameter sequence by a known speech analysis technique, and a known speech recognition process is performed on the acoustic feature parameter sequence obtained by the decomposition. Thus, a phoneme symbol string and phoneme boundary information are generated, and standard power information, standard phoneme duration time information, and standard pitch information are generated using the generated phoneme boundary information and the sound source parameter sequence. Then, the generated standard power information, standard phoneme duration time information, and standard pitch information are input to the prosody correction unit 3 and corrected power information, corrected prosody duration time information, correction pitch information output from the prosody correction unit 3 Using the acoustic feature parameter series, the sound source parameter series, and the phoneme symbol string obtained by the above-described decomposition, a known synthesized speech process can be performed to generate a corrected synthesized speech.

  In the above description, the case where the text analysis unit 1 is included as a constituent element has been described. For example, when the input text data does not include kana-kana mixed sentences and is known to be composed of katakana notation and control symbols such as accent symbols and end-of-sentence symbols, the text analysis unit 1 performs morphological analysis and syntax analysis processing. Is used to generate phonetic symbol strings that represent phoneme symbol sequences that represent reading, language attribute information that represents language attributes such as accent type and number of mora, and sentence end prosody control information that represents control methods of sentence end prosody such as questions You may make it do.

  In addition, the speech synthesizer can be configured such that phoneme symbol strings, language attribute information, and sentence end prosody control information are input. In this case, the text analysis unit 1 is not required in the speech synthesizer.

  In addition, some or all of the text analysis unit 1, prosody generation unit 2, prosody correction unit 3, and synthesis unit 4 described above are realized by a general computer such as a personal computer or a microprocessor, not a dedicated LSI. Also good. Further, if the central processing unit 110 includes a microprocessor, it may be realized by the microprocessor. In this case, for example, speech synthesis processing or prosody correction processing described later may be described as a program for causing a computer to execute. In this case, the program is recorded and distributed on a recording medium such as a CD-ROM (Compact Disc-Read Only Memory), read into a RAM provided in the computer by a CD-ROM drive, and the like (CPU). (Unit: central processing unit). In addition to the CD-ROM, the recording medium on which the program is recorded is, for example, a flexible disk, a cassette tape, a hard disk, an optical disk, an MO (Magneto-Optical disc), an MD (Mini Disc (registered trademark)), a DVD (Digital Versatile). Disc), an IC card (including a memory card), an optical card, a mask ROM, an EPROM, an EEPROM, a semiconductor memory such as a flash ROM, or a medium for recording a program fixedly. Furthermore, it may be downloaded from another device via a network such as the Internet.

  Further, part or all of the text analysis unit 1, the prosody generation unit 2, the prosody correction unit 3, and the synthesis unit 4 described above may be realized by combining a dedicated LSI with a computer or a microprocessor. A part or the whole may be configured as one dedicated LSI.

  Hereinafter, the parts described as an example in which the text analysis unit 1, the prosody generation unit 2, the prosody correction unit 3, and the synthesis unit 4 are configured by a dedicated LSI can be realized by various methods as described above. Can do.

[Process for generating corrected synthesized speech]
Next, specific processing contents executed in the speech synthesizer 100 will be described.

  FIG. 3 is a flowchart of a process (speech synthesis process) for generating a corrected synthesized speech executed in the speech synthesizer 100.

  Referring to FIG. 3, when text data is input to text analysis unit 1, text analysis unit 1 performs analysis including morphological analysis and syntax analysis on the input text data, and indicates phonemes representing how to read the text data. A phoneme symbol string representing a sequence of symbols is generated (step S101), and language attribute information representing language attributes such as accent type, part of speech, number of mora, and dependency is generated (step S102), and a method for controlling sentence end prosody such as questions Is generated (step S103).

  Here, the phoneme symbol string is “s” “o” “t” “o” “d” “e” “a” “s” “o” “ b ”“ u ”. The process of generating the phoneme symbol string may be generated by a single process for the entire input text data, or may be generated by a plurality of processes by dividing the input text data into parts. Note that the input text data may be another expression that can be expressed, such as “Sotode Asob?”, And is not limited to the above. Also, instead of the input text data, phoneme symbol strings “s” “o” “t” “o” “d” “e” “a” “s” “o” “b” “u” are directly input. You may do it. Regarding the expression of phoneme symbol strings, “s”, “o”, “t”, “o”, “d”, “e”, “a”, “s”, “o”, “b”, and “u” are examples, and the characteristics of the sound are shown. It goes without saying that other expressions (capital letters such as “S”, “O”, “T”, “O”, “D”...) May be used as long as the symbols are associated with the units shown.

  For example, when the input text data is “play outside”, the end-of-sentence prosody control information detects “?” Symbol to set “1” as a question flag to a storage device such as a RAM (not shown). Set. If no “?” Symbol is detected, “0” is set as the question flag.

  Next, the power generation unit 21 receives the phoneme symbol string (generated in step S101) and the language attribute information (generated in step S102), and uses them in the phoneme symbol string. Power information (information used when determining the amplitude of the waveform) corresponding to each phoneme symbol is generated (step S104).

  Here, in the generation of power information, for example, a power information table is created in advance using a known method (such as a known multivariate analysis), and the power information table is stored in a storage device (not shown) such as a ROM or a RAM. The power information corresponding to the phoneme symbol string and the language attribute information may be generated by referring to the created power information table. The power information generated here is the standard power information described above.

  Next, the phoneme duration generation unit 22 receives the phoneme symbol string (generated in step S101) and the language attribute information (generated in step S102), and uses them, Phoneme duration information representing the duration length corresponding to each phoneme symbol in the phoneme symbol string is generated (step S105).

