JP2004206144A - Fundamental frequency pattern generating method and program recording medium - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To solve problems wherein when a fundamental frequency pattern is generated, a fluctuation in fundamental frequency in a mora can not precisely be determined, and a rhythm represented by an accent becomes unnatural owing to distortion generation on a real-time axis. <P>SOLUTION: Provided are: a phoneme time length standardized fundamental frequency database 351 which stores the number of the moras of an accent phrase, a rise reference point of an (i)th mora as the peak of a fundamental frequency pattern, etc., in the form of positions relative to the time length of a phoneme of a specified mora; a fundamental frequency pattern deformation database 350 which stores peaks of an accent phrase by positions in phrases of the accent phrase and the quantity of deformation in fundamental frequency at the ending; a time length setting part 40 which sets a time length of each phoneme by referring to a phoneme time length database 30 according to phoneme information etc., outputted from a character string analysis part 20; a generation part 60 which generates a fundamental frequency pattern by referring to the above-mentioned database 351 according to the above-mentioned outputted phoneme information and the time length of a phoneme set by the time length setting part 40; and others. <P>COPYRIGHT: (C)2004,JPO&NCIPI

Description

  The present invention relates to a fundamental frequency pattern generation method used for speech synthesis, and a program recording medium.

As a conventional voice fundamental frequency pattern generation method, as in Patent Document 1, focusing on the type of accent, the starting point of the mora or the starting point of the vowel of the mora is used as a reference, and a critical braking quadratic linear on the logarithmic frequency axis is used. Some systems determine the fundamental frequency pattern. On the other hand, as in Patent Document 2, there is a method in which a fundamental frequency for each mora is determined by focusing on the type of accent, the type of phoneme, and the mora position of a word or phrase.
JP-A-5-173590 JP-A-5-88690

  However, with these conventional methods, fluctuations in the fundamental frequency within the mora cannot be determined accurately, or distortions on the real time axis due to differences in the time length of each mora cause prosody represented by accents. There was a problem of becoming natural.

  The present invention provides a fundamental frequency pattern generation method and a program recording medium capable of generating a more natural fundamental frequency pattern compared to the related art, in consideration of the above-described problem of the conventional audio fundamental frequency pattern generation method. With the goal.

A first aspect of the present invention is a fundamental frequency pattern generating method for generating a fundamental frequency pattern of an accent phrase using a fundamental frequency database storing fundamental frequency patterns classified by the number of phonemes and accent positions,
The basic frequency pattern corresponding to the number of phonemes and the accent type of the accent phrase for which the basic frequency pattern is to be generated is not stored in the basic frequency database, and furthermore, the basic frequency pattern is to be generated. If the accent position of the accent phrase is the same or earlier than the next phoneme position of the phoneme position including the peak of the fundamental frequency stored in the fundamental frequency database,
(1) A basic frequency pattern stored in the basic frequency database and having the same accent position as that of the accent phrase for which the basic frequency pattern is to be generated, and an accent for which the basic frequency pattern is to be generated. Using the fundamental frequency pattern stored in the fundamental frequency database, corresponding to the number of phonemes closest to the number of phonemes in the phrase,
(2) Applying a fundamental frequency pattern from the first phoneme to the next phoneme of the accent kernel by applying a fundamental frequency from the first phoneme of the fundamental frequency pattern stored in the fundamental frequency database to the next phoneme of the accent kernel. Generate
(3) The basic frequency from the accent nucleus to the phoneme immediately before the end of the accent phrase, which is a predetermined number of phonemes less than or equal to 4 from the second phoneme, is calculated as (a) in the fundamental frequency pattern stored in the fundamental frequency database. ) A fundamental frequency between a second phoneme from the accent kernel and the end of the accent phrase, or (b) a fundamental frequency between a phoneme next to the accent kernel and the end of the accent phrase, or (c) a second phoneme from the accent kernel. And (d) a fundamental frequency between the phoneme immediately before the end of the accent phrase and a fundamental frequency between the phoneme immediately before the end of the accent phrase and (d) a fundamental frequency between the phoneme immediately before the end of the accent phrase and the fundamental frequency of the next phoneme of the accent kernel.
(4) By applying the fundamental frequency pattern at the end of the accent phrase of the accent phrase for which the fundamental frequency pattern is to be generated to the fundamental frequency of the phoneme at the end of the accent phrase of the fundamental frequency pattern stored in the fundamental frequency database. This is a basic frequency pattern generation method to be generated.

A second invention is a fundamental frequency pattern generation method for generating a fundamental frequency pattern of an accent phrase using a fundamental frequency database storing fundamental frequency patterns classified according to the number of phonemes and accent positions,
The basic frequency pattern corresponding to the number of phonemes and the accent type of the accent phrase for which the basic frequency pattern is to be generated is not stored in the basic frequency database, and further, the basic frequency pattern is to be generated. The accent position of the accent phrase is located after the phoneme position next to the phoneme position including the peak of the fundamental frequency stored in the fundamental frequency database and before the end of the accent phrase which is a predetermined number of phonemes of 4 or less. If you do
(1) It is a fundamental frequency pattern having an accent nucleus after the second phoneme from the peak of the fundamental frequency stored in the fundamental frequency database and before the end of the accent phrase, and further attempts to generate the fundamental frequency pattern. Using the fundamental frequency pattern stored in the fundamental frequency database, corresponding to the number of phonemes closest to the number of phonemes of the accent phrase,
(2) A basic frequency pattern from the first phoneme of the accent phrase for which the basic frequency pattern is to be generated to the phoneme including the peak of the basic frequency is obtained from the first phoneme of the basic frequency pattern stored in the basic frequency database. Generate by applying the fundamental frequency up to the phoneme including the peak of the fundamental frequency,
(3) The fundamental frequencies from the phoneme following the phoneme including the peak of the fundamental frequency to the phoneme immediately before the accent nucleus are expressed by: (a) the fundamental frequency peak stored in the fundamental frequency database; (B) the fundamental frequency of the phoneme including the peak of the fundamental frequency and the fundamental frequency of the phoneme immediately preceding the phoneme including the accent kernel, or (c) the fundamental frequency of the phoneme including the accent kernel. A fundamental frequency of a phoneme next to the phoneme including the frequency peak and a phoneme including the accent kernel, or (d) a phoneme next to the phoneme including the peak of the fundamental frequency and a phoneme immediately before the phoneme including the accent kernel. Generated by interpolation with the fundamental frequency
(4) The basic frequencies of the phoneme including the accent nucleus of the accent phrase for which the basic frequency pattern is to be generated and the phoneme immediately following it are obtained by combining the phoneme including the accent nucleus of the basic frequency pattern stored in the basic frequency database with the phoneme. Generated by applying the fundamental frequency with the phoneme immediately after,
(5) The fundamental frequency from the accent nucleus to the phoneme immediately before the end of the accent phrase, which is the predetermined number of phonemes of 4 or less from the second phoneme, is calculated by using the basic frequency pattern stored in the fundamental frequency database as ( a) the fundamental frequency of the second phoneme from the accent kernel and the end of the accent phrase, or (b) the fundamental frequency of the next phoneme of the accent kernel and the end of the accent phrase, or (c) the fundamental frequency of 2 from the accent kernel. A fundamental frequency between a phoneme and a phoneme immediately before the end of the accent phrase, or (d) a fundamental frequency between a phoneme next to the accent kernel and the phoneme immediately before the end of the accent phrase, and generated by interpolation.
(6) applying the fundamental frequency pattern at the end of the accent phrase of the accent phrase for which the fundamental frequency pattern is to be generated to the fundamental frequency of the phoneme at the end of the accent phrase of the fundamental frequency pattern stored in the fundamental frequency database. This is a method of generating a fundamental frequency pattern generated by.

A third aspect of the present invention is a fundamental frequency pattern generation method for generating a fundamental frequency pattern of an accent phrase using a fundamental frequency database storing fundamental frequency patterns classified by the number of phonemes and accent positions,
If the fundamental frequency pattern corresponding to the number of phonemes and the accent type of the accent phrase for which the fundamental frequency pattern is to be generated is not stored in the fundamental frequency database, and the accent for which the fundamental frequency pattern is to be generated is When the accent position of the phrase is included in the phoneme at the end of the accent phrase having a predetermined number of phonemes of 4 or less,
(1) The basic position stored in the fundamental frequency database, wherein the accent position within the accent phrase end and the accent position within the accent phrase end of the accent phrase for which the fundamental frequency pattern is to be generated are the same. Using a fundamental frequency pattern stored in the fundamental frequency database, which is a frequency pattern and corresponds to the number of phonemes closest to the number of phonemes of the accent phrase to generate the fundamental frequency pattern,
(2) A basic frequency pattern from the first phoneme of the accent phrase for which the basic frequency pattern is to be generated to the phoneme including the peak of the basic frequency is obtained from the first phoneme of the basic frequency pattern stored in the basic frequency database. Generate by applying the fundamental frequency up to the phoneme including the peak of the fundamental frequency,
(3) The fundamental frequencies from the phoneme following the phoneme including the peak of the fundamental frequency to the phoneme immediately before the accent nucleus are expressed by: (a) the fundamental frequency peak stored in the fundamental frequency database; (B) the fundamental frequency of the phoneme including the peak of the fundamental frequency and the fundamental frequency of the phoneme immediately preceding the phoneme including the accent kernel, or (c) the fundamental frequency of the phoneme including the accent kernel. A fundamental frequency of a phoneme next to the phoneme including the frequency peak and a phoneme including the accent kernel, or (d) a phoneme next to the phoneme including the peak of the fundamental frequency and a phoneme immediately before the phoneme including the accent kernel. Generated by interpolation with the fundamental frequency
(4) The fundamental frequencies from the phoneme including the accent nucleus of the accent phrase for which the fundamental frequency pattern is to be generated to the final phoneme of the accent phrase are calculated from the phoneme including the accent nucleus of the fundamental frequency pattern stored in the fundamental frequency database. This is a method of generating a fundamental frequency pattern by applying a fundamental frequency up to the final phoneme of an accent phrase.

A fourth invention is a fundamental frequency pattern generation method for generating a fundamental frequency pattern of an accent phrase using a fundamental frequency database storing fundamental frequency patterns classified according to the number of phonemes and accent positions,
The basic frequency pattern corresponding to the number of phonemes and the accent type of the accent phrase for which the basic frequency pattern is to be generated is not stored in the basic frequency database, and further, the basic frequency pattern is to be generated. If the accent type of the accent phrase is flat,
(1) Using a fundamental frequency pattern stored in the fundamental frequency database, which corresponds to the number of phonemes closest to the number of phonemes of the accent phrase for which the basic frequency pattern is to be generated with the flat type accent,
(2) The fundamental frequency pattern from the first phoneme to the phoneme including the peak of the fundamental frequency is defined as the fundamental frequency from the first phoneme of the fundamental frequency pattern stored in the fundamental frequency database to the phoneme including the peak of the fundamental frequency. Apply and generate,
(3) The basic frequencies from the phoneme following the phoneme including the peak of the fundamental frequency to the end of the accent phrase or immediately before the final phoneme are calculated using the basic frequency pattern stored in the fundamental frequency database, A fundamental frequency of a phoneme including a frequency peak and the end of the accent phrase or the final phoneme, or (b) a phoneme including a peak of the fundamental frequency and a phoneme at the end of the accent phrase or immediately before the final phoneme. Or (c) a fundamental frequency between the phoneme following the phoneme including the peak of the fundamental frequency and the end of the accent phrase or the final phoneme, or (d) a phoneme including the peak of the fundamental frequency. The next phoneme of the, the fundamental frequency of the phoneme at the end of the accent phrase or immediately before the final phoneme, generated by interpolation,
(4) The end of the accent phrase of the accent phrase for which the fundamental frequency pattern is to be generated or the fundamental frequency pattern of the final phoneme is stored in the fundamental frequency pattern stored in the fundamental frequency database. This is a basic frequency pattern generation method generated by applying a basic frequency of a phoneme.

