JP2001166788A - Voice synthesis method and apparatus - Google Patents

Abstract

PROBLEM TO BE SOLVED: To solve the problem that a pitch pattern can not be extracted since an original natural voice has no pitch at its voiceless part when a pitch pattern is extracted from a natural voice and used to apply a different pronounced voice to synthesis. SOLUTION: A voice from which a pitch pattern is extracted is composed of a voice which is all composed of a voiced sound and when it is difficult to compose a sentence with all of a voiced sound, it is made possible to constitute the sentence all composed of the voiced sound with ease by using a composite meaningless document.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech synthesis method, and more particularly to a speech synthesis method for converting text into speech and an apparatus therefor.

[0002]

2. Description of the Related Art Conventionally, in a speech synthesizing method for synthesizing speech by connecting speech segments, as described in Japanese Patent Laid-Open No. 8-44391, a global pitch pattern is obtained from speech including unvoiced sound. The voice was extracted and used as a control target pitch pattern to synthesize speech.

[0003]

However, in the conventional speech synthesizing method, a large fluctuation appears in a pitch pattern of a voiced portion near an unvoiced sound, and a pitch pattern needs to be estimated for an unvoiced portion by interpolation. . For this reason, when vowel sections such as “sa” and “sha” are continuous and the vowel section is very short, there is not enough voiced section that has not received much fluctuation in pitch due to unvoiced consonants. When it is difficult to obtain a global pitch pattern or when vowels are unvoiced, it is necessary to obtain all pitch patterns of the syllables by interpolation processing, and the inflection of synthesized speech becomes unnatural. there were.

[0004] The present invention is to provide a speech synthesis method and an apparatus therefor which eliminate the unnaturalness of the inflection of the synthesized speech and improve the quality of the synthesized speech.

[0005]

In order to solve the above-mentioned problems, a speech synthesis method according to the present invention extracts a pitch pattern of a speech to be synthesized from natural speech, and according to the pitch pattern, calculates a pitch of a speech piece. When combining these and synthesizing an arbitrary speech, a meaningful natural speech composed of only voiced sounds is used as a natural speech for extracting a pitch pattern. By using pitch patterns extracted from meaningful natural voices composed only of voiced sounds, all voices can be corrected without correcting pitch patterns near unvoiced sounds or performing interpolation processing of pitch patterns in unvoiced parts. A stable pitch pattern extracted from natural speech in a section can be obtained.

A speech synthesis method according to the present invention is characterized in that a grammatical meaningless sentence is used as a natural speech for extracting a pitch pattern. The speech synthesis method is based on the pitch pattern extracted from the natural speech. When synthesizing a sentence, if the sentence has meaning, the pitch pattern will fluctuate due to the effect of emphasizing parts depending on the meaning of the sentence. By preparing voices composed of all voiced sounds with different mora number and accent type for each phrase, it is possible to easily compose sentences composed of voiced sounds by combining these in any pitch pattern. Become.

Further, in the speech synthesis method of the present invention, when it is not possible to prepare a voice composed of voiced sounds having the same number of phrases, mora and accent type as the voice to be synthesized, the number of phrases, mora number and accent type are similar. By changing a part of the pitch pattern of the voiced sound that is being voiced, the number of desired accent phrases, the number of mora in the accent phrase, and the accent type pitch pattern to be synthesized are created, and Synthesizes the voice of
When preparing voices consisting of voiced sounds for the number of all accent phrases, the number of mora in accent phrases, and the accent type to be synthesized, all mora numbers and accent type voices for all accent phrases are stored in a database. If you try to prepare it, 127,550 voices (50 + 50 × 50 + 50 × 50 × 50) will be needed even if the number of mora of each accent phrase is up to 9 mora for voices up to 3 phrases. Although the scale becomes extremely large, a stable pitch pattern is first obtained using meaningful and familiar voices such as meaningful words even if the number of mora is slightly different, and this is corrected to If a pitch pattern is created, an arbitrary pitch pattern can be obtained.

