JPH0887296A - Voice synthesizer - Google Patents

Abstract

PURPOSE: To correct the fault of a point pitch model, to generate a natural pitch pattern and to produce a high quality synthesized sound in a voice synthesizer. CONSTITUTION: The voice synthesizer is provided with a first database section 21 which stores accent type point pitch patterns, a second database section 24 which stores the rule corresponding to the combination of the accent type and phoneme and a pitch pattern generating section 8 which retrieves a point pitch pattern corresponding to the paragraph, that is an object of the voice synthesis, from the section 21 and generates a new point pitch pattern based on the rule of the section 24 from the obtained point pitch pattern. Thus, pitches are given to not only the vowel centroid point but also to a phoneme boundary and the naturality in a head word in a synthesized sound and the vowel chain in an accent kernel is improved.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech synthesizer,
In particular, it relates to a technique for generating a pitch pattern in a speech synthesizer. The speech synthesizer of the present invention is used, for example, as a speech synthesizer in security equipment.

[0002]

2. Description of the Related Art Conventionally, in pitch control of speech rule synthesis, a "point-pitch model" which gives a pitch frequency at the center position of a vowel and linearly interpolates between vowels (Japanese Patent Laid-Open No. 50-128)
40, see Japanese Patent No. 1087848). Speech synthesis using the point-pitch model is a numerical digitization of several patterns corresponding to Japanese accent types in advance, which is digitized based on the accent type information of the phrase to be synthesized. In this method, the point pitch pattern is searched and used as a synthesis parameter.

The point-pitch model is a model utilizing the characteristic that the pitch frequency at the energy center position of a vowel is dominant for human hearing, and can be said to be an efficient idea for pitch control.

[0004]

However, in natural utterance, even if the utterances of the same accent type and the same number of mora are used in the natural utterance, depending on the arrangement of the phonemes constituting the utterances,
Obviously, there may be a pitch pattern different from the point pitch pattern. This is due to the interaction of the pitch and articulatory mechanisms in the natural vocal system, resulting in a audible accent type.
This is because it does not always match the shape of the dot pitch pattern.

As a result, depending on the phrase, the retrieved point pitch pattern does not always match the original accent of the phrase, the accent is emphasized strangely, and the intonation may be strange. It was An object of the present invention is to improve the above-mentioned drawbacks of the conventional point-pitch model, generate a natural pitch pattern, and create a high-quality synthesized sound.

[0006]

In order to solve the above problems, the present invention expands the conventional interpretation of the point pitch pattern to provide a model in which not only the vowel center of gravity points but also the phonological boundaries are given pitches. This is called an extension point pitch model, and the given pitch is called an extension point pitch. The extension point pitch is determined by a rule from the point pitch given to the position of the center of gravity of the vowel and the arrangement of phonemes, and is used as a synthesis parameter by determining the pitch of the phoneme boundary (hereinafter referred to as “boundary pitch”).

In order to realize this, the present invention realizes, in a speech synthesizer, a first database section storing a point pitch pattern of an accent type and a second database section storing a rule corresponding to a combination of an accent type and a phoneme. The database unit and the point pitch pattern corresponding to the speech synthesis target phrase are searched from the first database unit, and a new point pitch pattern is obtained from the obtained point pitch pattern based on the rules of the second database unit. A pitch pattern generation unit for generating is provided.

Further, according to the present invention, the rule stored in the second database section can correspond to a combination of a combination of accent type and phoneme and a mora number. Further, after adding the descending component to the point pitch pattern stored in the first database section, the rule of the second database section can be applied to generate a new point pitch pattern.

[0009]

By applying the rule of the second database section to the point pitch pattern obtained by searching from the first database section, the boundary pitch is determined by the rule from the given arrangement of point pitch and phoneme. , Used as a synthesis parameter. As a result, not only the vowel center point of the point pitch pattern but also the phonological boundaries are given pitches, which improves the naturalness of the beginning of a synthesized voice and the vowel chain in the accent nucleus.