  A fifth aspect of the present invention is the fundamental frequency pattern generation method according to any one of the first to fourth aspects of the present invention, wherein the fundamental frequency pattern is extracted from a naturally uttered voice.

  A sixth aspect of the present invention is the fundamental frequency pattern generation method according to any one of the first to fourth aspects, wherein the interpolation is linear interpolation.

  According to a seventh aspect of the present invention, there is provided a method according to any one of the first to fourth aspects, wherein a fundamental frequency from a head of the accent phrase to a peak of the fundamental frequency is interpolated by a fundamental frequency pattern expressed on a real time axis. This is a frequency pattern generation method.

  An eighth aspect of the present invention is the fundamental frequency pattern generation method according to any one of the first to fourth aspects of the invention, wherein the phoneme is a mora or a syllable.

  A ninth aspect of the present invention is a program recording medium storing a program for causing a computer to execute each step of the fundamental frequency pattern generation method according to any one of the first to fourth aspects of the present invention.

  As is apparent from the above description, the present invention has an advantage that a fundamental frequency pattern with higher naturalness can be generated as compared with the related art.

  Hereinafter, an embodiment of the present invention and another embodiment of the present invention related to the present invention will be described with reference to FIGS.

(Embodiment 1)
FIG. 1 is a functional block diagram of a fundamental frequency pattern generation device showing another embodiment of the present invention related to the present invention. The configuration of the present embodiment will be described with reference to FIG.

  That is, in FIG. 1, reference numeral 10 denotes a character string input unit for inputting a character string to be subjected to speech synthesis. Reference numeral 20 denotes a character string analysis unit that analyzes a character string input from the character string input unit 10 and outputs phonological information of a voice to be synthesized and prosody information such as accent and pause. Reference numeral 30 denotes a phoneme time length database that stores the time length of each phoneme for each condition such as the speech speed and the position of the phoneme in the speech, and 40 is based on the phoneme information and the prosody information output from the character string analysis unit 20. A time length setting unit for setting the time length of each phoneme with reference to the phoneme time length database 30. Reference numeral 50 denotes a mora time length standardized fundamental frequency database that stores, for each mora, a fundamental frequency pattern standardized by the time length of each mora with respect to the conditions of determinants of prosody such as the number of mora, accent type, phoneme sequence, etc. Reference numeral 60 denotes a fundamental frequency for generating a fundamental frequency pattern by referring to the mora time length standardized fundamental frequency database 50 based on the prosody information output from the character string analysis unit 20 and the phoneme time length set by the time length setting unit 40. This is a pattern generation unit. Reference numeral 70 denotes a vocal cord vibration generating unit that generates vocal cord vibrations based on the fundamental frequency pattern output from the fundamental frequency pattern generating unit, and generates a sound source vibration of a synthesized voice. FIG. 2 is an example of a fundamental frequency pattern according to the present invention.

  The operation of the fundamental frequency pattern generator configured as described above will be described below.

  First, a character string to be converted into speech (a character string of “on-go-go-se” shown in FIG. 2) is input from the character string input unit 10. The character string analysis unit 20 analyzes the input character string, outputs phoneme information indicating the phoneme sequence to the time length setting unit 40, divides the character string into accent phrases, and indicates the number of mora and the accent type of each accent phrase. Prosody information and phoneme information indicating a phoneme sequence are output to the fundamental frequency pattern generation unit 60. The time length setting unit 40 sets the time length of each phoneme with reference to the phoneme time length database 30 based on the phoneme information input from the character string analysis unit 20, and outputs the time length information to the fundamental frequency pattern generation unit 60. . The basic frequency pattern generation unit 60 generates a basic frequency pattern for each accent phrase based on the prosody information and phoneme information input from the character string analysis unit 20 and the time length information input from the time length setting unit 40.

  First, the basic frequency pattern of the first mora of the accent phrase is acquired from the mora time length standardized basic frequency database 50 as shown in a) of FIG. Next, the mora having the maximum fundamental frequency is specified from the number of mora of the accent phrase and the accent type, and the basic frequency pattern of the specified mora as shown in b) in FIG. get. As shown in c) and d) in FIG. 2, the fundamental frequency pattern of the mora next to the accent nucleus and the accent nucleus and the fundamental frequency pattern of the last mora of the accent phrase are acquired from the mora time length standardized fundamental frequency database 50. Using linear interpolation on the real-time axis between reference mora as in b) and c), c) and d) in FIG. 2, the fundamental frequency patterns in e), f) and g) in FIG. decide. The vocal fold vibration generating unit 70 generates vocal fold vibration of the synthesized sound according to the fundamental frequency pattern output from the fundamental frequency pattern generating unit 60.

  By applying a fundamental frequency pattern that standardizes the timing and angle of the rise of the accent phrase and the fall at the accent nucleus with the time length of the mora, which greatly affects the naturalness of speech, the variation of the fundamental frequency within the mora By reproducing in detail and realizing high naturalness, by interpolating on the real time axis for parts that do not greatly affect the hearing, the sense of discontinuity when controlling in mora units is eliminated, and the fundamental frequency pattern database is also Can be smaller.

(Embodiment 2)
FIG. 4 is a functional block diagram of an apparatus showing another embodiment related to the present invention, in which a mora time length standardized fundamental frequency database 50 determines a prosody such as the number of mora of an accent phrase, an accent type and a phoneme sequence. Regarding the condition of the factor, except that the time length of the vowel part of each mora was divided into four equal parts, and the vowel time length standardized fundamental frequency database 150a storing the representative value of the fundamental frequency of each section as the value of the center point of the section was replaced. It is the same as FIG.

  FIG. 3 is an example of a fundamental frequency pattern according to the present invention. The operation will be described below. First, a character string to be converted to speech is input from the character string input unit 10. The character string analysis unit 20 analyzes the input character string, outputs phoneme information indicating the phoneme sequence to the time length setting unit 40, divides the character string into accent phrases, and indicates the number of mora and the accent type of each accent phrase. Prosody information and phoneme information indicating a phoneme sequence are output to the fundamental frequency pattern generation unit 60. The time length setting unit 40 sets the time length of each phoneme with reference to the phoneme time length database 30 based on the phoneme information input from the character string analysis unit 20, and outputs the time length information to the fundamental frequency pattern generation unit 60. . The basic frequency pattern generation unit 60 generates a basic frequency pattern for each accent phrase based on the prosody information and phoneme information input from the character string analysis unit 20 and the time length information input from the time length setting unit 40.

  First, according to the number of mora of the accent phrase, the accent type, the phoneme sequence, and the like, the vowel-equivalent portion of the mora whose fundamental frequency takes the maximum value from the vowel time length standardized fundamental frequency database 150a is divided into four equal parts a). B) The middle of the third section where the vowel equivalent part of the mora hitting the accent nucleus is divided into four equal parts b) The falling reference point and the center of the third section where the vowel equivalent part of the mora next to the accent nucleus is divided into four equal parts C) the falling reference point, the middle of the second section obtained by dividing the vowel equivalent part of the last mora of the accent phrase into four equal parts; d) the third reference point obtained by dividing the accent phrase end reference point and the vowel equivalent part of the final mora into four equal parts; E) The end reference point at the center of the section is obtained.

  Next, each reference point is set at a position relative to the vowel time length of the corresponding mora. a) From the beginning of the accent phrase to a) the rising reference point, the interpolation is performed on the real time axis using a critical damping quadratic linear system for the logarithmic frequency axis so that the rising reference point becomes the maximum value. Interpolation is performed between the reference points a) to d) in each section between two points on a real time axis using a critical braking quadratic linear system with respect to a logarithmic frequency axis. Further, when the end of the accent phrase is the end of the utterance, interpolation is performed between d) the reference point at the end of the accent phrase and e) the reference point at the end by an end function which is a function on the real time axis. The vocal fold vibration generating unit 70 generates vocal fold vibration of the synthesized sound according to the fundamental frequency pattern output from the fundamental frequency pattern generating unit 60.

  By setting the timing of the rise of the accent phrase and the timing of the fall at the accent nucleus on the time axis standardized by the vowel length of the mora, which greatly affects the naturalness of the voice, the timing of the fluctuation of the fundamental frequency within the mora By using the functions on the real-time axis for the rising and falling angles, the pattern of the smooth fundamental frequency with stable rising and falling without being affected by the difference in time length due to phonemes. To achieve high naturalness. Furthermore, by performing interpolation on the real time axis for a portion that does not significantly affect the hearing, a sense of discontinuity when controlling in units of mora can be eliminated, and the fundamental frequency pattern database can be made smaller.

(Embodiment 3)
A functional block diagram of an apparatus showing another embodiment related to the present invention is as follows. In the second embodiment, the database 150a according to the second embodiment is such that the mora time length standardized fundamental frequency database 50 uses the number of mora of accent phrase, accent type, phoneme. The condition of the determinant of the prosody such as a column is stored in the vowel time length standardized fundamental frequency database 150b which stores the fundamental frequency pattern of the vowel part and the leading fundamental frequency of the accent phrase for each mora standardized by the time length of the vowel part of each mora. Except for the replacement, it is the same as FIG.

  FIG. 5 is an example of a fundamental frequency pattern according to the present invention.

  The operation of the fundamental frequency pattern generator configured as described above will be described below.

  First, a character string to be converted to speech (a character string of “oNse-go-se-” shown in FIG. 5) is input from the character string input unit 10. The character string analysis unit 20 analyzes the input character string, outputs phoneme information indicating the phoneme sequence to the time length setting unit 40, divides the character string into accent phrases, and indicates the number of mora and the accent type of each accent phrase. Prosody information and phoneme information indicating a phoneme sequence are output to the fundamental frequency pattern generation unit 60. The time length setting unit 40 refers to the phoneme time length database 30 based on the phoneme information input from the character string analysis unit 20 and determines the vowel time length of each mora or the time length of a single vowel syllable, vowel sound or a vowel equivalent part in a long sound. And outputs the time length information to the fundamental frequency pattern generation unit 60. The basic frequency pattern generation unit 60 generates a basic frequency pattern for each accent phrase based on the prosody information and phoneme information input from the character string analysis unit 20 and the time length information input from the time length setting unit 40.