Further, the speech synthesizing apparatus of the present invention determines the number of phrases, the number of mora, and the accent type from the input means for inputting the reading of the voice to be synthesized,
A pitch pattern category determining means for determining a category of the pitch pattern, a pitch pattern database storing pitch patterns extracted from sentences of natural speech,
Pitch pattern determining means for selecting the pitch pattern of the pitch pattern category determined by the pitch pattern category determining means from the pitch pattern database; and pitch pattern matching the number of mora and accent type of each accent phrase of the voice to be synthesized. A voice pattern selecting means for selecting a voice fragment according to the reading to be synthesized; a voice fragment database storing the voice fragments; and A speech piece transformation means for transforming based on the pitch pattern determined by the pattern decision means, speech piece connection means for connecting speech pieces to synthesize speech, and synthetic speech output means for outputting synthesized speech. Yes, by synthesizing speech according to the pitch pattern extracted from natural speech, And thus capable of synthesizing a speech pitch pattern.

Hereinafter, an embodiment of the present invention will be described with reference to FIG.
This will be described with reference to FIG. (Embodiment 1) The invention described in claim 1 of the present invention will be specifically described. FIG. 1 is a conceptual diagram when a pitch pattern is extracted from a voice including an unvoiced sound. In FIG. 1, reference numeral 100 denotes a voice waveform, 101, 102, and 103 denote pitch patterns, and 104 to 108 denote vowel sections for each mora position. The voice waveform 100 is a voice uttered as “Aken Ken”, and the unvoiced sound portions are “ki”, “ta”, “ke”.
This is the consonant part and the unvoiced vowel part of "ki". To extract a global pitch pattern from natural speech, 104-
It is important to find a stable pitch pattern in the vowel section for each 108 mora positions. However, as shown in the conceptual diagram of FIG. 1, the pitch pattern near the unvoiced sound has large fluctuations, making it difficult to obtain a globally stable pitch. More specifically, the pitch near an unvoiced consonant tends to slightly increase as in the pitch pattern 102 and the pitch pattern 103, and the voiced sound portion sandwiched between non-plosive sounds as in the pitch pattern 102 is short. Therefore, there arises a problem that a stable portion cannot be sufficiently secured. Further, when the vowel part is unvoiced as in the section 105, the pitch pattern of this part needs to be obtained from the preceding and following pitches by interpolation, and the pitch pattern cannot be directly obtained from natural speech. Occurs.

However, the above problem can be solved by extracting a pitch pattern from a voice composed of all voiced sounds. FIG. 2 is a conceptual diagram in the case where pitch patterns are extracted from voiced sounds. In FIG. 2, 200
Is a voice waveform uttered as "Aomori", 201 is a pitch pattern, and 202 to 205 are vowel sections for each mora position. As shown in the pitch pattern 201, when no unvoiced sound is included, the pitch pattern is continuous over the entire word. Therefore, a stable pitch can be obtained globally.

(Embodiment 2) The invention described in claim 2 of the present invention will be specifically described. FIG. 3 shows a state in which a grammatical meaningless sentence for extracting a pitch pattern according to claim 2 is created. FIG. 3 shows the creation of a grammatical meaningless sentence composed of three accent phrases.
301 is a first accent phrase section, 302 is a second accent phrase section, and 303 is a third accent phrase section. At this time, if a noun word as shown in FIG. 4 is inserted into the noun substitution areas 304 and 305 and a verb as shown in FIG. 5 is inserted into the verb substitution area 306, a grammatical meaningless sentence composed entirely of voiced sounds can be easily obtained. , And an arbitrary mora number and accent type can be set in each accent phrase. Even in the case of two or four or more accent phrases, a syntactically meaningless sentence composed of an arbitrary number of mora and an accent type accent phrase can be created by the same operation. Since the syntactically meaningless sentence created in this way has no meaning at all, it is possible to obtain a pitch pattern having no change in pitch pattern due to emphasis or the like.

