JPH08160990A - Speech synthesizing device - Google Patents

Abstract

PURPOSE: To provide a speech synthesizing device which can generate a more natural synthesized speech. CONSTITUTION: A text analysis part 11 generate a phoneme and rhythm symbol sequence from input character information, and a speech unit dictionary 16 contains speech unit labels and storage positions in a speech element piece data storage part 14 as to speech units stored in a speech element piece data storage part 14. In a phoneme continuance table 17, naturally spoken continuous speeches are classified having phoneme environments of >=2 front and rear sounds according to the use frequencies of phonemes and phoneme continuance in the classified phoneme environments is described by a voiceless part, a consonant part, and a vowel part. A synthesis parameter generation part 13 retrieves the phoneme continuance table with the phoneme environment of each phoneme according to the phoneme and rhythm symbol sequence to determine phoneme continuance, thereby setting natural continuance by the consonant part and vowel part.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice synthesizing apparatus for synthesizing and outputting voice based on input character string information.

[0002]

2. Description of the Related Art A voice synthesizing device for inputting character information, converting it into voice and outputting it is output in various fields as a voice synthesizing technique which replaces a recording / playback type voice synthesizing technique because there is no limitation in output vocabulary. Can be expected to be applied in. For example, text data created by a word processor etc. can be converted into voice and output, or a response message can be created and changed simply by editing the text, so it can be used for communication services such as telephones. Available.

FIG. 2 shows a configuration of a conventional speech synthesizer (Japanese text-to-speech converter) which inputs Japanese (Kanji / Kana mixed sentence) as an input. The outline of the synthesizer will be described.

In FIG. 2, a text analysis unit (101)
Uses the pronunciation dictionary (102) to generate a phonological prosodic symbol string from a kanji-kana mixed sentence input from the character information input unit (100). This phonological prosodic symbol string describes the reading, accent, intonation, etc. of the input sentence as a character string and is called an intermediate language. The reading and accent of each word are registered in the pronunciation dictionary (102), and the text analysis unit (101) uses this pronunciation dictionary (10).
A phonological prosody symbol string is generated with reference to 2).

A synthesis parameter generation unit (103) extracts a voice segment based on a phoneme prosodic symbol string and generates a voice synthesis parameter such as a phoneme duration or a fundamental frequency pattern according to a predetermined rule. Of these, the speech unit is generated by analysis from the utterance data when a word or the like is pronounced, and is a basic unit of speech for speech synthesis, and a synthetic waveform is generated by superposing these. . In the following description, CV (consonant-vowel),
A combination itself of basic elements of voice such as VCV (vowel-consonant-vowel) is called a voice unit, and an element that realizes a waveform of the voice unit is called a voice unit. Each voice unit is, for example,
It corresponds to a set consisting of a plurality of speech units. The voice unit data is stored in a voice unit data storage unit (104) such as a ROM, and the synthesis parameter generation unit (103) recognizes the voice unit from the phonological prosodic symbol string and corresponds to the voice unit data. Take out.

The voice synthesis unit (105) generates a synthetic waveform (voice signal) based on the synthesis parameter generated by the synthesis parameter generation unit (103). Such a synthetic voice signal is output as voice through the speaker (106) or is transmitted to another device through a line.

In the above prior art, the synthesis parameters such as the phoneme duration are determined based on a predetermined rule. However, in order to enhance the naturalness of the synthesized speech, the phonology of the real voice is changed for each phonological environment before and after. A method of providing the result of analysis by statistical processing is disclosed in, for example, Japanese Patent Application Laid-Open No. 03-161800.

[0008]

However, in the above-mentioned first prior art, the phoneme duration is given by the phoneme symbol string obtained by converting the input text based on a predetermined rule. It was more monotonous than the duration of natural speech, and phoneme duration was often performed only by expanding and contracting the vowel stationary part.

