JPH07181995A - Device and method for voice synthesis - Google Patents

Abstract

PURPOSE:To generate a synthesized voice signal having a more human voice feeling and close to a natural voice. CONSTITUTION:A voice element piece data storage part 14B storing environment-unrelated voice element piece data generated by analyzing a voice signal generated by clear sound-by-sound vocalization without phoneme environment generation and a voice element piece data storage part 14A storing voice element piece data with extraction environment generated by analyzing a voice signal generated by vocalization with phoneme environment are prepared. Selection information storage means 16 and 18 storing selection information on the voice element piece data by the kinds of voice units are prepared. The selection information storage mean are fererred to in voice units of a phoneme sequence generated by converting inputted character information to select the voice element piece data with extraction environment which is high in similarity in phoneme environment to the voice unit of the phoneme sequence and provides an excellent connection with last voice element piece data when the data are found or the environment-unrelated voice element piece data when not (13, 17, and 19).

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 and a voice synthesizing method for converting input character string information into voice and outputting the voice.

[0002]

2. Description of the Related Art A voice synthesizer which receives character information (for example, text data) and converts it into voice and outputs the voice has no limitation on the output vocabulary. As a synthetic technology, application in various fields of application can be expected. For example, output text data created by a word processor, etc.,
It can be used, for example, when outputting text data in the target language obtained by the translation process in pronunciation.

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

In FIG. 2, the text analysis unit 101 uses the pronunciation dictionary 102 to make use of the character information input unit 100.
A phonological / prosodic symbol string is generated from a kanji-kana mixed sentence input by the user. Here, the phoneme / prosodic symbol string is a string (intermediate language) in which the reading, accent, intonation, etc. of the input sentence are described as a character string. The reading and accent of each word are registered in the pronunciation dictionary 102, and the text analysis unit 101 refers to the dictionary 102 to generate a phoneme / prosodic symbol string.

In the synthesis parameter generation unit 103, the phoneme /
Based on the prosodic symbol string, parameters for voice synthesis (referred to as synthesis parameters) such as a voice unit (sound type), phoneme duration (sound length), and fundamental frequency (voice pitch) pattern are generated. Of these, the voice unit is a basic unit of voice for connecting and creating a synthetic waveform, and is generated from vocal data when a word or the like is pronounced. In the following, a combination itself of basic elements of voice such as CV (consonant-vowel) and VCV (vowel-consonant-vowel) is called a voice unit, and an element that realizes a waveform of the voice unit is a voice unit. Call. One voice unit corresponds to, for example, a set of a plurality of voice units. The speech unit data is stored in the speech unit data storage unit 104 such as a ROM, and the synthesis parameter generation unit 103 recognizes the speech unit from the phoneme / prosodic symbol string and extracts the corresponding speech unit data. .

The voice synthesizing section 105 produces a synthetic waveform based on the synthesis parameters produced by the synthesis parameter producing section 103. Such a synthesized voice signal is output as a sound through a speaker or is transmitted to another device through a line.

By the way, since humans utter various phonemes, the shape of the vocal tract is adjusted according to the phoneme. However, in the case of general speech that is continuously uttered like conversational speech, the shape of the vocal tract is Since it cannot change suddenly, it has the property that it is affected by the preceding and following phonemes and deviates from its original frequency in the middle part between the phonemes. It is called articulatory coupling that the acoustic characteristics change continuously in the middle part of the phoneme and the phoneme. In recent years, in order to improve the quality of synthesized speech, a speech synthesizer also has a synthesis parameter considering this articulatory coupling. Is being considered.

A conceivable first synthesis parameter generation method is as a voice unit composed of the same CV (consonant-vowel), VCV (vowel-consonant-vowel), etc., of voice segment data corresponding to the voice unit. Prepare multiple phonetic units with different combinations, that is, different phoneme environments (environments where the preceding and following phonemes are different), and match the phoneme environment of the input sentence (and thus the phoneme / prosodic symbol string). The voice unit is selected and used.