Further, the rule stored in the second database is made to correspond to the combination of the accent type and the phoneme combination and the mora number, and further, by adding the descending component to the point pitch pattern, the accent is added. The naturalness of the vowel chain in the pattern can be further improved.

[0011]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 2 is a diagram showing the configuration of the Japanese speech synthesis system according to the embodiment of the present invention. In the figure, the text input device 1 receives a synthetic text for which voice synthesis is to be performed.

The text analysis unit 2 analyzes this synthetic sentence using the pronunciation information generation rule 3, and adds phoneme information such as accent information, pauses, and vowel devoicing necessary for speech synthesis processing. Convert to a column. The phoneme duration generating unit 4 determines the phoneme duration of the phoneme symbol string generated by the text analyzing unit 2 according to the rhythm rule 5. The output of the phoneme duration generation unit 4 is input to the sound source amplitude pattern generation unit 6, the pitch pattern generation unit 8, and the spectrum pattern generation unit 11.

The sound source amplitude pattern generator 6 determines the power envelope of the voice according to the power rule 7. The details of the sound source amplitude pattern generation unit 6 and the power rule 7 are described in Japanese Patent Application No. 5-247994 filed by the present applicant. The pitch pattern generation unit 8 determines point pitch patterns for each accent phrase from the prosody control rule 9 and interpolates them to generate continuous point pitch patterns.

The sound source generator 10 generates a sound source based on the power pattern and the pitch pattern. The spectrum pattern generation unit 11 searches the phonetic synthesis basic unit database 13 from the phoneme types such as vowels and consonants according to the phonological improvement rule 12, and combines the spectra of each phoneme according to the phonological combination rule 14 to create a formant pattern. To do.

The voice synthesizer 15 includes the sound source information obtained from the sound source generation unit 10 and the spectrum pattern generation unit 11.
Create a synthetic speech from the formant pattern obtained from. The created synthetic voice is uttered to the outside by the speaker 16. Next, the pitch pattern generator 8 and the prosody control rule 9, which are the features of the present invention, will be described with reference to FIG.

The prosody control rule 9 includes a first database 21 storing a normalized point pitch pattern, a descending component rule within accent phrase 22, a descending component rule between accent phrases 23, and a second database 24 storing modification rules. It is equipped with. First, the general operation of the pitch pattern generation unit 8 will be described. The pitch pattern generation unit 8 searches the normalized point pitch pattern stored in the first database 21 of the prosody control rule 9 for one normalized point pitch pattern determined by the mora number and accent type of the phrase to be synthesized. Yes (step S1).

Next, the pitch pattern generating unit 8 gives the descending component within the accent phrase to the retrieved point pitch pattern using the descending component within accent phrase 22 (step S2), and , The inter-accent phrase descending component is added using the accent phrase descending component rule 23 (step S3). Finally, the pitch pattern generation unit 8
Is the second database 24 for the point pitch pattern.
Using the rules stored in step S4, the accent type of the point pitch pattern and the phoneme are combined and the transformation corresponding to the number of mora is performed (step S4).

Each process briefly described above will be described in detail below. The normalized point pitch pattern stored in the first database 21 will be described. The normalized point pitch pattern is obtained by analyzing the point pitch of natural utterance in advance and normalizing it to create a database. 3 to 6 show examples of normalized point pitch patterns stored in the first database 21, which are collectively displayed for each accent type N. 3 is a 0-type normalized point pitch pattern, FIG. 4 is a 1-type normalized point pitch pattern,
FIG. 5 shows a type 2 normalized point pitch pattern, and FIG. 6 shows a type 3 normalized point pitch pattern. Although illustration is omitted below, in the present example, up to the 9-type normalized point pitch pattern is prepared.

Here, the accent type N represents the position of the accent nucleus. The accent nucleus is the mora just before the frequency drops when an accent is applied. That is,
Type 1 in FIG. 4 means that the accent nucleus is the first mora,
Type 2 in Fig. 5 means that the accent nucleus is the second mora,
The type 3 in FIG. 6 means that the accent nucleus is the third mora,
Type 0 in FIG. 3 means that there is no accent nucleus.