(Embodiment 3) The invention described in claim 3 of the present invention will be specifically described. FIG. 6 is an example of a creation table for creating a voice for extracting a pitch pattern according to the third aspect. This table shows an example of a voice composed of three accent phrases. In FIG. 6, the accent types are classified into three types: 0 type, 1 type, and n type. The state of this classification is shown in FIG. In FIG. 7, 701 to 7
03 represents a pitch pattern. The n-type accent indicates an accent type after the type 2 accent.
These voices are replaced for each accent phrase, and a total of 27 (3 × 3 × 3) sentence pitch patterns are prepared. For example, to find a pitch pattern of a voice composed of three accent phrases, "Tomorrow's weather is rain.", "Tomorrow" is type 0, and "Weather is" and "rain is." Therefore, the sentence "Lion is drinking mankind" is selected from FIG. However, since "morrow" and "lion" have different mora numbers, the pitch pattern is modified and used. "Tomorrow" and "Lion are" are type 0 accents, so 701 in FIG.
It becomes a pitch pattern like Since the 0-type accent has little change in the pitch pattern after the second mora, the expansion / contraction process is performed during this period (section indicated by the dotted line). Similarly, in the case of the type 1 accent, expansion / contraction processing is performed in a section indicated by a dotted line 702 in FIG. In the case of an n-type accent, expansion and contraction processing is performed in the same manner. In the above description, the case of a sentence composed of three accent phrases has been described as an example. However, the pitch pattern can be similarly obtained for a sentence composed of other accent phrases.

[0013] By generating a pitch pattern from natural speech in this way, an arbitrary pitch pattern can be generated from a very small number of speech databases.

(Embodiment 4) The invention described in claim 4 of the present invention will be specifically described. FIG. 8 is a block diagram of the voice synthesizing device according to claim 4, wherein the reading input means 800 for inputting the reading of the voice to be synthesized, the number of accent phrases, the number of mora, The pitch pattern category determining means 801 that determines the accent type and thereby determines the category of the pitch pattern, the pitch pattern database 808 storing pitch patterns extracted from sentences of natural speech, and the pitch pattern category determining means 801 Pitch pattern database 80 for pitch patterns in the pitch pattern category
8 and a pitch pattern deforming means 803 for transforming the pitch pattern obtained by the pitch pattern determining means 802 so that the mora number matches the accent type for each accent phrase. Voice segment selecting means 804 for selecting a voice segment according to the reading to be read, and a voice segment database 80 storing voice segments
9, a speech piece deforming means 805 for transforming the speech piece selected by the speech piece selecting means 804 based on the pitch pattern transformed by the pitch pattern deforming means 803, and a speech piece deformed by the speech piece deforming means 805. And a voice synthesis unit 806 for outputting synthesized voice.

As described in the third embodiment, the pitch pattern database 808 stores pitch patterns in which the accent type of each accent phrase is categorized into 0 type, 1 type, and n type. At this time, 363 ways (3 + 9 + 27 + 81) even when considering sentences up to five accent phrases
+243), so that a small-scale database can be obtained. The pronunciation input from the reading input means 801 is analyzed by the pitch category determining means 802 to determine which of the above categories the accent type of each accent phrase is classified into, and the pitch pattern determining means 802 forms one voice for the entire sentence. Select the pitch pattern extracted from. Since the pitch pattern determined by the pitch pattern determining means 802 differs in the number of moras in each accent phrase, the pitch pattern is changed by the pitch pattern deforming means 803 so that the number of moras matches. At this time, the pitch pattern of the n-type accent is changed so as to match the accent type. Based on the pitch pattern changed by the pitch pattern deforming unit 803, the voice unit selecting unit 804 selects a voice unit from the voice unit database 809, deforms it by the voice unit deforming unit 805, and connects it by the voice unit connecting unit 806. And synthesizes the sound, and outputs the synthesized sound from the synthesized sound output unit 807.

As described above, according to the present embodiment, the speech is synthesized in accordance with the pitch pattern extracted from the natural speech, so that the speech of the stable pitch pattern in the whole sentence can be obtained by using a small-scale pitch pattern database. Can be synthesized.