  Here, in the generation of phoneme duration information, for example, a phoneme duration information table is created in advance using a known method (well-known multivariate analysis or the like), and the table is stored in a storage device such as a ROM or RAM ( It may be stored in advance (not shown) and the phoneme duration information corresponding to the phoneme symbol string and the language attribute information may be generated by referring to the created phoneme duration information table. Further, the phoneme duration information generated here is the standard phoneme duration information described above.

  Next, the pitch generator 23 generates a phoneme symbol string (generated in step S101), language attribute information (generated in step S102), and phoneme duration information (generated in step S105). ) Is used to generate pitch information representing the pitch corresponding to each phoneme symbol in the phoneme symbol string (step S106).

  Here, the generation of the pitch information may be generated by using a known method such as pitch generation based on the Fujisaki model. The pitch information generated here is the standard pitch information described above.

  In the present embodiment, pitch control based on a point pitch that assigns pitch information to the vowel center is used. A specific pitch value assigned between the point pitches is calculated by, for example, linear interpolation.

  Next, the prosody correction unit 3 refers to the sentence end prosody control information generated in step S103 to determine whether the sentence is a question sentence. If it is determined to be a question sentence, the process proceeds to step S108. If so, the process proceeds to step S111 (step S107). In step S107, the prosody correction unit 3 refers to, for example, the setting value of the question flag stored in a storage device (not shown) such as a RAM. If the question flag is “1”, the process proceeds to step S108. To perform prosody correction processing (a generic name of step S108 to step S110 described later, or at least one of them). If the question flag is “0”, the processing proceeds to step S111 without performing prosody correction processing. . Here, when prosody correction processing is not performed, the prosody correction unit 3 directly uses the standard prosody information (standard power information, standard phoneme duration information, standard pitch information) generated in steps S104 to S106 as it is. The standard prosody information generated by the prosody generation unit 2 can be converted into a prosody by controlling the data flow with a control unit (not shown) provided in the speech synthesizer. You may comprise so that it may output directly to the synthetic | combination part 4 without going through the correction | amendment part 3. FIG.

  In step S108, the power correction unit 31 receives the power information generated in step S104 and corrects a part of the input standard power information (step S108). Details of the correction process will be described later.

  Next, the phoneme duration correction unit 32 receives the phoneme duration information generated in step S105 and corrects a part of the input standard phoneme duration information (step S109). Details of the correction process will be described later.

  Next, the pitch correction unit 33 receives the pitch information generated in step S106, thereby correcting a part of the input standard pitch information (step S110). Details of the correction process will be described later.

  In step S111, the synthesizing unit 4 includes the phoneme symbol string generated in step S101, the language attribute information generated in step S102, and the corrected prosodic information (corrected power information and corrected phoneme generated in steps S108 to S110). Continuous time length information, correction pitch information) or standard prosody information generated in steps S104 to S106, and a plurality of speech segments stored in advance in a storage device (not shown) such as a ROM or RAM. From the information, speech unit information corresponding to the phoneme symbol string is selected.

  The synthesizing unit 4 synthesizes the speech signal by connecting the speech unit information based on the input corrected prosody information (or standard prosody information) using the speech unit information selected in step S111. (Step S112).

  In the above description, power information, phoneme duration information, and pitch information are corrected in steps S108 to S110. However, one feature of the present invention is the configuration of the pitch correction unit 33 or details of processing (details). Is described later). That is, the prosody correction device and the speech synthesizer according to the present invention need only include at least the pitch correction unit 33. For this reason, part or all of the power correction processing in step S108 and the phoneme duration correction processing in step S109 may be omitted. In this case, the prosody correction unit 3 may be configured by excluding part or all of the power correction unit 31 and the phoneme duration correction unit 32.

[Configuration of power correction unit]
Next, a detailed configuration of the power correction unit 31 in the speech synthesizer 100 will be described.

FIG. 4 is a functional block diagram showing the configuration of the power correction unit 31 in FIG.
The power correction unit 31 includes a terminal phoneme power generation unit 311 and a terminal phoneme power setting unit 312. These are configured by, for example, dedicated LSIs, respectively, and all receive standard power information from the power generation unit 21. Then, the terminal phoneme power generation unit 311 generates a correction value of the power of the terminal phoneme using the input standard power information, and the terminal phoneme power setting unit 312 generates the power information of the terminal phoneme in the standard power information. By replacing (setting) with the correction value generated by the terminal phoneme power generation unit 311, correction power information is generated and output to the synthesis unit 4.

  The terminal phoneme is the terminal part of the phoneme symbol string. Specifically, in the above phoneme symbol string (“s” “o” “t” “o” “d” “e” “a” “s” “o” “b” “u”), “u” Is the terminal phoneme.

[Power correction processing]
Next, an example of specific processing contents of the power correction unit 31 will be described with further reference to FIGS. FIG. 5 is a flowchart of the subroutine of the power correction process (S108) in FIG. FIG. 6 is a diagram schematically showing an example of standard prosody information (standard power information, standard phoneme duration information, and standard pitch information), and FIG. 7 shows information shown in FIG. It is a figure which shows typically an example of corresponding correction prosody information (correction power information, correction phoneme duration information, and correction pitch information).

  FIG. 6 and FIG. 7 are described in a graph format. Table 1 (FIG. 6) and Table 2 (FIG. 7) show the data used as the basis of each graph.

  In FIG. 6 and FIG. 7, the pitch of each phoneme symbol is indicated by “●”, and the power is indicated by “◯”. In FIGS. 6 and 7, pitch and power are defined on the vertical axis, time is defined on the horizontal axis, and phoneme durations for each phoneme symbol are shown separated by vertical broken lines.