  First, as shown at A in FIG. 5, the head fundamental frequency of the accent phrase is obtained from the vowel time length standardized fundamental frequency database 150b. Next, as shown in FIG. 5A, the fundamental frequency pattern of the vowel part of the first mora of the accent phrase is obtained from the vowel time length standardized fundamental frequency database 150b. In this example, since the first mora is a single vowel syllable, the fundamental frequency pattern acquired from the vowel duration standardized fundamental frequency database 150b is applied to the latter half of the duration of the mora as shown in a) of FIG. . Similarly, for b), c), d), e), f), g) and h), the fundamental frequency pattern of the vowel part of the mora is acquired from the vowel time length standardized fundamental frequency database 150b. Similarly to a), the fundamental frequency pattern acquired from the vowel duration standardized fundamental frequency database 150b is applied to the latter half of the duration of the mora for b) which is a sound repellent and d), f) and h) which are long sounds. . Next, the fundamental frequencies of monovowel syllables, vowel sounds, the first half of long sounds or voiced consonants a '), b'), d '), e'), f '), and h') are calculated by the fundamental frequencies before and after. Generated using linear interpolation on the time axis. The vocal fold vibration generating unit 70 generates vocal fold vibration of the synthesized sound according to the fundamental frequency pattern output from the fundamental frequency pattern generating unit 60.

  By applying a fundamental frequency pattern that standardizes the timing and angle of the rise of the accent phrase and the fall at the accent nucleus, which greatly affects the naturalness of the voice, by the vowel length of the mora, the variation of the fundamental frequency within the mora is reduced. By reproducing in detail and realizing high naturalness, by interpolating on the real time axis for parts that do not greatly affect the hearing, the sense of discontinuity when controlling in mora units is eliminated, and the fundamental frequency pattern database is also Can be smaller.

(Embodiment 4)
In the fourth embodiment, the vowel time length standardized fundamental frequency database 150a determines the conditions of prosodic determinants such as the number of mora of an accent phrase, the accent type, the phoneme sequence, and the like; The falling reference point (accent kernel), D) the falling reference point (immediately after the accent kernel), E) the accent phrase end reference point, and F) the ending reference point relative to the vowel time length of the mora including each reference point. It is a vowel time length standardized fundamental frequency database stored by position. Otherwise, the configuration of the apparatus is the same as that of FIG. FIG. 6 is an example of a fundamental frequency pattern according to the present invention. The operation will be described below.

  First, a character string to be converted to speech is input from the character string input unit 10. The character string analysis unit 20 analyzes the input character string, outputs phoneme information indicating the phoneme sequence to the time length setting unit 40, divides the character string into accent phrases, and indicates the number of mora and the accent type of each accent phrase. Prosody information and phoneme information indicating a phoneme sequence are output to the fundamental frequency pattern generation unit 60. The time length setting unit 40 sets the time length of each phoneme with reference to the phoneme time length database 30 based on the phoneme information input from the character string analysis unit 20, and outputs the time length information to the fundamental frequency pattern generation unit 60. . The basic frequency pattern generation unit 60 generates a basic frequency pattern for each accent phrase based on the prosody information and phoneme information input from the character string analysis unit 20 and the time length information input from the time length setting unit 40. First, the reference points of A) to F) are obtained from the vowel time length standardized fundamental frequency database 150a based on the number of mora of the accent phrase, the accent type, the phoneme sequence, and the like. Next, each reference point is set at a position relative to the vowel length of the corresponding mora. The interval between A) the top fundamental frequency and B) the rising reference point is generated using a function on the real time axis. Further, B is generated by interpolating a fundamental frequency pattern between the respective reference points after that with a straight line on the real time axis.

  The vocal fold vibration generating unit 70 generates vocal fold vibration of the synthesized sound according to the fundamental frequency pattern output from the fundamental frequency pattern generating unit 60.

  By setting the timing of the rise of the accent phrase and the timing of the fall at the accent nucleus on the time axis standardized by the vowel length of the mora, which greatly affects the naturalness of the voice, the timing of the fluctuation of the fundamental frequency within the mora By using the functions on the real-time axis for the rising and falling angles, the pattern of the smooth fundamental frequency with stable rising and falling without being affected by the difference in time length due to phonemes. To achieve high naturalness. Furthermore, by performing interpolation on the real time axis for a portion that does not significantly affect the hearing, a sense of discontinuity when controlling in units of mora can be eliminated, and the fundamental frequency pattern database can be made smaller.

(Embodiment 5)
FIG. 7 is a functional block diagram of an apparatus showing another embodiment related to the present invention. The vowel time length standardized fundamental frequency database 150a stores the number of mora of an accent phrase and the condition of an accent type. , B) the falling reference point (accent kernel), c) the falling reference point (immediately after the accent kernel), d) the accent ending reference point, and e) the ending reference point, or the mora vowel containing each reference point or A value obtained by interpolating between the reference points and the reference points stored in the vowel time length standardized fundamental frequency database 150a by storing the relative variation with respect to the time length of the vowel equivalent part and the fine fluctuation of the fundamental frequency due to the phoneme or phoneme sequence. 4 is the same as FIG. 4 except that a microprosody database 250 for standardizing and storing the difference in the phoneme duration is added.

  FIG. 8 is a schematic diagram of the microprosody components stored in the microprosody database 250, and FIGS. 9A to 9C are examples of the fundamental frequency pattern according to the present invention.

  The operation of the fundamental frequency pattern generator configured as described above will be described below.

  First, a character string to be converted to speech is input from the character string input unit 10. The character string analysis unit 20 analyzes the input character string, outputs phoneme information indicating the phoneme sequence to the time length setting unit 40, divides the character string into accent phrases, and indicates the number of mora and the accent type of each accent phrase. Prosody information and phoneme information indicating a phoneme sequence are output to the fundamental frequency pattern generation unit 60. The time length setting unit 40 sets the time length of each mora phoneme with reference to the phoneme time length database 30 based on the phoneme information input from the character string analysis unit 20, and uses the time length information as the fundamental frequency pattern generation unit 60. Output to The basic frequency pattern generation unit 60 generates a basic frequency pattern for each accent phrase based on the prosody information and phoneme information input from the character string analysis unit 20 and the time length information input from the time length setting unit 40. First, according to the number of mora and the accent type of the accent phrase, a) a rising reference point at the center of a third section obtained by dividing the vowel equivalent part of the mora having the maximum fundamental frequency into four parts from the vowel time length standardized fundamental frequency database. B) Falling reference point, middle of the third section where the vowel equivalent part of the mora hitting the accent nucleus is divided into four equal parts, c) Standing at the center of the third section where the vowel equivalent part of the mora next to the accent nucleus is divided into four equal parts D) the descent reference point, the middle of the second section where the vowel equivalent part of the last mora of the accent phrase is divided into four equal parts d) the accent reference point reference point, and the middle of the third section where the vowel equivalent part of the final mora is divided into four equal parts e) Get the ending reference point.

  Next, each reference point is set to a relative position with respect to the phoneme time length of the corresponding mora. a) From the head of the accent phrase to a) the rising reference point, the interpolation is performed on the real time axis and in the critical damping quadratic linear system with respect to the logarithmic frequency axis so that the rising reference point becomes the maximum value. Interpolation between the reference points a) to e) is carried out by a critical braking quadratic linear system on the real time axis and the logarithmic frequency axis between the two points for each section, and the fundamental frequency as shown in FIG. Generate a pattern. Next, a minute variation of the fundamental frequency corresponding to each phoneme is obtained from the microprosody database 250, expanded and compressed in accordance with the time length of each phoneme, and applied as shown in FIG. 9B. The basic frequency pattern shown in FIG. 9B is added to the basic frequency pattern shown in FIG. 9A to generate a basic frequency pattern as shown in FIG. The vocal fold vibration generating unit 70 generates vocal fold vibration of the synthesized sound according to the fundamental frequency pattern output from the fundamental frequency pattern generating unit 60.

  By setting the timing of the rise of the accent phrase and the timing of the fall in the accent nucleus on an axis standardized by the duration of the phoneme of the mora, the timing of the fluctuation of the fundamental frequency in the mora is reproduced in detail, and High naturalness and clarity are achieved by adding small variations in the fundamental frequency that affect naturalness and clarity.

(Embodiment 6)
FIG. 10 is a functional block diagram of an apparatus showing an embodiment of the present invention. The i-th mora in which the mora time length standardized fundamental frequency database 50 is the peak of the fundamental frequency pattern for the number of mora of the accent phrase and the condition of the accent type. A) a rising reference point, b) a falling reference point (accent nucleus), c) a falling reference point (immediately after the accent nucleus), d) a k-mora at the end of the accent phrase, and d) an accent phrase end reference point. Is replaced with a phoneme time length standardized fundamental frequency database 351 stored at a relative position with respect to the time length of the phoneme of the mora, and the peak and the end of the accent phrase for each position in the phrase of the accent phrase for which the fundamental frequency is to be generated. 1. Same as FIG. 1 except that a fundamental frequency pattern deformation database 350 storing the amount of deformation of the fundamental frequency is added. It is.

  11, 12, 13, and 14 show the basics generated when there is no data of the basic frequency pattern corresponding to the number of mora and the accent type of the accent phrase to be generated in the phoneme duration standardized basic frequency database 351. It is a schematic diagram of a frequency pattern. FIG. 15 is a schematic diagram of a fundamental frequency pattern of a sentence generated by connecting fundamental frequency patterns of a plurality of accent phrases. The operation will be described below.

  First, a character string to be converted to speech is input from the character string input unit 10. The character string analysis unit 20 analyzes the input character string, outputs phoneme information indicating the phoneme sequence to the time length setting unit 40, divides the character string into accent phrases, and indicates the number of mora and the accent type of each accent phrase. Prosody information and phoneme information indicating a phoneme sequence are output to the fundamental frequency pattern generation unit 60. The time length setting unit 40 sets the time length of each phoneme with reference to the phoneme time length database 30 based on the phoneme information input from the character string analysis unit 20, and outputs the time length information to the fundamental frequency pattern generation unit 60. . The basic frequency pattern generation unit 60 generates a basic frequency pattern for each accent phrase based on the prosody information and phoneme information input from the character string analysis unit 20 and the time length information input from the time length setting unit 40.

  First, based on the mora number of the accent phrase, the accent type, the phoneme sequence, and the like, a) a rising reference point, b) a falling reference point, c) a falling reference point, and d) an accent phrase end from the phoneme time length standardized fundamental frequency database 351. Reference point or d ') Obtain final mora.

  When the number of mora of the accent phrase for which the fundamental frequency is to be generated and the data of the fundamental frequency pattern corresponding to the accent type are not in the phoneme time length standardized fundamental frequency database 351, the number of mora of the accent phrase for which the fundamental frequency is to be generated is Assuming that n mora and m-type accent type, if m is equal to or less than i + 1, as shown in FIG. 11 (A), a 1-mora m-type fundamental frequency pattern a of m-type accent type and mora number closest to n ) To d) are obtained from the phoneme time standardized fundamental frequency database 351, and d) obtained from the phoneme time standardized basic frequency database 351 as shown in FIG. It is set as a reference point of the (n−k + 1) mora to the n-th mora.