3 to 6 show a plurality of patterns with different mora numbers M for one accent type N. Generally, each pattern is called M-Mora N type. As is clear from the above explanation, the types of patterns are
Mora number M and accent type N (always M>N> =
It is the sum of the number of combinations of 0). Usually M is at most 1
Since <M <10, the total number of patterns does not exceed the sum (42) of 1 to 10.

In step S1 of FIG. 1, the pitch pattern generation unit 8 creates one normalized point pitch pattern determined from the number of mora of the phrase to be synthesized and the accent type from the plurality of stored normalized point pitch patterns. Search for. Next, the application of the pitch lowering component in steps S2 and S3 of FIG. 1 will be described.

The point pitch pattern analyzed from the natural utterance has a gradual pitch down component due to a decrease in expiratory pressure. Therefore, at the time of synthesis, a falling component is given to the point pitch pattern in order to bring it closer to a natural vocal sound. The pitch pattern generation unit 8 adds the descending component in the accent phrase to the point pitch pattern extracted by the search using the accent component descending component rule 22.

For patterns other than the 1st type, the descending component is shaped so that it becomes a sag to the right with reference to the 1st mora as shown in FIG. In the case of the 1st type, shift is performed so that the average pitch becomes 0 at the position of M / 2 at the center of the M mora. The dynamic range of each type should be a constant value. Further, the pitch pattern generation unit 8 uses the inter-accent phrase descending component rule 23 to attach the descending component between accent phrases to the point pitch pattern to which the in-accent phrase descending component is attached, as shown in FIG. To do.

Next, application of the transformation rule in step S4 of FIG. 1 will be described. The pitch pattern generation unit 8 uses the transformation rules stored in the second database 24 for the point pitch patterns created up to step S3, and newly creates a phonological boundary on the basis of the value of the obtained point pitch. The pattern is transformed by giving a different pitch or deleting a certain pitch from the point pitch pattern.

The deformation method will be described with reference to FIGS. In the following description, V is a vowel and N is a sound repellency. 1. An example of the expansion method at the beginning of a word will be described with reference to FIG. The rule of this example is applied when the accent type, phonological combination, and number of mora of bunsetsu satisfy the following conditions.

Condition (other than type 1) & (VV, VN, NV at the beginning of word) &
(4 or more mora) That is, the accent type is other than type 1, and the beginning of the word, that is, the first mora and the second mora are V-V, V-N,
The point pitch pattern is operated according to the following rules when three conditions are satisfied such that the number of moras is four or more and the number of moras is four or more.

Operation Rule a. The vowel center point pitch of the first mora is the accent kernel pitch of the accent phrase. The accent phrase is a minimum unit represented by one point pitch pattern. b. However, in the case of type 0, it matches the final mora.

C. The center of gravity pitch of the second mora is
Move to the boundary point pitch of the second mora. d. However, in the case of type 2, the average value with the original first mora pitch is the boundary point pitch between the first and second mora. e. Discarding the barycentric point pitch of the second mora As a result of the above operation, the point pitch pattern of FIG. 9A is transformed into the extended point pitch pattern of FIG. 9B. Note that C in the figure represents a consonant.

2. An example of the extension method in the accent kernel will be described with reference to FIG. The rule of this example is applied when the accent type and phonological combination of bunsetsu satisfy the following conditions.

Condition (other than type 1) & (accent kernel and subsequent mora are VV, V
N, NV) That is, when the accent type is other than type 1, and the accent kernel and the mora immediately after that satisfy the two conditions of being connected by either VV, VN, or NV,
It is operated according to the following rules.

Operating Rule a. The pitch of the center of gravity of the accent kernel is used as it is as the boundary point pitch between the accent kernel mora and the subsequent mora. b. The center of gravity pitch of the following mora is discarded with the following exceptions. b1. The next mora of the following mora is devoiced.