In the second prior art, too, since only the statistics for each phoneme before and after the phoneme of interest are taken into consideration, an appropriate duration may not be obtained. Also,
When the statistics are subdivided, it is difficult to collect statistics about all phoneme environments, and there is a problem that it is not possible to know what kind of duration table should be constructed.

Therefore, a main object of the present invention is to provide a speech synthesis / synthesis apparatus using a VCV or CVC unit, which has relatively little distortion due to connection of speech units, and has two or more mora before and after, based on analysis data of natural speech. A speech synthesizer capable of generating a synthetic speech having a more natural phoneme duration by creating a duration table efficiently subdivided according to the usage frequency and referring to this table during synthesis. To provide.

[0011]

SUMMARY OF THE INVENTION In order to solve the above problems, the present invention provides a text analysis unit for generating a phoneme prosodic symbol string from input character information, and a speech unit data storage unit for storing speech unit data. , A voice unit dictionary describing a voice unit label, a storage position in the voice unit data storage unit, and the like for a voice unit stored in the voice unit data storage unit; According to the frequency of use, the phoneme is classified into two or more phonological environments before and after, and a phoneme duration table in which the phoneme durations in the classified phonological environment are described for each silent part, consonant part, and vowel part, and a phonological prosody. A synthesis parameter generation unit that determines the phoneme duration by searching the phoneme duration table from the phoneme environment of each phoneme according to the symbol string and sets a natural duration for each consonant part and vowel part. Characterized by comprising.

[0012]

In the speech synthesizer according to the present invention, the text analysis section generates a phonological prosodic symbol string from the input character information, and the speech segment data storage section stores speech segment data which is a reference of the speech synthesis signal. Further, the voice unit dictionary describes a voice unit label, a storage position in the voice unit data storage unit, and the like for a voice unit stored in the voice unit data storage unit. Further, in the phoneme duration table, naturally uttered continuous voices are classified so as to have a phoneme environment of two or more sounds before and after according to the usage frequency of the phoneme, and the phoneme duration in the classified phoneme environment is It is described for each silent part, consonant part, and vowel part. Further, the synthesis parameter generation unit searches the phoneme duration table from the phoneme environment of each phoneme in accordance with the phoneme prosodic symbol string to determine the phoneme duration, and sets a natural duration for each consonant part and vowel part. Therefore, it becomes possible to generate synthetic speech having a natural phoneme duration.

[0013]

1 is a functional block diagram showing the configuration of a speech synthesizer according to the present invention. A character information input section 10, a text analysis section 11, a pronunciation dictionary 12, and a synthesis parameter generation section 1 are shown.
3, a voice unit data storage unit 14, a voice synthesis unit 15, a voice unit dictionary 16, a duration table 17, and a speaker 18. Of the above configuration, the character information input unit 10,
The pronunciation dictionary 12, the voice unit data storage unit 14, the voice synthesis unit 15, and the speaker 18 perform the same operation as the corresponding constituent elements of the conventional voice synthesis apparatus in FIG.

Synthesis parameter generator 1 in this embodiment
The duration table 17 used by No. 3 stores the duration that is analyzed and generated from the voice data that is naturally uttered based on the duration model.

In the voice unit dictionary 16, for voice units stored in the voice unit data storage unit 14, for example, if the voice unit is a VCV unit, voice unit labels such as / aki / and / eki / (voice units Name) and the storage location in the voice unit data storage unit 14 are described.

The synthesis parameter generation unit 13 refers to the voice unit dictionary 16 based on the phoneme symbol string, retrieves the corresponding voice unit data from the voice unit data storage unit 14 in accordance with the selected voice unit, and outputs the text data. The duration is determined by referring to the duration table 17 from the phonological environment and accent information, and parameters for speech synthesis such as power and fundamental frequency pattern are generated.