  As can be understood from FIGS. 6 and 7, in the present embodiment, power information, phoneme duration information, pitch information is obtained for the terminal phoneme of the phoneme symbol string by correcting the prosodic information. Are set to two. In this specification, the power information corresponding to the terminal phoneme in the corrected prosodic information, the phoneme duration information, the pitch information corresponding to the previously pronounced speech, Terminal phoneme first half duration, terminal phoneme first half pitch (or “terminal phoneme first pitch”), and the one corresponding to the sound that is pronounced later, terminal phoneme second half power, terminal phoneme second half duration, terminal phoneme This is called the latter half pitch (or “end phoneme second pitch”).

  Referring mainly to FIG. 5, in the power correction process, first, in step S1081, the terminal phoneme power generation unit 311 acquires the power of the terminal phoneme (PowL) from the standard power information.

  Next, the terminal phoneme power generation unit 311 calculates the terminal phoneme first half power (PowL1) in step S1082, and the terminal phoneme second power (PowL2) in step S1083. The terminal first phoneme power is the power of the terminal phoneme in the standard power information, and the terminal phoneme second power is a value obtained by subtracting a predetermined value (PWDOWN) from the power of the terminal phoneme in the standard power information.

  Then, the terminal phoneme power setting unit 312 replaces the power of the terminal phoneme with the power of the first half of the terminal phoneme and the power of the second half of the terminal phoneme with respect to the standard power information (S1084, S1085), and returns the processing to FIG.

  According to the power correction processing described with reference to FIG. 5, the power of the terminal phoneme (PowL) in the standard prosodic information is replaced with the power of the first half of the terminal phoneme (PowL1) and the power of the second half of the terminal phoneme (PowL2) in the corrected prosodic information. It is done.

  As can be understood from FIGS. 6, 7, Table 1, and Table 2, in this embodiment, PowL = PowL1 = 59 (dB) and PowL2 = 50 as specific examples of these values. (DB) is shown. That is, PWDOWN = 9. Note that the value of PWDOWN may be changed as appropriate by the user.

[Configuration of phoneme duration correction unit]
Next, the detailed configuration of the phoneme duration correction unit 32 in the speech synthesizer 100 will be described.

FIG. 8 is a functional block diagram showing the configuration of the phoneme duration correction unit 32 of FIG.
The phoneme duration correction unit 32 includes a terminal phoneme duration generation unit 321 and a terminal phoneme duration setting unit 322. These are configured by dedicated LSIs, for example, and all receive standard phoneme duration information from the terminal phoneme duration duration generator 22. Then, the terminal phoneme duration duration generation unit 321 generates a correction value for the phoneme duration length of the terminal phoneme using the input standard phoneme duration length information, and the terminal phoneme duration setting unit 322 By replacing (setting) the phoneme duration information of the terminal phoneme in the duration length information with the correction value generated by the terminal phoneme duration generation unit 321, the corrected phoneme duration information is generated, and the synthesis unit 4 is output.

[Phonological duration length correction]
Next, an example of specific processing contents of the phoneme duration correction unit 32 will be described with further reference to FIGS. 5, 6, and 9. FIG. 9 is a flowchart of a subroutine of the phoneme duration correction process (S109) of FIG.

  Referring mainly to FIG. 9, in the phoneme duration correction process, first, in step S1091, the terminal phoneme duration generation unit 321 calculates the phoneme duration (DurL) of the terminal phoneme from the standard phoneme duration information. get.

  Next, the terminal phoneme duration duration generator 321 calculates the terminal phoneme first half duration (DurL1) in step S1092, and calculates the terminal phoneme second half duration (DurL2) in step S1093. The first phoneme half power is a product of the phoneme duration length (DurL) of the terminal phoneme in the standard phoneme duration information and a predetermined first value (DRWGT1). It is the product of the phoneme duration (DurL) of the terminal phoneme in the standard phoneme duration information and a predetermined second value (DRWGT2). In this embodiment, 0.8 is adopted as an example of the first value (DRWGT1) and 0.5 is adopted as an example of the second value (DRWGT2). It is not limited and may be appropriately changed by the user.

  Then, the terminal phoneme duration setting unit 322 replaces the phoneme duration of the terminal phoneme with the first phoneme last duration and the last phoneme last duration for the standard phoneme duration information (S1094). S1095) The processing is returned to FIG.

  According to the phoneme duration correction process described with reference to FIG. 9, the phoneme duration (DurL) of the terminal phoneme in the standard prosodic information is the first phoneme duration (DurL1) and the terminal phoneme duration in the corrected prosodic information. The phonological second half duration (DurL2) is replaced.

  As can be understood from FIGS. 6, 7, Table 1 and Table 2, in the present embodiment, as specific examples of these values, DurL = 94 (msec), DurL1 = 75 (msec), And DurL2 = 47 (msec) is shown.

[Configuration of pitch correction unit]
Next, a detailed configuration of the pitch correction unit 33 in the speech synthesizer 100 will be described.

FIG. 10 is a functional block diagram showing the configuration of the pitch correction unit 33.
Referring to FIG. 10, pitch peak extraction unit 331 receives standard pitch information generated by pitch generation unit 21. The pitch peak extraction unit 331 extracts a peak pitch (maximum pitch value) from the pitch series included in the standard pitch information. Note that the range of the pitch sequence from which the peak pitch is extracted may be an accent phrase at the end of the sentence or an exhalation paragraph at the end of the sentence. The pitch peak extraction unit 331 is connected to the terminal phoneme pitch generation unit 332 and outputs the extracted peak pitch to the terminal phoneme pitch generation unit 332. Note that the pitch peak extraction unit 331 is configured by a dedicated LSI, for example.