Further, the second conceivable method of generating the synthesis parameter is that, when speech units (henceforth speech units) are connected to synthesize speech, distortion at the connection point between speech units cannot be avoided. , In order to reduce the number of connection points, the input phoneme sequence is divided so as to avoid connection at locations where connection distortion is large, and a voice unit of arbitrary length is selected (for example, see the following references). .

Reference “Naoto Iwahashi and Yoshinori Sakasaka,“ Subjective and Objective Evaluation of Distortion-Minimized Speech Synthesis Method ”Proceedings of Acoustical Society of Japan 2-2-15, March 1992.

[0011]

According to the above-mentioned first and second synthesis parameter generation methods in consideration of the property of articulatory combination of voices, the quality of synthesized voices is improved from the viewpoint of naturalness of voice and feeling of real voice. It can be expected to improve.

However, if the first synthesis parameter generation method is applied in consideration of the phonological environment and the second synthesis parameter generation method is applied in an attempt to reduce distortion (connection point) due to connection, 1 For each input sentence,
There has been a problem that an optimal one is obtained from various combinations of voice units, and many calculations are required.

Further, considering the increase of the calculation amount and the limitation of the memory, it is impossible to arrange all the phoneme environments for each voice unit. Therefore, if there is no voice unit with a suitable phonological environment, one with another environment will be substituted, but since the voice units that cause different articulation coupling are acoustically different, these When was connected, the connection distortion increased and the sound quality of the synthesized sound was impaired.

That is, in order to improve the sound quality, it is necessary to consider the phonological environment of each voice unit. However, when a voice unit having a phonological environment that is not stored in the storage unit is requested from the input sentence, I can't handle it properly.

The present invention has been made in consideration of the above points, and a speech synthesizer and a speech synthesizer capable of generating a synthesized speech signal having a real feeling of a natural voice with a small memory capacity and processing amount. It is intended to provide a speech synthesis method.

[0016]

In order to solve such a problem, in the first aspect of the present invention, the following means are provided in the voice synthesizing device for converting the input character information into a voice signal.

(1) An environment-irrelevant speech unit data storage unit that stores environment-unrelated speech unit data that is analyzed and generated from a speech signal in which one sound is clearly pronounced so as not to have a phonological environment , (2) a speech element data storage unit with an extraction environment, which stores speech element data with an extraction environment, which is analyzed and generated from a speech signal uttered so as to have a phonological environment,
(3) Selection information storage means that stores selection information of these speech unit data for each type of speech unit, and (4) Selection information storage for each speech unit in a phoneme string into which input character information is converted. Refer to the means, it is approximated to the phoneme environment of the voice unit in this phoneme sequence, if there is a speech unit data with an extraction environment that has a good connection with the immediately preceding speech unit data, select it, If not, a synthesis parameter generating means for selecting speech unit data irrelevant to the environment is provided.

Here, the selection information storage means includes (3-1) a voice unit dictionary in which the phoneme environment of the voice unit data with an extraction environment is stored for each voice unit, and (3-2) a voice unit dictionary with an extraction environment. And a connection part information storage part that stores connection part information of the voice unit data for each voice unit.
-1) A voice unit selection check unit that refers to the voice unit dictionary for each voice unit in the phoneme sequence in which the input character information is converted and narrows down candidates based on the degree of approximation of the voice unit in this phoneme sequence to the phoneme environment, 4-2) A similarity calculation unit that obtains the similarity between the narrowed down speech unit data with extraction environment and the immediately preceding speech unit data from the stored contents of the connection information storage unit, and (4-3) Phonological environment Based on the similarity of the connection part, the number of candidates for the speech unit data with the extraction environment is narrowed down to one, and it is determined whether to select the speech unit data with the extraction environment or the speech unit data unrelated to the environment. It is preferable that the image forming apparatus further comprises a synthesis parameter generating unit.