B2. When the accent type is (number of mora in accent phrase-1). As a result of the above operation, the point pitch pattern shown in FIG. 10A is transformed into the extended point pitch pattern shown in FIG. An example of the extension method in the 3.1 type accent phrase will be described with reference to FIG. The rule of this example is applied when the accent type and phonological combination of bunsetsu satisfy the following conditions.

Condition (Type 1) & (First and second mora are VV, V
That is, when the accent type is type 1, when the two conditions that the first mora and the second mora are connected by VV, VN, NV are satisfied, the operation is performed according to the following rule. To be done.

Operating Rule a. The pitch of the center of gravity of the first mora is moved to the boundary pitch of the first and second mora. b. Discard the second mora centroid pitch. c. However, in the case of the two-mora phrase, the center-of-gravity point pitch of the second mora is reserved. As a result of the above operation, the point pitch pattern of FIG. 11A is transformed into the extended point pitch pattern of FIG. 11B.

These rules are obtained empirically from the result of observation of the analysis value of natural speech, but the generated extension point pitch pattern has a pattern comparable to that of natural sound. However, the synthesized speech has improved naturalness as compared with the conventional point-pitch model. It should be noted that the above rules may be rules corresponding only to the combination of accent type and phoneme.

As described above, the extended point pitch model uses the same pitch pattern database as the conventional point pitch model, but the degree of freedom of control is increased so as to match the phenomenon of natural speech, and it is realized by the rule. It is a thing. The above-mentioned transformation rules are VV coupling at the beginning of a word or accent nucleus, which greatly affects the quality, but in addition to this, various prominence control, voice quality control, consonantization, devoicing, and various changes in the speech speed speed are also possible. This model can express the pitch pattern.

[0037]

As described above, according to the present invention, it is possible to obtain a pitch pattern closer to natural speech by using the point pitch pattern of the conventional method and setting the boundary pitch according to the rule. become. As a result, the naturalness of the synthetic sound is improved.

[Brief description of drawings]

FIG. 1 is a block diagram showing details of a pitch pattern generation unit and a prosody control rule according to an embodiment of the present invention.

FIG. 2 is a diagram showing the overall configuration of a Japanese speech synthesis system according to an embodiment of the present invention.

FIG. 3 is a diagram (1) showing a normalized point pitch pattern.

FIG. 4 is a diagram showing a normalized point pitch pattern (No. 2).

FIG. 5 is a diagram showing a normalized point pitch pattern (No. 3).

FIG. 6 is a diagram showing a normalized point pitch pattern (No. 4).

FIG. 7 is a diagram illustrating a method of adding a descending component in an accent phrase.

FIG. 8 is a diagram illustrating a method of adding a descending component between accent phrases.

9 is a diagram for explaining an expansion method at the beginning of a word, which is executed by the pitch pattern generation unit in FIG.

10A and 10B are views for explaining an extension method in an accent kernel executed by the pitch pattern generation unit in FIG.

FIG. 11 is a diagram for explaining an expansion method in a type 1 accent phrase executed by the pitch pattern generation unit in FIG. 1;

[Explanation of symbols]

 8 ... Pitch pattern generation unit 9 ... Prosody control rule 21 ... First database (normalized point pitch pattern) 22 ... Accent phrase descending component rule 23 ... Accent phrase descending component rule 24 ... Second database (transformation rule)

Claims (3)

[Claims] 1. A speech synthesizer, comprising: a first database section storing a point pitch pattern according to an accent type; a second database section storing a rule corresponding to a combination of an accent type and a phoneme; and an object of speech synthesis. And a pitch pattern generation unit that generates a new point pitch pattern based on the rule of the second database unit from the obtained point pitch pattern based on the obtained point pitch pattern. A speech synthesizer characterized by: 2. The rules stored in the second database unit correspond to combinations of accent types, phoneme combinations, and mora numbers.
The described speech synthesizer. 3. The new point pitch pattern is generated by applying the rule after adding a descending component to the point pitch pattern stored in the first database section. Speech synthesizer.