  The terminal phoneme pitch generation unit 332 receives standard pitch information (generated by the pitch generation unit 21) and a peak pitch (extracted by the pitch peak extraction unit 331). The terminal phoneme pitch generation unit 332 calculates a reference pitch using the input peak pitch, calculates a pitch change amount, and uses the reference pitch and the pitch change amount to allocate two pitches ( A terminal phoneme first pitch and a terminal phoneme second pitch) are generated. Detailed processing contents for generating these two pitches will be described later. The terminal phoneme pitch generation unit 332 is connected to the terminal phoneme pitch setting unit 333, and outputs the generated terminal phoneme first pitch and terminal phoneme second pitch to the terminal phoneme pitch setting unit 333. The terminal phoneme pitch generation unit 332 is configured by a dedicated LSI, for example.

  The terminal phoneme pitch setting unit 333 receives standard pitch information (generated by the pitch generation unit 21), and the terminal phoneme first pitch and the terminal phoneme second pitch (generated by the terminal phoneme pitch generation unit 332). Is done. The terminal phoneme pitch setting unit 333 does not modify any pitch other than the pitch assigned to the terminal phoneme (the last pitch in the time direction) in the pitch series included in the input standard pitch information, and assigns it to the terminal phoneme. The standard pitch information is corrected by replacing the input pitch with the input terminal phoneme first pitch and terminal phoneme second pitch. In other words, the corrected pitch information is generated by changing the terminal phoneme pitch from the one input value to the two generated values. The terminal phoneme pitch setting unit 333 is connected to the synthesis unit 4 and outputs correction pitch information including the first terminal phoneme first pitch and the second terminal phoneme pitch to the synthesis unit 4. The terminal phoneme pitch setting unit 333 is configured by a dedicated LSI, for example.

  Note that some or all of the above-described pitch peak extraction unit 331, termination phoneme pitch generation unit 332, and termination phoneme pitch setting unit 333 are realized not by a dedicated LSI but by a general computer such as a personal computer or a microprocessor. May be. In this case, for example, a pitch correction process described later may be described as a program for causing a computer to execute.

[Pitch correction processing]
Next, an example of specific processing contents of the pitch correction unit 33 will be described with further reference to FIGS. 5, 6, and 11. FIG. 11 is a flowchart of the subroutine of the pitch correction process (S110) of FIG.

  Referring to FIG. 11, in the pitch correction process, first, pitch peak extraction section 331 extracts a peak pitch (PitP) from the pitch series in the standard pitch information (step S1101).

  Next, the terminal phoneme pitch generation unit 332 acquires the pitch (PitL) assigned to the terminal phoneme from the standard pitch information (step S1102), and uses the peak pitch (PitP) extracted in step S1101 to end the phoneme pitch. A reference pitch (PitS) used as a reference value in the generation is calculated by the equation (1) (step S1103).

PitS = PitP (1)
Next, the terminal phoneme pitch generation unit 332 uses a predetermined pitch change amount (PTDIFF) stored in advance in a storage device such as a ROM or a RAM, and uses a predetermined pitch change amount (PTDIFF) to process the terminal phoneme first pitch and the terminal phoneme second pitch in a process described later. A pitch change amount (PDif) used as the change amount is calculated by the equation (2) (step S1104).

PDif = PTDIFF Expression (2)
Here, the predetermined pitch change amount (PTDIFF) is a value indicating the change amount when the pitch is expressed in logarithm, and as an example, when expressing the pitch in natural logarithm, “0.6” may be used. .

  Next, the terminal phoneme pitch generation unit 332 uses the reference pitch (PitS) calculated in step S1103 and the pitch change amount (PDif) calculated in step S1104 to end the first phoneme pitch assigned to the first half of the terminal phoneme. (PitL1) is calculated by equation (3) (step S1105).

Log (PitL1) = Log (PitS) −PDif / 2 Equation (3)
Next, the terminal phoneme pitch generation unit 332 uses the reference pitch (PitS) calculated in step S1103 and the pitch change amount (PDif) calculated in step S1104, and the terminal phoneme second pitch assigned to the second half of the terminal phoneme. (PitL2) is calculated by equation (4) (step S1106).

Log (PitL2) = Log (PitS) + PDif / 2 Equation (4)
Then, the terminal phoneme pitch setting unit 333 uses the terminal phoneme first pitch (PitL1) calculated in step S1105 and the terminal phoneme second pitch (PitL2) calculated in step S110 as the pitch corresponding to the terminal phoneme in the standard pitch information. By replacing them, these pitches are set (steps S1107 and S1108).

  In the above description, Expressions (1) to (4) are merely examples, and all the following five conditions may be satisfied, and the detailed calculation method in the present invention is not limited.

・ Calculate the reference pitch (PitS) using the peak pitch (PitP) ・ Calculate the pitch change (PDif) using the default pitch change (PTDIFF) ・ The first pitch (PitL1) of the final phoneme is the reference pitch (PitS) The terminal phoneme second pitch (PitL2) is larger than the reference pitch (PitS). The difference (or ratio) between the terminal phoneme first pitch (PitL1) and the terminal phoneme second pitch (PitL2) may be: There is a correlation with the pitch change amount (PDif).

  Each equation may be changed as long as the same result can be obtained. For example, when calculating the terminal phoneme second pitch (PitL2) in step S1106, instead of equation (4), in step S1105 You may calculate by Formula (5) using the calculated terminal phoneme 1st pitch (PitL1).

Log (PitL2) = Log (PitL1) + PDif Expression (5)
In the above description, a series of processing steps from step S1101 to step S1108 has been described. However, the order of processing is not limited except when referring to the previous processing result. For example, the order of the pitch peak extraction processing in step S1101 and the terminal phoneme pitch acquisition processing in step S1102 may be interchanged, or may be executed simultaneously.