  When m is larger than i + 1 and equal to or smaller than n−k, as shown in FIG. 12A, the mora position j of the accent nucleus is larger than i + 1 and equal to or smaller than 1−k, and the number of mora is closest to n. The frequency patterns a) to d) are obtained from the phoneme time length standardized fundamental frequency database 351, and the basic frequencies b) and c) obtained from the phoneme time length standardized fundamental frequency database 351 as shown in FIG. The d) obtained from the m-th mora and the (m + 1) -th mora of the accent phrase to be obtained and obtained from the phoneme time length standardized fundamental frequency database 351 is obtained from the (nk−1) th mora of the accent phrase for which the fundamental frequency is to be generated. It is set as the reference point of the nth mora.

  When m is larger than n−k, as shown in FIG. 13A, the mora position j of the accent nucleus is larger than 1−k and the number of mora is closest to n. ') From the phoneme time length standardized fundamental frequency database 351 and generate a fundamental frequency d') including b) and c) obtained from the phoneme time standardized fundamental frequency database 351 as shown in FIG. Are set as reference points of the (n−k + 1) th mora to the nth mora of the accent phrase. When the accent phrase for which the fundamental frequency is to be generated is an n-mora flat plate type, as shown in FIG. 14A, the 1-mora flat plate basic frequency pattern a) whose accent type is a flat type and whose mora number is closest to n is d) is obtained from the phoneme time length standardized fundamental frequency database 351, and d) obtained from the phoneme time length standardized fundamental frequency database 351 as shown in FIG. It is set as a reference point from k + 1 moras to n-th moras.

  Next, the fundamental frequency pattern of the accent phrase obtained from the reference point acquired from the phoneme time length standardized fundamental frequency database 351 or generated from the phoneme time standardized fundamental frequency database 351 is stored in the fundamental frequency deformation database 350. The maximum value of the fundamental frequency of each accent phrase, that is, the fundamental frequency of the reference point from a) to d) or d ') is changed according to the amount of deformation stored for each position in the phrase.

  First, as shown in A) in FIG. 15, the difference between the fundamental frequencies a) and d) is obtained from the phoneme time length standardized fundamental frequency database 351 based on the amount of transformation of the first accent phrase stored in the fundamental frequency transformation database 350. The basic frequencies b), c) and d) are changed so as to be 90% of the difference between the fundamental frequencies. For the second accent phrase, as shown in B) in FIG. 15, the fundamental frequency of a) is changed to a value of 75% of the fundamental frequency obtained from the phoneme time length standardized fundamental frequency database 351; The fundamental frequencies b), c) and d) are changed so that the difference between the fundamental frequencies of the above becomes 70% of the difference between the fundamental frequencies obtained from the phoneme time length standardized fundamental frequency database 351. Similarly, the third accent phrase also changes the fundamental frequency of a) to 70% of the fundamental frequency obtained from the phoneme time length standardized fundamental frequency database 351 as shown in C) in FIG. The fundamental frequencies b), c) and d) are changed so that the difference between the fundamental frequencies of the above becomes 68% of the difference between the fundamental frequencies obtained from the phoneme time length standardized fundamental frequency database 351.

  If the transformation amount corresponding to the n-th accent phrase is not stored in the fundamental frequency transformation database 350, the transformation amount corresponding to the accent position whose accent position value is smaller than n and closest to n is applied. This example shows a case where the deformation amount of the fourth accent phrase is not stored in the fundamental frequency deformation database 350.

  The amount of deformation of the third accent phrase in which the value of the accent position is smaller than 4 and closest to 4 is applied, and the same deformation as in the third accent phrase is added as shown in D) in FIG. For the final accent phrase at the end of the phrase, the amount of deformation corresponding to the final accent phrase is obtained from the basic frequency deformation database 350, and the basic frequency of a) is converted to the phoneme time length standardized basic frequency database as shown in E) in FIG. 351 is changed to a value of 48% of the fundamental frequency acquired so that the difference between the fundamental frequencies a) and d) becomes 60% of the difference between the fundamental frequencies acquired from the phoneme time length standardized fundamental frequency database 351. Change the fundamental frequencies of b), c) and d).

  Next, for each accent phrase, a fundamental frequency from the beginning of the accent phrase to a) is generated using a function on the real time axis as in the second or fourth embodiment, and furthermore, between each reference point. Is interpolated on the real time axis to generate a fundamental frequency pattern up to the end point of the accent phrase.

  The vocal fold vibration generating unit 70 generates vocal fold vibration of the synthesized sound according to the fundamental frequency pattern output from the fundamental frequency pattern generating unit 60.

  By setting the timing of the rise of the accent phrase and the fall timing at the accent nucleus on the time axis standardized by the phoneme time length of the mora, which greatly affects the naturalness of the voice, it is influenced by the difference in time length due to phonemes. This makes it possible to obtain a pattern with a stable fundamental frequency with a stable rise and fall without causing a high naturalness. Furthermore, the database can be reduced by extending the fundamental frequency pattern. Also, by transforming the fundamental frequency pattern based on the position of the accent phrase in the phrase, it is possible to form a unity as a phrase and realize a natural sentence voice.

(Embodiment 7)
FIG. 17 is a schematic diagram of a fundamental frequency pattern of a sentence generated by connecting fundamental frequency patterns of a plurality of accent phrases. The configuration of the device is the same as in FIG. The operation will be described below.

  First, a character string to be converted to speech is input from the character string input unit 10. The character string analysis unit 20 analyzes the input character string, outputs phoneme information indicating the phoneme sequence to the time length setting unit 40, divides the character string into accent phrases, and indicates the number of mora and the accent type of each accent phrase. Prosody information and phoneme information indicating a phoneme sequence are output to the fundamental frequency pattern generation unit 60. The time length setting unit 40 sets the time length of each phoneme with reference to the phoneme time length database 30 based on the phoneme information input from the character string analysis unit 20, and outputs the time length information to the fundamental frequency pattern generation unit 60. . The basic frequency pattern generation unit 60 generates a basic frequency pattern for each accent phrase based on the prosody information and phoneme information input from the character string analysis unit 20 and the time length information input from the time length setting unit 40.

  As shown in FIG. 17, first, a fundamental frequency pattern 1711 corresponding to the number of mora and the accent type of the first accent phrase 1701 is obtained from the mora time length standardized fundamental frequency database 50 and applied.

The maximum value a of the first accent phrase 1701 decreases by 10% every time the value of i indicating the position of the n-th accent phrase increases through the maximum value a of the fundamental frequency of the first accent phrase 1701. Expression 1 is obtained which indicates the maximum fundamental frequency of the accent phrase for the n accent phrase.
(Equation 1)
(−0.1i + 1) a Equation 1
Here, a is the maximum value of the fundamental frequency of the first accent phrase 1701. The number of accent phrases i is a number indicating the number of the accent phrase counted from the first accent phrase, and is n-1.

Further, every time the value of i indicating the position of the n-th accent phrase increases through the frequency b at the end of the accent phrase of the first accent phrase 1701, the frequency b at the end of the accent phrase of the first accent phrase 1701 decreases by 5%. Equation 2 indicating the frequency at the end of the accent phrase for the n-th accent phrase is obtained.
(Equation 2)
(−0.05i + 1) b Equation 2
Here, b is the frequency at the end of the accent phrase of the first accent phrase 1701.

Next, a basic frequency pattern 1712 (represented by a dotted line in the figure) corresponding to the number of moras and the accent type of the second accent phrase 1702 is obtained from the moras time length standardized basic frequency database 50. Since the number of accent phrases i of the second accent phrase is 1, this is substituted into Expression 1 to obtain the transformed maximum value a ₂ of the fundamental frequency pattern 1712. Similarly, the frequency b ₂ at the end of the accent phrase after the transformation of the fundamental frequency pattern 1712 is obtained from Expression 2.

The fundamental frequency pattern 1712 obtained from the mora time length standardized fundamental frequency database 50 so as to match the maximum value a ₂ after the transformation obtained in this way and the frequency b ₂ at the end of the accent phrase after the transformation. Is transformed, the transformed fundamental frequency pattern 1713 is used as the fundamental frequency pattern of the second accent phrase 1702.

  For the n-th accent phrase, if the accent phrase is not the final accent phrase (end of sentence), the number of mora of the n-th accent phrase and the fundamental frequency pattern corresponding to the accent type are obtained from the mora time length standardized fundamental frequency database 50. I do. Then, the maximum value of the obtained fundamental frequency pattern matches the value obtained from Expression 1, and the frequency at the end of the accent phrase of the obtained fundamental frequency pattern matches the value obtained from Expression 2. In this manner, the fundamental frequency pattern acquired from the database 50 is transformed and used as the fundamental frequency pattern of the n-th accent phrase.

  Further, when the accent phrase for which the fundamental frequency is to be generated is at the end of the sentence, the fundamental frequency pattern corresponding to the number of mora and the accent type is obtained from the mora time length standardized fundamental frequency database 50, and the maximum value is obtained as the corresponding accent. In order to match the maximum value of the accent phrase immediately before the phrase by 15%, and to match the frequency at the end of the accent phrase with the value of the accent phrase of the immediately preceding accent phrase by 10%, The fundamental frequency pattern acquired from the database 50 is modified and applied.

  If the data of the corresponding fundamental frequency pattern is not in the mora time length standardized fundamental frequency database 50, the fundamental frequency pattern of the accent phrase is generated and modified as in the sixth embodiment.

  The vocal fold vibration generating unit 70 generates vocal fold vibration of the synthesized sound according to the fundamental frequency pattern output from the fundamental frequency pattern generating unit 60.

  By setting on the time axis standardized by the mora time length of the mora, it is possible to obtain a smooth fundamental frequency pattern with stable rising and falling without being affected by the difference in time length due to phonemes, and Realize naturalness. Also, by transforming the fundamental frequency pattern based on the position of the accent phrase in the phrase, it is possible to form a unity as a phrase and realize a natural sentence voice.

Note that, in the above embodiment, only when the accent phrase for which the basic frequency is to be generated is at the end of a sentence, the frequency is reduced at a predetermined ratio with respect to the frequency at a predetermined position of the immediately preceding accent phrase. Was mentioned. Therefore, as a modified example of the above embodiment, each frequency value may be compressed according to the same rule as in the above example also for an accent phrase existing at a position other than the end of a sentence. That is, in this case, for example, as shown in FIG. 18, for the second accent phrase to the n-th accent phrase, excluding the end of the sentence, a value obtained by reducing the maximum value of the immediately preceding accent phrase by 10% (in the figure, for example, a _2), in the frequency of 5% reduction is allowed values of accent phrase end of the immediately preceding accent phrase (Figure, for example, determine the b _2).

Then, for example, for the second accent phrase, and the maximum value a ₂ after the deformation obtained in this manner, to match the frequency b ₂ of the accent phrase end after deformation, mora time length standardized fundamental frequency data After transforming the fundamental frequency pattern 1712 obtained from the source 50, the transformed fundamental frequency pattern 1713 is used as the fundamental frequency pattern of the second accent phrase 1702. The same applies to the n-th accent phrase. If the accent phrase for which the fundamental frequency is to be generated is at the end of a sentence, a method similar to that in FIG. 17 is used.