Further, in the second aspect of the present invention, the voice synthesizing method for converting the input character information into a voice signal is as follows.

In other words, one note 1 does not have a phonological environment.
Sounds generated from analysis of speech signals uttered clearly and generated from speech signals uttered to have an environment-independent speech unit data storage unit that stores environment-unrelated speech unit data. A speech element data storage unit with an extraction environment, which stores speech element data with an extraction environment, and a selection information storage unit that stores selection information of these speech element data for each type of speech unit. I have it. Then, the selection information storage means is referred to for each voice unit in the phoneme string in which the input character information is converted, and the phoneme environment of the phoneme unit in this phoneme string is approximated to the immediately preceding phoneme data. If there is speech element data with a good extraction environment that has a good connection, the speech element data that does not relate to the environment is selected.

Here, the selection information storage means stores a voice unit dictionary in which the phoneme environment of the voice unit data with the extraction environment is stored for each voice unit, and the connection part information of the voice unit data with the extraction environment. The connection unit information storage unit stores each voice unit, and first, the voice unit dictionary is referred to for each voice unit in the phoneme sequence in which the input character information is converted, and the voice unit dictionary in this phoneme sequence is stored. The candidates are narrowed down from the degree of approximation to the phonological environment, and the similarity between the narrowed down speech unit data with the extraction environment and the immediately preceding speech unit data is obtained from the stored contents of the connection information storage unit, and the phonological environment And whether to select the speech unit data with the extraction environment or the speech unit data with no environment, after narrowing down the candidates of the speech unit data with the extraction environment to one based on the similarity of the connection part. To decide The preferable.

[0022]

In the voice synthesizing apparatus and the voice synthesizing method according to the present invention, environment-independent voice segment data generated by analysis and generation from voice signals uttered clearly one by one without storing a phonological environment are stored. An environment-independent speech unit data storage unit, and an extraction-environment-based speech unit data storage unit that stores speech unit data with an extraction environment that is analyzed and generated from a speech signal uttered to have a phonological environment; Is prepared. Furthermore, a selection information storage means for storing the selection information of these voice segment data for each type of voice unit is prepared.

Then, the selection information storage means is referred to for each voice unit in the phoneme sequence into which the input character information is converted, and the phoneme environment of the voice unit in this phoneme sequence is approximated to the immediately preceding phoneme. If there is a voice segment data with an extraction environment that has a good connection with the voice data, it is selected, and if not, environment-independent voice segment data is selected.

As a result, since the property of articulatory coupling of voices is emphasized and synthesis is performed in consideration of the phonological environment of the input sentence (phoneme sequence) and the voice unit (henceforth voice segment data), it is possible to obtain a real voice. Inappropriate voice unit that causes distortion due to connection because standard voice units are available even when there is no voice unit in an appropriate phonological environment that can obtain synthetic voice close to natural voice. Is never used for synthesis. Further, since it is sufficient to prepare the voice segment data with the high-frequency extraction environment as the center, it is possible to reduce the memory capacity and the processing amount.

[0025]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described in detail below with reference to the drawings. In addition, this embodiment is also intended for a Japanese sentence (kanji / kana mixed sentence).

FIG. 1 is a block diagram showing the functional configuration of this embodiment. In FIG. 1, in this embodiment, a character information input unit 10, a text analysis unit 11, a pronunciation dictionary 12, a synthesis parameter generation unit 13, a speech unit data storage unit 14,
Speech synthesizer 15, speech unit dictionary 16, distance value calculator 1
7, a connection unit acoustic parameter data storage unit 18 and a voice unit selection check unit 19.

Character information input unit 10 and text analysis unit 1
1, the pronunciation dictionary 12 and the voice synthesizing unit 15 perform the same operations as the conventional corresponding configuration.

The basic function of the synthesis parameter generation unit 13 is the same as that of the conventional synthesis parameter generation unit 103. Based on the phoneme / prosodic symbol string, the voice synthesis of a voice unit, a phoneme duration, and a basic frequency pattern is performed. It is for generating a parameter for use.