  Next, how the pitch information is corrected by the pitch correction processing described with reference to FIG. 11 will be specifically described with reference to FIGS. 12 to 15 show the natural logarithm of the pitch corresponding to the phoneme symbol string “s” “o” “t” “o” “d” “e” “a” “s” “o” “b” “u”. It is a figure which shows the change of. In these figures, the natural logarithm of the pitch is defined on the vertical axis, and the time is defined on the horizontal axis. FIG. 12 shows an example of a pitch sequence according to the standard pitch information. FIGS. 13 to 15 show pitch sequences generated by correcting the pitch sequence of FIG. First to third examples are shown.

  The phoneme duration length and pitch of each phoneme symbol in FIG. 12 are shown in Table 1. The phoneme duration length and pitch of each phoneme symbol in FIG. 13 are shown in Table 2. 14 shows the phoneme duration and pitch of each phoneme symbol in FIG. 14, and Table 4 shows the phoneme duration and pitch of each phoneme symbol in FIG. In Tables 3 and 4, the power of each phoneme symbol is also shown for the purpose of explanation to be described later.

First, description will be made with reference to FIGS.
FIG. 15 shows the corrected pitch information (black circle in the figure) corrected in step S110 and the standard pitch information before correction (broken line circle in the figure). In FIG. 15, PitP is , The pitch peak extracted in step S1101, and PitL is the terminal phoneme pitch acquired in step S1102. PitS is the reference pitch (PitP itself) calculated in step S1103, and PDif is the pitch change amount (0.6) calculated in step S1104. Also, PitL1 is the terminal phoneme first pitch calculated in step S1105, and PitL2 is the terminal phoneme second pitch calculated in step S1106.

  As understood from FIG. 12 and FIG. 15, the standard end phoneme pitch PitL is replaced with the end phoneme first pitch PitL1 and the end phoneme second pitch PitL2 by the pitch correction process, and the intonation of the ending that was in a downward trend Can be corrected to a rising intonation (intonation considered to be suitable for a question sentence).

  Here, a method of correcting PitL by adding the pitch change amount PDif to the pitch PitL corresponding to the terminal phoneme in the standard pitch information can be considered. However, when the accent type is 0 type (accent nucleus does not exist), the pitch PitL is close to the pitch peak PitP, and it is considered that the sentence end pitch becomes unnaturally high (too high). On the other hand, according to the present embodiment, the pitch at the end of the sentence is expressed by a smaller value (PitL1) and a larger value (PitL2) than the reference pitch PitS calculated based on the pitch peak PitP. It is possible to obtain a stable, natural questioning intonation.

  10 and 11, the pitch is shown in units of frequency (Hz). However, in FIGS. 12 to 15, the pitch is shown in natural logarithm of frequency. When discussing the tendency of the pitch so that the pitch of the sound is doubled by increasing the pitch by one octave when playing an instrument, it is more natural to compare the pitch value than to compare the pitch value itself. It may be considered appropriate to compare logarithms. Therefore, in the pitch correction according to the present embodiment, the pitch is handled in the form of its natural logarithm.

[Modification of pitch correction processing]
FIG. 16 is a flowchart illustrating a modification of the pitch correction process described with reference to FIG.

  In this modification, the processing contents of steps S2101, 1022, 2105 to 2108 are the same as the processing contents of steps S1101, 1102, 1105 to 1108 described with reference to FIG. In this modification, step S1103 is changed to step S2103, and step S1104 is changed to step S2104, compared to the processing content shown in FIG. That is, changes are made to the calculation of the reference pitch and the calculation of the pitch change amount.

  In the calculation of the reference pitch in step S2103, the terminal phoneme pitch generation unit 332 calculates the reference pitch (PitS) according to the following equation (6).

PitS = PitL + (PitP−PitL) × m / M (6)
Further, in the calculation of the pitch change amount in step S2104, the terminal phoneme pitch generation unit 332 calculates the pitch change amount (PDif) according to the following equation (7).

PDif = PTDIFF × n / N (7)
In the modification shown in FIG. 16, “m” and “M” in equation (6) and “n” and “N” in equation (7) adjust the pitch of the final phoneme after correction. For example, the value can be input by the user via the input unit 140.

  Specifically, when m is set small or M is set large, the value of the reference pitch (PitS) becomes small, and therefore the pitch of the terminal phoneme becomes low as a whole. Become.

  Further, when n is set to be small or N is set to be large, the amount of pitch change (PDif) is reduced, so that the degree of rising in the terminal phoneme is reduced.

Further, the processing content shown in FIG. 11 can be said to be an example in which m = M and n = N.
13, 14, and 15 are “m / M = 1/3, n / N = 1/3”, “m / M = 2/3, n / N = 2/3”, It is a figure which shows a pitch series when it is set as "m / M = 3/3, n / N = 3/3". In each figure, D11, D21, and D31 are the differences between the maximum pitch value (PitP) and the pitch of the terminal phoneme (PitL) in the pitch sequence according to the standard pitch information, and D11 = D21 = D31. D12, D22, and D32 are differences between the pitch (PitL) of the terminal phoneme in the pitch series according to the standard pitch information (PitS) and the standard pitch information. In the example shown in FIG. 15, since the reference pitch (PitS) is the maximum pitch value (PitP), D31 = D32.

  When comparing FIGS. 13 to 15, the value of m (m / M) and the value of n (n / N) are set larger in the order of FIGS. 13, 14, and 15. Correspondingly, in this order, the values of PitL1 and PitL2 increase in the terminal phoneme, and the amount of change from PitL1 to PitL2 increases.

  As described above, in this embodiment, the values of m, n, M, and N are adjusted to adjust the values of PitL1 and PitL2, that is, the degree of ending when the input text data is pronounced. It is also possible to adjust the amount of change from PitL1 to PitL2, that is, the degree of inflection at the end of the word (the place that was originally regarded as the terminal phoneme). Thereby, in this Embodiment, the intensity of intonation expressing a question can be easily adjusted by adjusting the value of m, n, M, and N.