(Embodiment 8)
FIG. 19 is a schematic diagram of a fundamental frequency pattern of a sentence generated by connecting fundamental frequency patterns of a plurality of accent phrases. The configuration of the device is the same as in FIG. The operation will be described below.

  First, a character string to be converted to speech is input from the character string input unit 10. The character string analysis unit 20 analyzes the input character string, outputs phoneme information indicating the phoneme sequence to the time length setting unit 40, divides the character string into accent phrases, and indicates the number of mora and the accent type of each accent phrase. Prosody information and phoneme information indicating a phoneme sequence are output to the fundamental frequency pattern generation unit 60. The time length setting unit 40 sets the time length of each phoneme with reference to the phoneme time length database 30 based on the phoneme information input from the character string analysis unit 20, and outputs the time length information to the fundamental frequency pattern generation unit 60. . The basic frequency pattern generation unit 60 generates a basic frequency pattern for each accent phrase based on the prosody information and phoneme information input from the character string analysis unit 20 and the time length information input from the time length setting unit 40.

  As shown in FIG. 19, first, the basic frequency pattern 1811 corresponding to the number of moras and the accent type of the first accent phrase 1801 is obtained from the moras time length standardized basic frequency database 50 and applied.

The maximum value a of the accent phrase 1801 decreases by 2% every time the number of mora from the mora position including the maximum value a of the fundamental frequency of the first accent phrase 1801 passes through the maximum value a of the fundamental frequency of the first accent phrase 1801. Expression 3 indicating the maximum fundamental frequency of the accent phrase with respect to the accumulated number of mora j is obtained.
(Equation 3)
(−0.02j + 1) a Equation 3
Here, a is the maximum value of the fundamental frequency of the first accent phrase 1801, and the cumulative number of moras j is the mora position including the maximum value a of the fundamental frequency of the first accent phrase 180 (in FIG. ) Is the number of mora counted based on).

Further, each time the number of mora from the mora position including the frequency b at the end of the accent phrase of the first accent phrase and passing through the frequency b at the end of the accent phrase of the first accent phrase 1801 increases, the end of the accent phrase of the first accent phrase 1801 is increased. Equation 4 is obtained which indicates the frequency at the end of the accent phrase with respect to the accumulated number of moras j such that the frequency b decreases by 1%.
(Equation 4)
(-0.01j + 1) b ... Equation 4
Here, b is the frequency at the end of the accent phrase of the first accent phrase 1801.

Next, the number of mora of the second accent phrase 1802 and the basic frequency pattern 1812 (represented by a dotted line in the figure) corresponding to the accent type are obtained from the mora time length standardized basic frequency database 50, and the maximum value 1812a is obtained. mora taking is determined to become the origin of mora to j _2a mora th, which is substituted as the cumulative number of moras in equation 3, the maximum value a ₂ after the deformation of the fundamental frequency pattern 1812. Further, it is determined that the end of the accent phrase 1812b of the second accent phrase 1802 is j _2b mora from the origin mora, and this is substituted into Expression 4 as the cumulative number of mora, and the transformed accent phrase of the fundamental frequency pattern 1812 is obtained. determine the end of the frequency b _2.

The fundamental frequency pattern 1812 obtained from the mora time length standardized fundamental frequency database 50 so as to match the maximum value a ₂ after the transformation thus obtained and the frequency b ₂ at the end of the accent phrase after the transformation. Is used as the fundamental frequency pattern of the second accent phrase 1802.

  For the n-th accent phrase, if the accent phrase is not the final accent phrase (end of sentence), the number of mora of the n-th accent phrase and the fundamental frequency pattern corresponding to the accent type are obtained from the mora time length standardized fundamental frequency database 50. Then, the number of the mora that takes the maximum value is counted from the mora at the origin, and the number of the mora is obtained. This is substituted into Equation 3 as the number of the accumulated mora, and the maximum value after the deformation of the fundamental frequency pattern is obtained. Further, the number of moras counted from the mora at the origin is determined, and this is substituted into Equation 4 as the cumulative number of moras, to obtain the frequency of the accent phrase end after transformation of the fundamental frequency pattern.

  The fundamental frequency pattern obtained from the mora time length standardized fundamental frequency database 50 is transformed so as to match the maximum value after the transformation obtained in this way and the frequency at the end of the accent phrase after the transformation, Used as the fundamental frequency pattern of the nth accent phrase.

  If the accent phrase for which the fundamental frequency is to be generated is at the end of the sentence, the fundamental frequency pattern corresponding to the number of mora and the accent type is obtained from the mora time length standardized fundamental frequency database 50, and the maximum value is obtained for the accent. A fundamental frequency pattern acquired so that the maximum value of the accent phrase immediately before the phrase is reduced by 15% and the frequency at the end of the accent phrase matches the value obtained by decreasing the accent phrase end of the immediately preceding accent phrase by 10%. And apply it. If the data of the corresponding fundamental frequency pattern is not in the mora time length standardized fundamental frequency database 50, the fundamental frequency pattern of the accent phrase is generated and modified as in the sixth embodiment.

  The vocal fold vibration generating unit 70 generates vocal fold vibration of the synthesized sound according to the fundamental frequency pattern output from the fundamental frequency pattern generating unit 60.

  By setting on the time axis standardized by the mora time length of the mora, it is possible to obtain a smooth fundamental frequency pattern with stable rising and falling without being affected by the difference in time length due to phonemes, and Realize naturalness. Also, by transforming the fundamental frequency pattern based on the accumulated mora positions in the phrase, it is possible to form a unity as a phrase and realize a natural sentence voice.

(Embodiment 9)
FIG. 16 is a functional block diagram of an apparatus showing another embodiment related to the present invention. In the mora time length standardized fundamental frequency database 50, the accent phrases from the first accent phrase to the third accent phrase are sentence endings. The basic frequency pattern of the vowel part for each mora, standardized by the time length of the vowel part of each mora, classified according to the prosodic determining factors such as the number of mora of the accent phrase, the accent type, the phoneme sequence, etc. It is the same as FIG. 1 except that the accent phrase position fundamental frequency database 450 to be stored is replaced.

  The operation of the fundamental frequency pattern generator configured as described above will be described below.

  First, a character string to be converted to speech is input from the character string input unit 10. The character string analysis unit 20 analyzes the input character string, outputs phoneme information indicating the phoneme string to the time length setting unit 40, divides the character string into accent phrases, and sets the number of mora and accent type of each accent phrase, and The prosody information indicating the position of the accent phrase in the phrase and the phoneme information indicating the phoneme sequence are output to the fundamental frequency pattern generation unit 60.

  The time length setting unit 40 refers to the phoneme time length database 30 based on the phoneme information input from the character string analysis unit 20 and determines the vowel time length of each mora or the time length of a single vowel syllable, vowel sound or a vowel equivalent part in a long sound. And outputs the time length information to the fundamental frequency pattern generation unit 60. The basic frequency pattern generation unit 60 generates a basic frequency pattern for each accent phrase based on the prosody information and phoneme information input from the character string analysis unit 20 and the time length information input from the time length setting unit 40. In this example, generation of a fundamental frequency of a sentence composed of five accent phrases will be described.

  First, for the first accent phrase, a basic frequency pattern corresponding to the number of mora of the accent phrase for which the fundamental frequency is to be generated and the accent type, which is not the end of the sentence in the first accent phrase, is obtained from the accent phrase position fundamental frequency database 450. . Similarly, a fundamental frequency pattern is obtained from the accent phrase position fundamental frequency database 450 for the second accent phrase and the third accent phrase.

  As for the fourth accent phrase, since there is no fundamental frequency pattern corresponding to the fourth accent phrase in the accent phrase position fundamental frequency database 450, the third accent phrase whose accent phrase is closest to the fourth accent phrase is not at the end of the sentence. The basic frequency pattern corresponding to the number of mora and the accent type is obtained from the frequency pattern.

  As for the fifth accent phrase, which is the final accent phrase, there is no fundamental frequency pattern corresponding to the accent phrase position fundamental frequency database 450. Therefore, the number of mora from the fundamental frequency pattern at the end of the sentence of the third accent phrase having the closest accent phrase position is calculated. And the fundamental frequency pattern corresponding to the accent type. As in the third or fourth embodiment, a portion having no fundamental frequency pattern is interpolated on the real time axis to generate a fundamental frequency pattern.

  The vocal fold vibration generating unit 70 generates vocal fold vibration of the synthesized sound according to the fundamental frequency pattern output from the fundamental frequency pattern generating unit 60.

  By using the fundamental frequency pattern standardized by the vowel length of the mora, the variation of the fundamental frequency in the mora is reproduced in detail, and applied to the condition of the accent phrase, the condition of the end of the sentence or not, and the fundamental frequency of the phrase unit Can be accurately reproduced, so that a unity as a phrase can be formed and a natural sentence voice can be realized.

(Embodiment 10)
FIGS. 20A and 20B are schematic diagrams of a connection portion of a fundamental frequency pattern when a sentence is generated by connecting fundamental frequency patterns of accent phrases. The configuration of a fundamental frequency pattern generation device according to another embodiment of the present invention relating to the present invention is the same as that shown in FIG. The operation will be described below.

  First, a character string to be converted to speech is input from the character string input unit 10. The character string analysis unit 20 analyzes the input character string, outputs phoneme information indicating the phoneme sequence to the time length setting unit 40, divides the character string into accent phrases, and indicates the number of mora and the accent type of each accent phrase. Prosody information and phoneme information indicating a phoneme sequence are output to the fundamental frequency pattern generation unit 60. The time length setting unit 40 sets the time length of each phoneme with reference to the phoneme time length database 30 based on the phoneme information input from the character string analysis unit 20, and outputs the time length information to the fundamental frequency pattern generation unit 60. . The basic frequency pattern generation unit 60 generates a basic frequency pattern for each accent phrase based on the prosody information and phoneme information input from the character string analysis unit 20 and the time length information input from the time length setting unit 40.

  First, the number of mora of each accent phrase for which a basic frequency pattern is to be generated and the basic frequency pattern corresponding to the accent type are obtained from the mora time length standardized basic frequency database 50 and applied. According to the method of the sixth, seventh, or eighth embodiment, the fundamental frequency pattern acquired from the mora time length standardized fundamental frequency database 50 is modified for each accent phrase.

  Among the transformed fundamental frequency patterns of each accent phrase, the n-th accent phrase that is not the end of the sentence is shown in FIG. 20e) the fundamental frequency of the vowel part of the last mora of the accent phrase and the first mora of the n + 1-th accent phrase. Find the difference between the fundamental frequencies of the vowels.