The speech synthesizing apparatus of this embodiment, which is composed of the respective units functioning as described above, operates as follows as a whole. This operation procedure will be described with reference to FIG. First, character information (text data such as kanji and kana mixed sentences)
Is input (step S201), the character information is analyzed and converted into a phonological prosodic symbol string for each phrase (step S202).

Next, along the phonological prosodic symbol string, the phonetic word dictionary 16 is sequentially created according to the type of the phonetic unit at the beginning of the phrase.
To retrieve the voice segment data. (Step S2
03).

After that, the duration of each phoneme is determined for each phrase by referring to the duration table 17 based on the phoneme prosodic symbol string (step S20).
4), prosody parameters (parameters that define phoneme duration, fundamental frequency pattern, power, etc.) are set (step S205).

When the synthesis parameter composed of the prosody parameter and the voice segment data is determined as described above, the voice signal is synthesized (step S206) and output (step S207). The output method may be output from the speaker 18 or may be transmission to another device via a line.

Next, a method of creating the duration table 17 will be described in detail. First, using natural speech labeled for each part of the silent part, consonant part, and vowel part for each phoneme,
The duration is calculated for each phoneme environment before and after. FIG.
This is a part of the classification of phonological environments for "ka". A label table in which labels are spread in a tree shape is constructed from a phoneme environment close to "ka", but the same applies to other phonemes. First, the consonant part in the [K] group is the top of the classification tree. The next classification is the phoneme of "ka, ki, ku, ke, ko". For the next classification, an example in which the subsequent phonological environment is “sa” is shown. First, the group of [S] and then “sa, shi, su, se,
That's the group. The next is the immediately preceding environment, but here is the average of the vowel group, and the next is for each vowel "A, I, U, E, O", and so on. Labels are written in a tree shape. When there is no phoneme that follows the beginning or end of a word, classification is performed in one direction from the beginning of the word to the succeeding environment and from the ending of the word to the immediately preceding environment, and a label is created. This will be the label for the table.

The voice data is divided into silent parts, consonant parts, and vowel parts according to the label, and the duration is calculated and described in the table. Although the amount of data decreases as going down the tree, the data is repeatedly classified until the variation in the data is not statistically absorbed, for example, the data becomes about 10. In this case, the data runs out in the middle of the tree, but at that time, the terminal symbol is described in the label at the time when the data runs out (* symbol is used in FIG. 4). The table created in this way can describe the phoneme duration more finely for phonemes with a large amount of data, that is, for phonemes that are frequently used.

A specific example of the process (step S204 in FIG. 3) for determining the phoneme duration using this duration table 17 is shown below. Here, the input sentence will be described as "this is a speech synthesizer". Further, although the present invention does not make the synthesis unit a problem, the VCV unit will be used for explanation here.

The above-mentioned input sentence is analyzed by the text analysis unit 11 as P1 corewa, P2 onsaegoseo'-chidespo. The text analysis unit 11 converts the input sentence into a phonological prosodic symbol string with reference to the pronunciation dictionary 12 as described above. At this time, a phrase symbol (P0, P1, P2) or the like is inserted at the beginning or in the sentence or at the end of the sentence, if necessary. These phrase symbols indicate rising and falling edges of a phrase at the beginning of a sentence, the inside of a sentence, and the end of a sentence.

First, for the first phrase "corewa", the synthesis parameter generator 13 refers to the voice unit dictionary 16 based on the phoneme symbol string, and outputs the corresponding voice segment data according to the selected voice unit. It is taken out from the segment data storage unit 14. In VCV voice unit, / ko /, /
Speech units corresponding to three speech units of ore / and / ewa / are taken out.

Next, referring to the duration table 17, the duration of each phoneme is determined. Table 1 shows each phoneme environment.

[0027]

[Table 1] Here, the duration table 17 is referred from the phoneme symbol string, and the frame length and the number of frames are set so that the duration on the table is realized from the configuration (consonant length, vowel length) of the speech unit in the voice unit. By determining, the duration is comprehensively determined. By using such a duration table, it is possible to obtain the duration for each of the consonant part and the vowel part, and to determine the duration to be close to the balance between the consonant length and the vowel length of natural speech.