  Further, the reference pitch (PitS) and the pitch change amount (PDif) are not adjusted independently, but m = n and M = N, and the reference pitch (PitS) and the pitch change amount (PDif) are set for the same adjustment. By adjusting based on the value, the intensity of intonation expressing a question can be adjusted more easily.

  For reference, FIG. 17 and FIG. 18 show the power corresponding to the corrected pitch series and the corrected phoneme symbols, respectively, corresponding to FIG. 14 and FIG. The horizontal and vertical axes in FIGS. 17 and 18 are the same as those in FIGS. 12 and 13. In FIGS. 17 and 18, the pitch is indicated by “●” and the power is indicated by “◯” as in FIGS. 12 and 13.

  In the present embodiment described above, the pitch peak extraction unit 331 extracts the maximum value of the pitch in the pitch sequence according to the standard prosodic information, and the terminal phoneme pitch generation unit 332 uses the maximum value to terminate the pitch. A phoneme first pitch and a terminal phoneme second pitch (a pitch given to the terminal phoneme of the input text data) are generated, and the terminal phoneme pitch setting unit 333 converts the text data input to the text analysis unit 1 into text data. Corresponding pitch sequences and corrected pitch information are generated.

  In this way, the maximum pitch value (PitP) in the pitch sequence according to the standard prosodic information is used for correcting the pitch sequence, so that various pattern data are created in advance for correcting the pitch sequence. It is possible to generate intonation expressing a question without having to store a large amount of data such as pattern data for prosody correction.

  Also, the above-mentioned first phoneme first pitch is calculated by subtracting the pitch change amount (PDif) divided by 2 from PitP, and the final phoneme second pitch is obtained by dividing PDif by Pif by 2 Calculated by adding together. This makes it possible to generate an intonation that expresses a question in the terminal phoneme, and to generate a natural intonation along the intonation basically given to the input text data.

  Further, in the present embodiment, since the terminal phoneme first pitch is set to a value smaller than the terminal phoneme second pitch, it is possible to generate intonation expressing a question in the terminal phoneme.

  In the present embodiment, the terminal phoneme first pitch is set to a value smaller than the terminal phoneme pitch according to the standard pitch information (PitL: a pitch given in advance to the terminal phoneme), and the terminal phoneme second pitch. Is set to a value larger than the above PitL, it is possible to generate a natural intonation along the sentence intonation expressing a question in the terminal phoneme.

  Further, in the present embodiment, as shown in FIG. 13 or FIG. 14, the reference pitch is calculated so as to be between the above PitP and PitL, and the terminal phoneme first pitch is smaller than the reference pitch. Since the terminal phoneme second pitch is set to a value larger than the reference pitch, the terminal phoneme generates a natural intonation along the sentence intonation that expresses a question regardless of the accent type of the sentence. Is possible.

  In the present embodiment, the terminal phoneme is subjected to power correction in addition to pitch correction, so that an intonation expressing a question can be generated.

  Further, in the present embodiment, the power correction is performed by using the power of the terminal phoneme (PowL: power given in advance to the terminal phoneme) according to the standard power information to generate a correction value of the power of the terminal phoneme. Therefore, it is possible to generate a natural intonation expressing a question in the terminal phoneme.

  Further, in the present embodiment, the terminal phoneme first half power (PowL1) is set to the above-mentioned PowL, and the terminal phoneme second half power (PowL2) is set to a value smaller than PowL1. It is possible to generate natural intonation along the sentence intonation to be expressed.

  In the present embodiment, the corrected phoneme duration information is generated using the phoneme duration of the terminal phoneme according to the standard phoneme duration information (the phoneme duration previously given to the terminal phoneme). This makes it possible to generate a natural intonation that expresses a question in the terminal phoneme.

  Further, in the present embodiment, based on the phoneme duration information (DurL) of the terminal phoneme according to the standard phoneme duration time information, and DRWGT1 and DRWGT2 that are both smaller than 1, the first phoneme last duration ( By generating the DurL1) and the terminal phonological second half duration (DurL2), these are shorter than the above DurL, and in the terminal phonology, a natural intonation according to the sentence intonation expressing the question is generated. It becomes possible.

  In the present embodiment described above, the prosody correction unit 3 shows an embodiment of the prosody correction device of the present invention. Note that the prosody correction device may include at least the pitch correction unit 33.

  In the present embodiment described above, an embodiment of the speech synthesizer 100 of the present invention is shown by the speech synthesizer 100 including the correction / synthesis unit 111. Note that the speech synthesizer may include at least the pitch correction unit 33 and the synthesis unit 4.

  The embodiment disclosed this time should be considered as illustrative in all points and not restrictive. The scope of the present invention is defined by the terms of the claims, rather than the description above, and is intended to include any modifications within the scope and meaning equivalent to the terms of the claims.

  The present invention relates to text data recorded in advance or dynamically read via a network or the like, such as telephones, home appliances, game machines, game machine software, personal computers, and personal computer software. Alternatively, the present invention can be applied to a device that outputs sound via headphones or the like.

  In addition, the present invention can be recorded in advance, such as a telephone, a game machine, game machine software, a personal computer, or personal computer software, or dynamically loaded via a network, or via a microphone. The present invention can be applied to a device that outputs voice data to be recorded through a speaker or headphones.

  In addition, the present invention outputs text data or voice data recorded in advance or dynamically read via a network or the like, such as an interactive voice guidance server or an interactive game server. It can be applied to such a device.