  If there is no pause between the n-th accent phrase and the n + 1-th accent phrase, e) the fundamental frequency of the vowel part of the last mora of the relevant accent phrase and the fundamental frequency of the vowel part of the first mora of the n + 1-th accent phrase Is greater than 40 Hz and the accent nucleus of the n-th accent phrase is not in the last three moras in the accent phrase, the first mora of the accent phrase end reference point or the preceding mora and the (n + 1) -th accent The fundamental frequency pattern from the mora having a fundamental frequency exceeding the value obtained by subtracting 40 from the fundamental frequency of the vowel part of the first mora of the phrase to the final mora of the n-th accent phrase is compressed in the frequency axis direction, and FIG. , The connection between the n-th accent phrase and the (n + 1) -th accent phrase is smoothly connected. e) The difference between the fundamental frequency of the vowel part of the last mora of the accent phrase and the fundamental frequency of the vowel part of the first mora of the n + 1th accent phrase is 40 Hz or more, and the accent nucleus of the nth accent phrase is within the accent phrase. If the mora is within the last three moras, the mora having the fundamental frequency exceeding the value obtained by subtracting 40 from the fundamental frequency of the vowel part of the first mora of the n + 1 accent phrase in the mora of the accent kernel or the preceding mora, The fundamental frequency pattern up to the final mora of the n-th accent phrase is compressed in the frequency axis direction, and a smooth connection is made between the n-th accent phrase and the (n + 1) -th accent phrase.

  If there is a pause of less than 50 msec between the nth accent phrase and the (n + 1) th accent phrase, e) the fundamental frequency of the vowel part of the last mora of the accent phrase and the vowel part of the first mora of the (n + 1) th accent phrase If the fundamental frequency difference is 50 Hz or more, and the accent nucleus of the n-th accent phrase is not in the last three moras in the accent phrase, the n-th mora is the first morah of the accent phrase end reference point or a mora preceding it. A fundamental frequency pattern from a mora having a fundamental frequency exceeding a value obtained by subtracting 50 from a fundamental frequency of a vowel part of a first mora of the +1 accent phrase to a final mora of the nth accent phrase is compressed in the frequency axis direction. e) the difference between the fundamental frequency of the vowel part of the last mora of the accent phrase and the fundamental frequency of the vowel part of the first mora of the n + 1th accent phrase is 50 Hz or more, and the accent nucleus of the nth accent phrase is within the accent phrase If the mora is within the last three moras, the mora with the fundamental frequency exceeding the value obtained by subtracting 50 from the fundamental frequency of the vowel part of the first mora of the n + 1 accent phrase in the mora of the accent kernel or the preceding mora, The fundamental frequency pattern up to the last mora of the nth accent phrase is compressed in the frequency axis direction.

  If there is a pause of 50 msec or more and less than 100 msec between the nth accent phrase and the n + 1th accent phrase, e) the fundamental frequency of the vowel part of the last mora of the accent phrase and the first mora of the n + 1th accent phrase If the difference between the fundamental frequencies of the vowels is 70 Hz or more and the accent nucleus of the n-th accent phrase is not in the last three moras of the accent phrase, the first mora of the accent phrase end reference point or the mora preceding it , Compressing in the frequency axis direction a fundamental frequency pattern from a mora having a fundamental frequency exceeding a value obtained by subtracting 70 from the fundamental frequency of the vowel part of the head mora of the n + 1th accent phrase to the last mora of the nth accent phrase . e) the difference between the fundamental frequency of the vowel part of the last mora of the accent phrase and the fundamental frequency of the vowel part of the first mora of the n + 1th accent phrase is 70 Hz or more, and the accent nucleus of the nth accent phrase is within the accent phrase If the mora is within the last three moras, the mora having the fundamental frequency exceeding the value obtained by subtracting 70 from the fundamental frequency of the vowel part of the first mora of the n + 1 accent phrase in the mora of the accent kernel or the preceding mora, The fundamental frequency pattern up to the last mora of the nth accent phrase is compressed in the frequency axis direction.

  If there is a pause between 100 msec and less than 150 msec between the nth accent phrase and the n + 1th accent phrase, e) the fundamental frequency of the vowel part of the last mora of the accent phrase and the first mora of the n + 1th accent phrase If the difference between the fundamental frequencies of the vowels is 80 Hz or more and the accent nucleus of the n-th accent phrase is not in the last three moras in the accent phrase, the first mora of the accent phrase end reference point or a mora preceding it, A fundamental frequency pattern from a mora having a fundamental frequency exceeding a value obtained by subtracting 80 from the fundamental frequency of the vowel part of the head mora of the (n + 1) th accent phrase to the last mora of the nth accent phrase is compressed in the frequency axis direction. e) The difference between the fundamental frequency of the vowel part of the last mora of the accent phrase and the fundamental frequency of the vowel part of the first mora of the n + 1th accent phrase is 80 Hz or more, and the accent nucleus of the nth accent phrase is within the accent phrase. If the mora is within the last three moras, the mora having the fundamental frequency exceeding the value obtained by subtracting 70 from the fundamental frequency of the vowel part of the first mora of the n + 1 accent phrase in the mora of the accent kernel or the preceding mora, The fundamental frequency pattern up to the last mora of the nth accent phrase is compressed in the frequency axis direction.

  The vocal fold vibration generating unit 70 generates vocal fold vibration of the synthesized sound according to the fundamental frequency pattern output from the fundamental frequency pattern generating unit 60.

  By deforming the end of the fundamental frequency pattern generated for each accent phrase based on the pause length between the accent phrase and the subsequent accent phrase, the connection between the accent phrases can be smoothed, and a natural sentence voice can be realized.

  In the above description, the first, third, and fourth embodiments have been described using an example in which a straight line is used as an interpolation function, and the second embodiment uses a critical damping quadratic linear system with respect to a logarithmic frequency axis as the interpolation function. A second-order critical braking linear system may be used in the first, third, and fourth embodiments, and a straight line may be used in the second embodiment. Other functions on the real time axis can be similarly implemented.

  In the second embodiment, the fundamental frequency from the beginning of the accent phrase to the rising reference point is interpolated using a critical damping quadratic linear system on the logarithmic frequency axis, and is represented on the real time axis in the fourth embodiment. Although the interpolation is performed by applying the fundamental frequency pattern, the fundamental frequency pattern expressed on the real time axis may be applied to the second embodiment, and the critical damping quadratic linear system for the logarithmic frequency axis may be used in the fourth embodiment. .

  In the second embodiment, the vowel time length standardized fundamental frequency database 150a divides the time length of the vowel part of each mora into four equal parts and stores the representative value of the fundamental frequency in each section. Any other standardized time length for each phoneme may be used.

  In the second and fifth embodiments, the center of the third section obtained by dividing the vowel length of the mora by four is used as the accent rising reference point. However, any other relative position corresponding to the latter half of the vowel may be used. Value may be used.

  In the fifth embodiment, the vowel time length standardized fundamental frequency database 150a divides the time length of the vowel part of each mora into four equal parts and stores the representative value of the fundamental frequency in each section. Any other vowel may be used as long as it is standardized by the time length.

  In the second and fifth embodiments, the center of the third section obtained by dividing the vowel length of the mora by four is used as the accent rising reference point. However, any other relative position corresponding to the latter half of the vowel may be used. Value may be used.

  In the second and fifth embodiments, the center of the third section obtained by dividing the vowel length of the mora corresponding to the accent nucleus into four parts, and the center of the third section obtained by dividing the vowel length of the next mora following the accent nucleus into four parts Are set as falling reference points, but other values may be used as long as they are relative positions corresponding to the latter half of the vowel.

  In the second and fifth embodiments, the center of the second section obtained by dividing the vowel length of the last mora of the accent phrase into four is set as the accent phrase end reference point. Value may be used.

  In Embodiments 2 and 5, the center of the third section obtained by dividing the vowel length of the last mora of the speech into four equal parts is used as the ending reference point. However, any other value may be used as long as the relative position corresponds to the latter half of the vowel. good.

  In the fifth embodiment, the elementary pattern is generated in the same manner as in the second embodiment for the basic week on which the microprosody is added, but may be the same as in the first, third, and fourth embodiments.

  In the sixth embodiment, the fundamental frequency pattern of the accent phrase is generated in the same manner as in the second embodiment, but may be similar to the first, third, and fourth embodiments.

  In the sixth embodiment, interpolation is performed after the reference point of the fundamental frequency pattern is changed according to the deformation amount acquired from the database. However, the fundamental frequency pattern may be modified after the interpolation.

  In the sixth embodiment, as the amount of deformation of the fundamental frequency pattern, the difference between the maximum value and the end of the accent phrase is compressed to 90% in the first accent phrase, but other values within the range of 70% to less than 100% are used. But it's fine.

  In the sixth embodiment, the maximum value of the second accent phrase is compressed to 75%, and the maximum value of the third accent phrase and the n-th accent phrase is reduced to 70% as the amount of deformation of the fundamental frequency pattern. May be another value within the range of 50% to 90%.

  In the sixth embodiment, the difference between the maximum value and the end of the accent phrase in the second accent phrase is compressed to 70% as the amount of deformation of the fundamental frequency pattern, and the difference in the third accent phrase and the n-th accent phrase is the maximum. Although the difference between the value and the end of the accent phrase is compressed to 68%, another value within the range of 50% to 90% may be used.

  In the sixth embodiment, as the amount of deformation of the fundamental frequency pattern, the maximum value of the final accent phrase is compressed to 48%, but another value within the range of 30% to 70% may be used.

  In the sixth embodiment, as the amount of deformation of the fundamental frequency pattern, the difference between the maximum value and the end of the accent phrase is compressed to 60% for the final accent phrase, but other values within the range of 40% to 80% are compressed. But it's fine.

  In the seventh embodiment, the coefficient of i in Equation 1 is -0.1, but may be another value in the range of -0.05 to -0.4.

  In the seventh embodiment, the coefficient of j in Equation 2 is set to -0.05, but other values within the range of -0.2 with 0 as the maximum may be used.

  In Embodiments 7 and 8, in the final accent phrase, the maximum value of the fundamental frequency is set to a value obtained by reducing the maximum value of the immediately preceding accent phrase by 15%, but within a range of 10% to 40%. Other values may be used.

  Although the end of the accent phrase is set to a value obtained by reducing the end of the accent phrase of the immediately preceding accent phrase by 10%, another value within the range of 5% to 40% may be used.

  In Embodiment 8, the coefficient of i in Equation 3 is -0.02, but is not limited to this, and may be another value in the range of -0.01 to -0.2.

  In Embodiment 8, the coefficient of j in Equation 4 is -0.01, but is not limited to this, and may be another value in the range from -0.01 to -0.1.

  In the tenth embodiment, the fundamental frequency pattern acquired from the mora time length standardized fundamental frequency database 50 is modified in the same manner as in the sixth, seventh, or eighth embodiment. The fundamental frequency pattern may be obtained from the frequency database 450 based on the position of the accent phrase.

  In the tenth embodiment, when there is no pause between the n-th accent phrase and the (n + 1) -th accent phrase, the basics of the vowel center of the last mora of the n-th accent phrase and the vowel center of the first mora of the (n + 1) -th accent phrase are different. Although the fundamental frequency pattern is modified so that the frequency difference becomes 40 Hz or less, another value between 20 Hz and 60 Hz may be used.

  In the tenth embodiment, the duration of the pause between the n-th accent phrase and the (n + 1) -th accent phrase is less than 50 msec, and 50 msec or more and less than 100 msec, as the basis for changing the fundamental frequency of the accent phrase fall, the accent phrase end, and the ending. , 100 msec to less than 150 msec, and 150 msec or more, but may be classified into 1 to 8 other numbers.