In the case of this embodiment, the speech element data storage unit 14 used by the synthesis parameter generation unit 13 analyzes from the voice data naturally including the nature of voice such as articulatory coupling produced naturally. The generated voice segment data with an extraction environment 14A and environment-independent voice segment data 14B analyzed and generated from the data of voices clearly pronounced for each sound are stored.

Further, the synthesis parameter generation unit 13 of this embodiment is provided with a voice unit selection check unit 19 and a distance value calculation unit 17 in association. It can be said that the synthesis parameter generation unit 13, the voice unit selection check unit 19, and the distance value calculation unit 17 constitute a synthesis parameter generation unit.

The phonetic unit selection checking unit 19 selects a phonetic unit having a high degree of similarity in phoneme environment to the input phoneme string while searching the phonetic unit dictionary 16 by the phoneme string.

FIG. 3 shows the structure of the voice unit dictionary 16. The voice unit dictionary 16 includes, for example, a voice unit pointer table with extraction environment 16A, a voice unit storage unit with extraction environment 16B, and a voice unit pointer table 1 not related to the environment.
6C and an environment-independent voice unit storage unit 16D.

The extraction environment-added voice unit storage unit 16B stores a phoneme unit (hereinafter referred to as an extraction environment-added voice unit) together with a phonological string (extraction environment) about a uttered voice when the above-mentioned extraction-unit-added voice unit data 14A is generated. Is called so that the phonological environment of the voice unit with the extraction environment can be understood, and the voice unit in which the voice unit data 14A with the extraction environment for the voice unit with the extraction environment is stored. The address of the data storage unit 14 is also described. Information about the same voice unit (hereinafter, the same type of voice unit is referred to as the same) although the phonological environments are different is continuously stored, for example. The voice unit pointer table with extraction environment 16A is
The start address of the information group of the voice unit with the extraction environment of the same type stored in the voice unit storage unit with phoneme 16B is
It is configured so that it can be taken out based on the type (voice unit name) of the input voice unit.

In FIG. 3, "/ i / iki" is "ii".
"i" means the initial voice unit / i / when pronounced "ki", and "/ i / NdeaN" means the initial voice unit / i / when pronounced "iNdeaN" (where N is It is shown to distinguish "n" from "na line". Further, P / i / represents the head address (pointer) of the phonetic environment-attached voice unit information group for the voice unit / i / for the beginning of a word.

Environment-independent voice unit pointer table 16
The C and environment-irrelevant voice unit storage unit 16D is for the environment-unrelated voice unit data 14B and has substantially the same configuration as the extraction environment-attached voice unit pointer table 16A and the extraction environment-attached voice unit storage unit 16B. The description is omitted. Note that there is only one environment-unrelated voice unit for the same type of voice unit.

The distance value calculation unit 17 is a connection unit between a candidate voice unit with an extraction environment and a previously selected voice unit (not necessarily a voice unit with an extraction environment) at the immediately preceding position. Is calculated by reading the connection part acoustic parameter data from the connection part acoustic parameter data storage unit 18 (the distance value is used as the similarity in this embodiment). When generating the voice element data with extraction environment 14A from the uttered natural voice, the acoustic parameter data of the connection unit for each voice unit is also generated,
The connection section acoustic parameter data is stored in the connection section acoustic parameter data storage unit 18.

Based on the phoneme sequence, the synthesis parameter generation unit 13 determines the phoneme environment of each voice unit candidate narrowed down by the voice unit selection check unit 19 and the distance of the connection unit between the voice units calculated by the distance value calculation unit 17. Based on the value (similarity), if there is a speech unit with an extraction environment having an appropriate phonological environment and distance value, the speech unit data with extraction environment 14A corresponding to the speech unit with an extraction environment is stored as speech unit data. If there is no voice unit with an extracted environment having a proper phonological environment and distance value, the voice unit data 14B irrelevant to the environment is fetched from the voice unit data storage unit 14.