It is a figure which shows the hardware constitutions of the speech synthesizer which is one embodiment of this invention. It is a functional block diagram which shows the structure of the correction | amendment / synthesis | combination part of the speech synthesizer of FIG. It is a flowchart of the speech synthesis process performed in the speech synthesizer of FIG. It is a functional block diagram which shows the structure of the power correction part of FIG. It is a flowchart of the subroutine of the power correction process of FIG. It is a figure which shows typically an example of the standard prosodic information handled in the speech synthesizer of FIG. It is a figure which shows typically an example of the correction | amendment prosody information handled in the speech synthesizer of FIG. It is a functional block diagram which shows the structure of the phoneme duration time correction | amendment part of FIG. It is a flowchart of the subroutine of the phoneme duration correction process of FIG. It is a functional block diagram which shows the structure of the pitch correction | amendment part of FIG. It is a flowchart of the subroutine of the pitch correction process of FIG. It is a figure which shows an example of the pitch series according to standard pitch information handled in the speech synthesizer of FIG. It is a figure which shows the 1st example of the pitch series produced | generated by correcting the pitch series of FIG. 12 in the pitch correction | amendment part of FIG. It is a figure which shows the 2nd example of the pitch series produced | generated by correcting the pitch series of FIG. 12 in the pitch correction part of FIG. It is a figure which shows the 3rd example of the pitch series produced | generated by correcting the pitch series of FIG. 12 in the pitch correction | amendment part of FIG. It is a flowchart of the modification of the process shown by FIG. It is a figure which shows typically the other example of the correction | amendment prosody information handled in the speech synthesizer of FIG. It is a figure which shows typically the further another example of the correction | amendment prosody information handled in the speech synthesizer of FIG.

Explanation of symbols

  1 text analysis unit, 2 prosody generation unit, 3 prosody correction unit, 4 synthesis unit, 21 power generation unit, 22 phoneme duration duration generation unit, 23 pitch generation unit, 31 power correction unit, 32 phoneme duration length correction unit, 33 Pitch correction unit, 100 Speech synthesizer, 110 Central processing unit, 111 Correction / synthesis unit, 120 Antenna, 121 Communication control unit, 130 Flash memory, 140 Input unit, 150 Prosodic information table storage unit, 160 ROM, 170 Display unit , 180 amplifier, 181 speaker, 311 terminal phoneme power generation unit, 312 terminal phoneme power setting unit, 321 terminal phoneme duration generation unit, 322 terminal phoneme duration setting unit, 331 pitch peak extraction unit, 332 terminal phoneme pitch generation Part 333, terminal phoneme pitch setting part.

Claims (19)