For example, for the phoneme / ko /, one frame has a standard speed of 8 milliseconds and the length required in the table is as follows: Table: {ko} = {k} + {o} = 6 frames + 9
It is a frame, and the actual length of the speech segment is as follows: Element configuration: {ko} = {k} + {o} = 5 frames + 1
When the frame length is 1 frame, the frame length of the consonant part of the speech unit is 9.6.
If it is set to milliseconds, the duration of the same consonant as in the table can be realized, and for the vowel portion, by adjusting the number of frames of the vowel stationary portion as in the conventional case, a more natural duration can be set.

Depending on the phoneme environment of each phoneme, how far the phoneme is classified on the duration table is different, but the data at the final end where the phoneme is classified is used.
As a result, for phonemes that are frequently used, it is possible to obtain more detailed and appropriate phoneme duration,
The sound becomes more natural. In addition, it is possible to automatically set the unvoiced sound by having the environment with two or more sounds before and after and the silence, consonant, and vowel length separately, and the pause length between phrases and the length of silent intervals are all statistical. It can be decided from the data. The use of the data at the end of the tree is, for example, "to reveal. (Akasu
To refer to the duration of "ka" in the phoneme sequence "nodeatta)" from FIG. 4, check the label of the subsequent [su], then the immediately preceding [a], and then the second [] of the subsequent environment. n
Looking at the label of [o], since the terminal symbol (*) is added, the average data of the group of group [n] before [no] is used.

[0030]

As described above in detail, according to the present invention, a text analysis unit for generating a phoneme prosodic symbol string from input character information, and a voice unit data storage unit for storing voice unit data. A voice unit dictionary describing a voice unit label, a storage position in the voice unit data storage unit, and the like for a voice unit stored in the voice unit data storage unit,
Naturally uttered continuous speech is classified so as to have a phonological environment of two or more sounds before and after according to the frequency of use of the phonology, and the phoneme duration in the classified phonological environment is divided into silent parts, consonant parts, and vowel parts. According to the described phoneme duration table and the phoneme prosodic symbol string, the phoneme duration table is searched from the phoneme environment of each phoneme to determine the phoneme duration, and a natural duration is set for each consonant part and vowel part. Since it is configured to include the synthesis parameter generation unit, it has a very effective effect on the rise and fall of phrases and phonemes with accents, and the effect of increasing the real voice and giving a more natural impression.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of an embodiment of a speech synthesizer of the present invention.

FIG. 2 is a block diagram showing a configuration of a conventional speech synthesizer.

FIG. 3 is a flowchart showing a voice synthesizing operation procedure of the voice synthesizing device of the embodiment.

FIG. 4 is a diagram showing a configuration example of a duration table.

[Explanation of symbols]

 10 character information input unit 11 text analysis unit 12 pronunciation dictionary 13 synthesis parameter generation unit 14 speech unit data storage unit 15 speech synthesis unit 16 speech unit dictionary 17 duration table 18 speaker

Claims (1)

[Claims] 1. A text analysis unit for generating a phonological prosodic symbol string from input character information, a voice unit data storage unit for storing voice unit data, and a voice unit stored in the voice unit data storage unit. For each piece, a voice unit dictionary describing the voice unit label and the storage position in the voice unit data storage unit, and a continuous voicing spontaneously have a phonological environment of two or more tones before and after, depending on the phoneme usage frequency. A phoneme duration table in which the phoneme duration in the classified phoneme environment is described for each of the silent part, the consonant part, and the vowel part, and the phoneme continuation from the phoneme environment of each phoneme according to the phoneme prosodic symbol string. A speech synthesizing apparatus comprising: a synthesis parameter generating unit that searches a time table to determine a phoneme duration and sets a natural duration for each consonant part and vowel part.