The speech synthesizing apparatus according to the embodiment, which is composed of the respective units functioning as described above, operates as shown in FIG. 4 throughout.

First, character information (text data) is taken in (step 201), the character information is analyzed and decomposed into phrases, and each phrase is converted into a phoneme / prosodic symbol string (step 202).

Then, along with the phoneme sequence in the phoneme / prosodic symbol sequence, a certain voice unit type is targeted, and the voice unit dictionary 16 is searched by the voice unit type to extract the voice unit with the extraction environment related to the voice unit type. Take out (Step 2)
03).

Here, when one or more speech units with an extraction environment can be extracted, the phonological environments of the input phoneme sequence and each extracted speech unit with an extraction environment are compared, and the extraction environment with the closest phonological environment is added. Leave the voice unit as a candidate (step 20)
4). At the time of this selection, if there is a current voice unit with extraction environment related to the same utterance (extraction environment) as the voice unit with extraction environment selected immediately before, this is prioritized. Also,
The degree of approximation of the phonological environment is not only the coincidence of the following phonemes,
Although it is preferable that the matching of the preceding phonemes is taken into consideration, the following phonemes may be prioritized over the preceding phonemes.

Then, the distance value between the connection portion of the immediately preceding voice unit (the voice unit with the extracted environment or the voice unit not related to the environment) which has already been selected and the connection portion of the candidate voice unit with the extracted environment is calculated, and the step is performed. If a plurality of speech units with an extraction environment remain as candidates by the processing of 204 (if there are a plurality of candidates with the same phoneme environment), one extraction environment with the smallest distance value (largest similarity) The candidates are narrowed down by the attached voice unit (step 205). The distance value calculation is not executed for the voice unit of the beginning of the word.

Then, with respect to the remaining one speech unit with extraction environment, it is judged whether or not the phonological environment and the connection distance value are appropriate (step 206). This judgment condition for the phonological environment may be appropriately set according to the number of voice units with the extraction environment stored in the voice unit dictionary 16, the target sound quality, and the like. Two
It can be mentioned that the above) matches the phoneme at the corresponding position in the input phoneme sequence. Further, the judgment of the connection part distance value is made by comparison with a predetermined threshold value.

If only one voice unit with an extraction environment that remains as a candidate is appropriate, select that voice unit with an extraction environment,
The voice segment data with extraction environment 14A corresponding to this voice unit with extraction environment is adopted (step 207).

On the other hand, if the voice unit with the extraction environment, which remains only one candidate, is inappropriate, the above step 20 is repeated.
If the extracted environment-added voice unit is not found even after searching the voice unit dictionary 16 by the process of 3, the environment-independent voice unit corresponding to the target voice unit type and formed from the voiced voice data without articulation coupling is generated. The data 14B is adopted (step 208).

After that, it is judged whether or not the voice unit data to be adopted for all the voice unit types (also for the ending voice unit types) related to the target phrase have been decided (step 209), and if not decided, the above-mentioned step. 20
Return to 3.

Then, the prosodic information of the phonological / prosodic symbol string,
Prosodic parameters (parameters that define phoneme duration, fundamental frequency pattern, power, etc.) are also set based on the determined speech unit data (step 210).

Although the series of steps 203 to 210 described above does not show a repeating loop line for each phrase, it is repeated for each phrase. The series of processes in steps 203 to 210 may be executed in parallel with the process in step 202 or the process in step 211 described later.

When the synthesis parameters (prosodic parameters, voice segment data, etc.) are determined as described above, the voice signals are synthesized and output (steps 211 and 212).
The output method may be sound output from a speaker or transmission to another device through a line.

Next, the voice synthesizing operation of the embodiment, especially the operation of determining the voice segment data to be used will be described with reference to a specific example. Here, it is assumed that the input sentence (phrase) is “cannot be” shown in FIG. 5 (1). The voice unit is VC
The description will be made on the basis of V (vowel-consonant-vowel).