[0017]

As described above, according to the present invention, speech is synthesized in accordance with a pitch pattern extracted from a natural speech composed only of voiced sounds, whereby a pitch pattern appearing near an unvoiced sound is synthesized. Since large fluctuations can be suppressed, and pitch pattern interpolation processing is eliminated, and a stable global pitch pattern can be obtained as it is, a natural inflection sound can be synthesized. Further, in the speech synthesis of a sentence, there is obtained an advantageous effect that it is possible to provide a speech synthesis device in which the pitch pattern is stable over the entire sentence.

[Brief description of the drawings]

FIG. 1 is a conceptual diagram of a pitch pattern of a voice including an unvoiced sound according to the first embodiment.

FIG. 2 is a conceptual diagram of a pitch pattern of a voice composed of only voiced sounds according to the first embodiment.

FIG. 3 is a conceptual diagram of creating a grammatical meaningless sentence according to the second embodiment.

FIG. 4 is a list showing examples of nouns composed only of voiced sounds for each mora number and accent type according to the second embodiment.

FIG. 5 is a list showing an example of a verb composed of only voiced sounds for each mora number and accent type according to the second embodiment.

FIG. 6 is a list showing an example of pitch pattern extraction voice creation according to the third embodiment.

FIG. 7 is a conceptual diagram of pitch pattern categorization in a third embodiment.

FIG. 8 is a block diagram showing a configuration of a speech synthesizer using a pitch pattern database according to a fourth embodiment.

[Explanation of symbols]

 100 voice waveform 101-103 pitch pattern 104-108 vowel section 200 voice pattern 201 pitch pattern 202-205 vowel section 301-303 accent phrase section 304, 305 noun substitution area 306 verb substitution area 701 ... 703 pitch pattern 800 reading input means 801 pitch pattern category determining means 802 pitch pattern determining means 803 pitch pattern deforming means 804 voice unit selecting means 805 voice unit deforming means 806 synthetic unit connecting means 807 … Synthesized sound output means 808… pitch pattern database 809… voice segment database

Continuation of the front page (72) Inventor Ryo Mochizuki 4-3-1 Tsunashima Higashi, Kohoku-ku, Yokohama-shi, Kanagawa Prefecture F-term in Matsushita Communication Industrial Co., Ltd. 5D045 AA09

Claims (4)

[Claims] 1. A method of extracting a pitch pattern of a speech to be synthesized from natural speech, changing a pitch of a speech piece according to the pitch pattern, and connecting these to synthesize an arbitrary speech. A speech synthesis method characterized by using a meaningful natural speech composed of only voiced sounds as a natural speech for extracting a pattern. 2. The speech synthesis method according to claim 1, wherein a grammatical meaningless sentence is used as natural speech for extracting a pitch pattern. 3. The same number of phrases as the speech to be synthesized,
When a voice composed of voiced sounds of the number of mora and accent type cannot be prepared, a part of a pitch pattern of a voice composed of voiced sounds having a similar number of phrases, mora number and accent type is changed to obtain a desired accent phrase. 2. The speech synthesis method according to claim 1, wherein an arbitrary speech is synthesized using the number of characters, the number of mora in an accent phrase, and an accent type pitch pattern to be synthesized. 4. Input means for inputting a reading of a speech to be synthesized, pitch pattern category determining means for determining the number of phrases, mora number and accent type from the input, and thereby determining the category of the pitch pattern, A pitch pattern database storing pitch patterns extracted from sentences of natural speech, and pitch pattern determining means for selecting a pitch pattern of the pitch pattern category determined by the pitch pattern category determining means from the pitch pattern database are to be synthesized. Pitch pattern transformation means for transforming the pitch pattern so as to match the mora number and accent type of each accent phrase of speech, speech piece selection means for selecting speech pieces according to the reading to be synthesized, and speech pieces storing speech pieces Database and selected by the voice segment selecting means A speech piece deformation means for transforming the speech piece based on the pitch pattern determined by the pitch pattern determination means, a speech piece connection means for connecting the speech pieces to synthesize a voice, and a synthesized sound output means for outputting a synthesized voice; A speech synthesizer provided with.