Pitch peak extracting means for extracting the maximum value from the end of the pitch sequence given in advance to the phoneme symbol sequence;
Using the maximum value of the extracted pitch, terminal phoneme pitch generation means for generating a terminal phoneme pitch that is a pitch to be given to the terminal phoneme in the phoneme symbol sequence;
A terminal phoneme pitch correction unit that corrects a pitch sequence previously given to the phoneme symbol sequence by giving the generated pitch for the end phoneme to a pitch of the end phoneme in the phoneme symbol sequence;
The terminal phoneme pitch generating means includes
Calculating a reference pitch corresponding to a value between the maximum value of the extracted pitch and a pitch given in advance to the terminal phoneme in the phoneme symbol sequence;
As the terminal phoneme pitch, a first pitch and a second pitch are generated,
The first pitch is generated using a predetermined value representing the amount of change in pitch in the terminal phoneme of the phoneme symbol sequence, and is a value smaller than the reference pitch,
The second pitch is generated using the predetermined value and is larger than the reference pitch.
The amount of change between the first pitch and the second pitch is calculated using the predetermined value,
The prosody correction device, wherein the amount of change between the first pitch and the second pitch is a difference or ratio between the first pitch and the second pitch.   The prosody correction device according to claim 1, wherein the pitch peak extraction unit extracts a maximum value from an accent phrase at the end of a pitch sequence or an exhalation paragraph at the end of a sequence given to a phoneme symbol sequence in advance.   The terminal phoneme pitch generation means sets the first pitch to a value smaller than the maximum value of the extracted pitch, and sets the second pitch to a value larger than the maximum value of the extracted pitch. The prosody correction device according to claim 1 or 2. The terminal phoneme pitch generation means sets the extracted maximum value as PitP, a pitch given in advance to the terminal phoneme in the phoneme symbol sequence as PitL, a predetermined value representing the amount of change in the pitch as PDTIFF, and the phoneme When values regarding the degree of correction of the pitch of the terminal phoneme in the symbol series are m, M, n, and N, using these values, PitS that is the reference pitch is set as follows.
PitS = PitL + (PitP−PitL) × m / M
According to
PDif, which is the amount of change between the first pitch and the second pitch,
PDif = PTDIFF × n / N
The prosody correction device according to claim 1, wherein the prosody correction device according to claim 1 is calculated.   5. The prosody correction device according to claim 4, wherein the values m, M, n, and N relating to the degree of correcting the pitch of the terminal phoneme are m = n and M = N.   Power correcting means for correcting power given in advance to a terminal phoneme in the phoneme symbol sequence, or phoneme duration length correcting means for correcting a duration given in advance to a terminal phoneme in the phoneme symbol sequence The prosody correction apparatus according to claim 1, further comprising:   The prosody correction apparatus according to claim 6, wherein the power correction unit corrects the power of the terminal phoneme using power preliminarily given to the terminal phoneme. The power correction means includes
Correcting the power of the terminal phoneme to generate a first power to be given to the terminal phoneme and a second power to be given to the terminal phoneme;
Generating the second power based on a predetermined value representing an amount of attenuation with respect to the power given in advance to the terminal phoneme, and generating the second power so as to be smaller than the power given in advance to the terminal phoneme; The prosody correction device according to claim 7, characterized in that it is characterized in that:   The prosody correction according to claim 6, wherein the phoneme duration correction means corrects the phoneme duration of the terminal phoneme using a phoneme duration length previously given to the terminal phoneme. apparatus.   The phoneme duration correction means corrects the phoneme duration of the terminal phoneme, thereby correcting the phoneme duration based on a predetermined value representing a ratio to the phoneme duration length previously given to the terminal phoneme. 10. The prosody correction device according to claim 9, wherein a first phoneme duration length and a second phoneme duration length value that are smaller than a predetermined phoneme duration length are generated. Text analysis means for analyzing input text data and obtaining phonological symbols related to the text data and ending prosodic control information representing a controlling method of ending prosody;
Prosody generation means for generating prosody information including at least a pitch sequence of the text data based on the acquired phoneme symbol;
Pitch peak extraction means for extracting the maximum value of the pitch from the end of the pitch sequence generated by the prosody generation means;
Using the maximum value of the extracted pitch, a terminal phoneme pitch generating means for generating a terminal phoneme pitch that is a pitch to be given to the terminal phoneme of the text data;
Prosody correction means for correcting the prosodic information of the text data generated by the prosody generation means by giving the pitch for the terminal phoneme to the pitch of the terminal phoneme of the text data;
Using the acquired phonological symbol and the corrected prosodic information, comprising a synthesis means for synthesizing a speech signal related to the text data,
The terminal phoneme pitch generating means includes
Calculating a reference pitch corresponding to a value between the maximum value of the extracted pitch and a pitch given in advance to the terminal phoneme in the phoneme symbol sequence;
As the terminal phoneme pitch, a first pitch and a second pitch are generated,
The first pitch is generated using a predetermined value representing the amount of change in pitch in the terminal phoneme of the phoneme symbol sequence, and is a value smaller than the reference pitch,
The second pitch is generated using the predetermined value and is larger than the reference pitch.
The amount of change between the first pitch and the second pitch is calculated using the predetermined value,
The speech synthesis apparatus, wherein the amount of change between the first pitch and the second pitch is a difference or ratio between the first pitch and the second pitch.   12. The speech synthesizer according to claim 11, wherein the pitch peak extracting unit extracts a maximum value from an accent phrase at the end of a pitch sequence or an exhalation paragraph at the end of a pitch sequence given in advance to a phoneme symbol sequence.   The speech synthesizer according to claim 11 or 12, further comprising speech unit storage means for storing a plurality of speech units representing speech information corresponding to one or more phoneme symbols. . Extracting a maximum pitch value from the end of a pitch sequence given in advance to the phoneme symbol sequence;
Generating a terminal phoneme pitch that is a pitch to be given to a terminal phoneme of the phoneme symbol sequence using the maximum value of the extracted pitch;
Providing the terminal phoneme pitch to the terminal phoneme of the phoneme symbol sequence,
Generating the terminal phoneme pitch comprises:
Calculating a reference pitch corresponding to a value between the maximum value of the extracted pitch and a pitch given in advance to a terminal phoneme in the phoneme symbol sequence;
As the terminal phoneme pitch, a first pitch and a second pitch are generated,
The first pitch is generated using a predetermined value representing the amount of change in pitch in the terminal phoneme of the phoneme symbol sequence, and is a value smaller than the reference pitch,
The second pitch is generated using the predetermined value and is larger than the reference pitch.
The amount of change between the first pitch and the second pitch is calculated using the predetermined value,
The prosody correction method, wherein the amount of change between the first pitch and the second pitch is a difference or ratio between the first pitch and the second pitch.   The prosody correction method according to claim 14, wherein the step of extracting the maximum value of the pitch extracts the maximum value from an accent phrase at the end of a pitch sequence or an exhalation paragraph at the end of the sequence given to a phoneme symbol sequence in advance. A prosody correction program for correcting the prosody of a terminal phoneme of a phoneme symbol sequence,
A prosody correction program for realizing the prosody correction method according to claim 14 or 15 on a computer. Analyzing the input text data, obtaining a phonological symbol related to the text data and a sentence end prosody control information representing a sentence end prosody control method;
Generating prosodic information including at least a pitch sequence of the text data based on the acquired phonological symbol;
Extracting the maximum pitch value from the end of the generated pitch sequence;
Generating a terminal phoneme pitch that is a pitch to be given to the terminal phoneme of the text data using the maximum value of the extracted pitch;
Correcting the generated prosodic information by assigning the terminal phoneme pitch to the terminal phoneme of the text data;
Using the acquired phonological symbol and the corrected prosodic information, synthesizing a speech signal related to the text data,
Generating the terminal phoneme pitch comprises:
Calculating a reference pitch corresponding to a value between the maximum value of the extracted pitch and a pitch given in advance to a terminal phoneme in the phoneme symbol sequence;
As the terminal phoneme pitch, a first pitch and a second pitch are generated,
The first pitch is generated using a predetermined value representing the amount of change in pitch in the terminal phoneme of the phoneme symbol sequence, and is a value smaller than the reference pitch,
The second pitch is generated using the predetermined value and is larger than the reference pitch.
The amount of change between the first pitch and the second pitch is calculated using the predetermined value,
The speech synthesis method, wherein the amount of change between the first pitch and the second pitch is a difference or ratio between the first pitch and the second pitch.   The speech synthesizing method according to claim 17, wherein the step of extracting the maximum value of the pitch extracts the maximum value from an accent phrase at the end of a pitch sequence or an exhalation paragraph at the end of the sequence given in advance to a phoneme symbol sequence. A speech synthesis program for synthesizing speech signals based on input text data,
A speech synthesis program for implementing the speech synthesis method according to claim 17 or 18 on a computer.

Images