  In the tenth embodiment, when the duration of the pause between the n-th accent phrase and the (n + 1) -th accent phrase is 150 msec or more, the accent phrase falls, and the fundamental frequency of the accent phrase end and the ending is not changed. However, the upper limit of the duration of the pause to be changed may be another value between 120 msec and 200 msec.

In the tenth embodiment, the duration of the pause between the n-th accent phrase and the (n + 1) -th accent phrase is classified into four stages as the basis for changing the fundamental frequency of the accent phrase falling edge, the accent phrase end, and the ending. The upper limit of the difference between the fundamental frequency of the center of the vowel part of the last mora of the n accent phrase and the center of the vowel part of the head mora of the (n + 1) th accent phrase is set for each step of the pause duration. (Equation 5)
(Equation 5)
at + b (Hz) ... Equation 5
However, 0 <a <0.4 20 <b <60
It may be set by setting.

  The present invention is realized by a program, and is recorded on a recording medium such as a floppy (registered trademark) disk, an optical disk, an IC card, a ROM cassette, or the like on which a program can be recorded, and is transferred. It can be easily implemented in a computer system.

  Further, in the above embodiment, the phoneme of the present invention has been described as mainly corresponding to a mora, but the present invention is not limited to this, and may be, for example, a syllable. That is, as described above, the fundamental frequency database is not limited to the case where data is stored in mora units or phoneme units. Of course, this is also good, and in this case, the same effect as described above is exerted. That is, in all of the above-described embodiments, the same effect as described above is exerted even if the configuration is such that “mora” is read as “syllable”.

  Further, in the above-described embodiment, a case has been described where the fundamental frequency database holds the fundamental frequency patterns from the end to 3 moras. However, if the fundamental frequency database holds the fundamental frequency patterns from the end to 4 moras at the maximum. Demonstrate sufficient effect.

  As described above, the first method of the present invention uses a phoneme time length standardized fundamental frequency database that stores a fundamental frequency pattern standardized by the time length of the phoneme of the mora for each mora position of the accent phrase, and For the mora containing the maximum of the fundamental frequency, the mora next to the accent kernel and the accent kernel, and one or more mora at the end of the accent phrase, the fundamental frequency pattern in each mora is referred to the above database. Is set, and for a section in which the basic frequency is not set from the database, a basic frequency pattern is generated by interpolating between the basic frequencies set from the database using a function on the real time axis.

  The second method uses a phoneme time length standardized basic frequency database that stores a fundamental frequency pattern standardized by the time length of the phoneme of the mora for each mora position of the accent phrase, and determines the maximum value of the fundamental frequency of the accent phrase. The rise reference point to give, the fall reference point to give the fall of the accent, the accent reference point to give the fundamental frequency at the end of the accent phrase, and the ending reference point to give the fundamental frequency at the end of the utterance to the vowel length of the mora. For each of the reference points, the fundamental frequency is set with reference to the database, and the fundamental frequency between these reference points is interpolated by a function on the real time axis. This is a basic frequency pattern generation method for generating a basic frequency pattern.

  The third method uses a phoneme time length standardized basic frequency database that stores a fundamental frequency pattern standardized by the time length of a vowel or a vowel equivalent part of the mora for each mora position of the accent phrase, and uses the basic frequency of the accent phrase. For each of the mora containing the maximum value of mora, the mora next to the accent nucleus and the accent nucleus, and one or more mora at the end of the accent phrase, the fundamental frequency pattern in each mora is set by referring to the database, In a section in which the fundamental frequency is not set from the database, the basic frequency pattern is generated by interpolating between the fundamental frequencies set from the database using a function on the real time axis to generate a fundamental frequency pattern.

  The fourth method uses a phoneme time length standardized fundamental frequency database storing a fundamental frequency pattern standardized by the time length of a vowel or a vowel equivalent part of the mora for each mora position of the accent phrase, and uses the fundamental frequency of the accent phrase. The rising reference point that gives the maximum value of, the falling reference point that gives the fall of the accent, the accent reference point that gives the fundamental frequency at the end of the accent phrase, and the ending reference point that gives the fundamental frequency at the end of the utterance Is set to a time point that is a fixed ratio to the vowel length of each, the fundamental frequency is set with reference to the database for each reference point, and the fundamental frequency between those reference points is a function on the real time axis. This is a basic frequency pattern generation method of generating a basic frequency pattern by interpolation.

  In the fifth method, a phoneme time length standardized fundamental frequency database storing a fundamental frequency pattern standardized by the phoneme time length of the mora for each mora position of the accent phrase, and a fundamental frequency for each phoneme or phoneme sequence are stored in the phoneme time. Using a microprosody database that stores the difference between the value standardized by the length and the fundamental frequency pattern, the fundamental frequency pattern is obtained by adding or subtracting the microprosody data to or from the fundamental frequency pattern obtained from the phoneme time length standardized fundamental frequency database. Is a basic frequency pattern generation method for generating

  The sixth method generates a fundamental frequency pattern for each accent phrase by using a phoneme time length standardized fundamental frequency database storing a fundamental frequency pattern standardized by a phoneme time length of the mora for each mora position of the accent phrase. In the basic frequency pattern generation method, if the basic frequency pattern corresponding to the mora number and accent type of the accent phrase for which the basic type fraction is to be generated is not in the phoneme time length standardized basic frequency database, the basic frequency pattern in the database is used. Then, the accent phrase for which the basic frequency is to be generated is n mora m type, the basic frequency pattern obtained from the database is 1 mora j type, the position of the mora including the maximum value of the obtained basic frequency pattern is i, and the obtained basic frequency is When the number of mora at the end of the accent phrase of the pattern is k, m ≦ i + 1 Then, from the first to the (m + 1) th mora, the first to j + 1th mora of the fundamental frequency pattern obtained from the database are applied, and from the (n−k + 1) th to the nth mora, the (l−k + 1) th of the fundamental frequency pattern obtained from the database is applied. To the first mora, and for the mora in between, a fundamental frequency pattern is generated by interpolating on the real time axis. When i + 1 <m ≦ nk + 1, the first to i-th mora of the basic frequency pattern obtained from the database are applied to the first to i-th mora, and the basic data obtained from the database are applied to the m-th and m + 1-th mora. The j-th and j + 1-th mora of the frequency pattern are applied, and from the (n-k + 1) -th to the n-th mora, the l-k + 1 to the l-th mora of the fundamental frequency pattern acquired from the database are applied. A fundamental frequency pattern is generated by interpolating on the time axis. When m> nk + 1, the first to i-th mora of the fundamental frequency pattern obtained from the database are applied to the first to i-th mora, and the basic frequency pattern obtained from the database is applied to the m-th to n-th mora. This is a fundamental frequency generation method of generating a fundamental frequency pattern by applying the j-th to the l-th mora and interpolating the mora between them on the real time axis.

  A seventh method is a basic frequency generation method for generating a basic frequency pattern using a basic frequency database in which the basic frequency pattern of an accent phrase is classified according to the position of the accent phrase of the phrase and whether or not it is the end of a sentence.

  The eighth method uses a fundamental frequency database that stores a fundamental frequency pattern of an accent phrase and a deformation database that stores the amount of deformation of the fundamental frequency pattern depending on the position of the accent phrase of the phrase and whether or not it is the end of the sentence. This is a basic frequency pattern generation method for generating a basic frequency pattern by deforming a basic frequency pattern obtained from basic frequency data according to a deformation amount obtained from a deformation database.

  A ninth method uses a fundamental frequency database storing fundamental frequency patterns of accent phrases, and transforms the fundamental frequency pattern obtained from the fundamental frequency data by a function of the position i of the accent phrase in the phrase. This is a basic frequency pattern generation method.

  Further, a tenth method uses a fundamental frequency database storing fundamental frequency patterns of accent phrases, and converts a fundamental frequency pattern acquired from the fundamental frequency data into a mora phrase used as a reference for determining the fundamental frequency pattern. This is a method of generating a fundamental frequency pattern in which the fundamental frequency pattern is deformed by a function of the position j in.

  In the eleventh method, a fundamental frequency pattern is generated for each accent phrase, and the difference between the frequency at the end and the end point of the accent phrase and the frequency at the start point of the next accent phrase is equal to or less than a predetermined value. This is a fundamental frequency pattern generation method that changes the characteristics of the fall of the accent, the end of the accent, and the end point of the accent phrase so that

  As described above, according to the present invention, the fundamental frequency pattern in the mora is applied by applying the fundamental frequency pattern in which the timing and angle of the rise of the accent phrase and the fall in the accent nucleus are standardized by the vowel length of the mora. In addition to realizing the high naturalness of the fluctuations of the data, the interpolation on the real-time axis to which the pattern of the database is not applied eliminates the sense of discontinuity when controlling in mora units. Can also be smaller. Alternatively, by setting the timing of the rise of the accent phrase and the fall of the accent nucleus on the time axis standardized by the vowel length of the mora, the timing of the fluctuation of the fundamental frequency in the mora is reproduced in detail, and the rise, By using the function on the real-time axis for the fall angle, it is possible to obtain a pattern with a stable fundamental frequency with a stable rise and fall without being affected by the difference in time length due to phonemes. Eliminates the sense of discontinuity when controlling with, and achieves high naturalness. Further, by using interpolation, the fundamental frequency pattern database can be made smaller, and its practical effect is great.

  The fundamental frequency pattern generation method and the like according to the present invention has an advantage that a fundamental frequency pattern with higher naturalness can be generated as compared with the related art, and is useful as a fundamental frequency pattern generation method and the like.

Functional block diagram of a fundamental frequency generation device according to the present invention and / or another invention related to the present invention (hereinafter, referred to as the present invention, etc.) FIG. 4 is a diagram illustrating an example of a fundamental frequency pattern generated according to the first embodiment of the present invention. FIG. 9 is a diagram illustrating an example of a fundamental frequency pattern generated according to the second embodiment of the present invention. Functional block diagram of an apparatus showing one embodiment of the present invention etc. FIG. 4 is a diagram illustrating an example of a fundamental frequency pattern according to the present invention. FIG. 4 is a diagram illustrating an example of a fundamental frequency pattern according to the present invention. Functional block diagram of an apparatus showing one embodiment of the present invention etc. Schematic diagram of microprosody components stored in microprosody database 250 (A): A diagram showing a fundamental frequency pattern generated from the fundamental frequency database of the fifth embodiment. (B): A diagram showing microprosody components obtained from the microprosody database of the same embodiment. (C): FIG. The figure which shows the fundamental frequency pattern generated by adding the pattern of FIG.9 (B) to the pattern of A). Functional block diagram of an apparatus showing an embodiment of the present invention (A), (B): A diagram showing an example of a fundamental frequency pattern according to the present invention. (A), (B): A diagram showing an example of a fundamental frequency pattern according to the present invention. (A), (B): A diagram showing an example of a fundamental frequency pattern according to the present invention. (A), (B): A diagram showing an example of a fundamental frequency pattern according to the present invention. Schematic diagram of the fundamental frequency pattern of the present invention Functional block diagram of an apparatus showing one embodiment of the present invention etc. FIG. 1 is a schematic diagram of a fundamental frequency pattern according to an embodiment of the present invention. Schematic diagram of a fundamental frequency pattern according to a modified example of the present invention. Schematic diagram of the fundamental frequency pattern of the present invention (A), (B): Schematic diagrams of the accent phrase connection part of the fundamental frequency pattern of the present invention.

Explanation of reference numerals

DESCRIPTION OF SYMBOLS 10 Character string input part 20 Character string analysis part 30 Phoneme time length database 40 Time length setting part 50 Mora time length standardized fundamental frequency database 60 Basic frequency pattern generation part 70 Vocal cord vibration generation part 150, 150a, 150b Vowel time length standardized fundamental frequency Database 250 Microprosody database 350 Fundamental frequency pattern deformation database 450 Accent phrase position fundamental frequency database

Claims (9)

A fundamental frequency pattern generation method for generating a fundamental frequency pattern of an accent phrase using a fundamental frequency database storing fundamental frequency patterns classified according to the number of phonemes and accent positions,
The basic frequency pattern corresponding to the number of phonemes and the accent type of the accent phrase for which the basic frequency pattern is to be generated is not stored in the basic frequency database, and furthermore, the basic frequency pattern is to be generated. If the accent position of the accent phrase is the same or earlier than the next phoneme position of the phoneme position including the peak of the fundamental frequency stored in the fundamental frequency database,
(1) A basic frequency pattern stored in the basic frequency database and having the same accent position as that of the accent phrase for which the basic frequency pattern is to be generated, and an accent for which the basic frequency pattern is to be generated. Using the fundamental frequency pattern stored in the fundamental frequency database, corresponding to the number of phonemes closest to the number of phonemes in the phrase,
(2) Applying a fundamental frequency pattern from the first phoneme to the next phoneme of the accent kernel by applying a fundamental frequency from the first phoneme of the fundamental frequency pattern stored in the fundamental frequency database to the next phoneme of the accent kernel. Generate
(3) The basic frequency from the accent nucleus to the phoneme immediately before the end of the accent phrase, which is a predetermined number of phonemes less than or equal to 4 from the second phoneme, is calculated as (a) in the fundamental frequency pattern stored in the fundamental frequency database. ) A fundamental frequency between a second phoneme from the accent kernel and the end of the accent phrase, or (b) a fundamental frequency between a phoneme next to the accent kernel and the end of the accent phrase, or (c) a second phoneme from the accent kernel. And (d) a fundamental frequency between the phoneme immediately before the end of the accent phrase and a fundamental frequency between the phoneme immediately before the end of the accent phrase and (d) a fundamental frequency between the phoneme immediately before the end of the accent phrase and the fundamental frequency of the phoneme immediately before the end of the accent phrase.
(4) By applying the fundamental frequency pattern at the end of the accent phrase of the accent phrase for which the fundamental frequency pattern is to be generated to the fundamental frequency of the phoneme at the end of the accent phrase of the fundamental frequency pattern stored in the fundamental frequency database. The method of generating the fundamental frequency pattern to be generated. A fundamental frequency pattern generation method for generating a fundamental frequency pattern of an accent phrase using a fundamental frequency database storing fundamental frequency patterns classified according to the number of phonemes and accent positions,
The basic frequency pattern corresponding to the number of phonemes and the accent type of the accent phrase for which the basic frequency pattern is to be generated is not stored in the basic frequency database, and further, the basic frequency pattern is to be generated. The accent position of the accent phrase is located after the phoneme position next to the phoneme position including the peak of the fundamental frequency stored in the fundamental frequency database and before the end of the accent phrase which is a predetermined number of phonemes of 4 or less. If you do
(1) It is a fundamental frequency pattern having an accent nucleus after the second phoneme from the peak of the fundamental frequency stored in the fundamental frequency database and before the end of the accent phrase, and further attempts to generate the fundamental frequency pattern. Using the fundamental frequency pattern stored in the fundamental frequency database, corresponding to the number of phonemes closest to the number of phonemes of the accent phrase,
(2) A basic frequency pattern from the first phoneme of the accent phrase for which the basic frequency pattern is to be generated to the phoneme including the peak of the basic frequency is obtained from the first phoneme of the basic frequency pattern stored in the basic frequency database. Generate by applying the fundamental frequency up to the phoneme including the peak of the fundamental frequency,
(3) The fundamental frequencies from the phoneme following the phoneme including the peak of the fundamental frequency to the phoneme immediately before the accent nucleus are expressed by: (a) the fundamental frequency peak stored in the fundamental frequency database; (B) the fundamental frequency of the phoneme including the peak of the fundamental frequency and the fundamental frequency of the phoneme immediately preceding the phoneme including the accent kernel, or (c) the fundamental frequency of the phoneme including the accent kernel. A fundamental frequency of a phoneme next to the phoneme including the frequency peak and a phoneme including the accent kernel, or (d) a phoneme next to the phoneme including the peak of the fundamental frequency and a phoneme immediately before the phoneme including the accent kernel. Generated by interpolation with the fundamental frequency
(4) The basic frequencies of the phoneme including the accent nucleus of the accent phrase for which the basic frequency pattern is to be generated and the phoneme immediately following it are obtained by combining the phoneme including the accent nucleus of the basic frequency pattern stored in the basic frequency database with the phoneme. Generated by applying the fundamental frequency with the phoneme immediately after,
(5) The fundamental frequency from the accent nucleus to the phoneme immediately before the end of the accent phrase, which is the predetermined number of phonemes of 4 or less from the second phoneme, is calculated by using the basic frequency pattern stored in the fundamental frequency database as ( a) the fundamental frequency of the second phoneme from the accent kernel and the end of the accent phrase, or (b) the fundamental frequency of the next phoneme of the accent kernel and the end of the accent phrase, or (c) the fundamental frequency of 2 from the accent kernel. A fundamental frequency between a phoneme and a phoneme immediately before the end of the accent phrase, or (d) a fundamental frequency between a phoneme next to the accent kernel and the phoneme immediately before the end of the accent phrase, and generated by interpolation.
(6) applying the fundamental frequency pattern at the end of the accent phrase of the accent phrase for which the fundamental frequency pattern is to be generated to the fundamental frequency of the phoneme at the end of the accent phrase of the fundamental frequency pattern stored in the fundamental frequency database. The method of generating the fundamental frequency pattern generated by. A fundamental frequency pattern generation method for generating a fundamental frequency pattern of an accent phrase using a fundamental frequency database storing fundamental frequency patterns classified according to the number of phonemes and accent positions,
If the fundamental frequency pattern corresponding to the number of phonemes and the accent type of the accent phrase for which the fundamental frequency pattern is to be generated is not stored in the fundamental frequency database, and the accent for which the fundamental frequency pattern is to be generated is When the accent position of the phrase is included in the phoneme at the end of the accent phrase having a predetermined number of phonemes of 4 or less,
(1) The basic position stored in the fundamental frequency database, wherein the accent position within the accent phrase end and the accent position within the accent phrase end of the accent phrase for which the fundamental frequency pattern is to be generated are the same. Using a fundamental frequency pattern stored in the fundamental frequency database, which is a frequency pattern and corresponds to the number of phonemes closest to the number of phonemes of the accent phrase to generate the fundamental frequency pattern,
(2) A basic frequency pattern from the first phoneme of the accent phrase for which the basic frequency pattern is to be generated to the phoneme including the peak of the basic frequency is obtained from the first phoneme of the basic frequency pattern stored in the basic frequency database. Generate by applying the fundamental frequency up to the phoneme including the peak of the fundamental frequency,
(3) The fundamental frequencies from the phoneme following the phoneme including the peak of the fundamental frequency to the phoneme immediately before the accent nucleus are expressed by: (a) the fundamental frequency peak stored in the fundamental frequency database; (B) the fundamental frequency of the phoneme including the peak of the fundamental frequency and the fundamental frequency of the phoneme immediately preceding the phoneme including the accent kernel, or (c) the fundamental frequency of the phoneme including the accent kernel. A fundamental frequency of a phoneme next to the phoneme including the frequency peak and a phoneme including the accent kernel, or (d) a phoneme next to the phoneme including the peak of the fundamental frequency and a phoneme immediately before the phoneme including the accent kernel. Generated by interpolation with the fundamental frequency
(4) The fundamental frequencies from the phoneme including the accent nucleus of the accent phrase for which the fundamental frequency pattern is to be generated to the final phoneme of the accent phrase are calculated from the phoneme including the accent nucleus of the fundamental frequency pattern stored in the fundamental frequency database. A method of generating a fundamental frequency pattern by applying a fundamental frequency up to the final phoneme of an accent phrase. A fundamental frequency pattern generation method for generating a fundamental frequency pattern of an accent phrase using a fundamental frequency database storing fundamental frequency patterns classified according to the number of phonemes and accent positions,
The basic frequency pattern corresponding to the number of phonemes and the accent type of the accent phrase for which the basic frequency pattern is to be generated is not stored in the basic frequency database, and further, the basic frequency pattern is to be generated. If the accent type of the accent phrase is flat,
(1) Using a fundamental frequency pattern stored in the fundamental frequency database, which corresponds to the number of phonemes closest to the number of phonemes of the accent phrase for which the basic frequency pattern is to be generated with the flat type accent,
(2) The fundamental frequency pattern from the first phoneme to the phoneme including the peak of the fundamental frequency is defined as the fundamental frequency from the first phoneme of the fundamental frequency pattern stored in the fundamental frequency database to the phoneme including the peak of the fundamental frequency. Apply and generate,
(3) The basic frequencies from the phoneme following the phoneme including the peak of the fundamental frequency to the end of the accent phrase or immediately before the final phoneme are calculated using the basic frequency pattern stored in the fundamental frequency database, A fundamental frequency of a phoneme including a frequency peak and the end of the accent phrase or the final phoneme, or (b) a phoneme including a peak of the fundamental frequency and a phoneme at the end of the accent phrase or immediately before the final phoneme. Or (c) a fundamental frequency between the phoneme following the phoneme including the peak of the fundamental frequency and the end of the accent phrase or the final phoneme, or (d) a phoneme including the peak of the fundamental frequency. The next phoneme of the, the fundamental frequency of the phoneme at the end of the accent phrase or immediately before the final phoneme, generated by interpolation,
(4) The end of the accent phrase of the accent phrase for which the fundamental frequency pattern is to be generated or the fundamental frequency pattern of the final phoneme is stored in the fundamental frequency pattern stored in the fundamental frequency database. A method of generating a fundamental frequency pattern by applying a fundamental frequency of a phoneme. The method according to claim 1, wherein the fundamental frequency pattern is extracted from a naturally uttered voice. 5. The method according to claim 1, wherein the interpolation is linear interpolation. The basic frequency pattern generation method according to any one of claims 1 to 4, wherein a basic frequency from a head of the accent phrase to a peak of the basic frequency is interpolated by a basic frequency pattern expressed on a real time axis. The method according to claim 1, wherein the phoneme is a mora or a syllable. A program recording medium on which a program for causing a computer to execute each step of the fundamental frequency pattern generation method according to claim 1 is recorded.

Images