This input sentence “I can't” is shown in FIG.
As shown in, the phoneme sequence is converted into a phoneme sequence of "iraranai".

First, the voice unit dictionary 16 is searched for the voice unit type [i] at the beginning of the word, and the voice unit /
The voice unit with extraction environment having i / is retrieved. That is, the voice unit pointer table unit 16A with extraction environment of the voice unit dictionary 16 is accessed by using the information of the voice unit type [i] at the beginning of the word as an address to extract the pointer value P / i /, and the pointer value P / i / The voice unit with extraction environment is retrieved by accessing the voice unit storage unit with extraction environment 16B.

FIG. 5A shows the voice unit with the extraction environment extracted at this time. The first searched "/ i / ik"
i "and the second" / i / NdeaN "are the third" /
Since i / rechigau is the phoneme unit / i / the phoneme "r" immediately after it matches the corresponding phoneme of the input phoneme sequence, the third phoneme "/ i / rechigau" is excluded from the candidates at this point. The third one is a candidate, but "/ i / rasya" retrieved after this
Since i ”has a better phoneme environment with the input phoneme sequence, this extraction environment-added voice unit is replaced with the candidate.

Such an operation is repeated, and here, "/
It is assumed that only “i / rassyai” remains as a candidate.
The distance value is not calculated because it is the beginning of a word, but the validity from the phonological environment is judged for this voice unit with extraction environment "/ i / rassyai". Although it depends on the determination condition, it is determined to be appropriate because two succeeding phonemes match. Therefore, with respect to the voice unit type [i] at the beginning of the word, the voice element data with extraction environment 14A for the voice unit / i / at the beginning of the word when the articulated combined natural sound "irassyai" is uttered is adopted. .

Next, the voice unit dictionary 16 is searched for the voice unit type [ira] in the word. in this case,
As shown in FIG. 5B, it is assumed that “/ ira / sseyai” and “s / ira / byouosi” have been taken out. Here, the former "/ ira / sssaii" remains as a candidate, but the subsequent phoneme is the input phoneme sequence "irarena".
Since it does not match i ", it is judged to be inappropriate by the validity judgment.

Therefore, the environment-independent voice unit / ira / is selected and the environment-independent voice unit data 14 corresponding thereto is selected.
It is decided to adopt B. To extract the environment-irrelevant voice unit data 14B, the environment-irrelevant voice unit pointer table unit 16C of the unit dictionary 16 is used for the voice unit type [ir.
a] is used as an address to access the pointer value and the pointer value is used to extract the environment-independent voice unit storage unit 1
6D is accessed to take out the storage address of the voice unit data 14B, and the voice unit data storage unit 14 is used at this address.
To access.

Next, the next voice unit type [are] in the word
Is the target. In this case, detailed description is avoided, but FIG.
Of the plurality of voice units with extraction environment shown in (C), only the voice unit with extraction environment "emoiw / are / nu" is narrowed down to candidates and is judged to be valid, and a natural sound "emoiwarenu" with articulation is uttered. Voice unit when
It is assumed that the voice segment data with extraction environment 14A for are / is adopted.

Next, the next voice unit type [ena] in the word
5 is obtained by searching the voice unit dictionary 16.
The four voice units with extraction environment shown in (D) are extracted.
In narrowing down candidates by matching the phoneme environment with the input phoneme sequence, "katajik / ena / i" and "a / ena" are selected.
/ I ”remains. Therefore, the magnitude of the distortion (distance value) when connected to the voice unit / are / in the voice unit“ emoiw / are / nu ”with the extraction environment immediately before selected
Is compared between both candidates.

FIG. 6 is an image-like explanatory diagram of distance value calculation. Speech unit with extraction environment "emoiw / are / n"
In the audio unit / are / in the "u", the connection section acoustic parameter data CT at the rear part is extracted,
Atajik / ena / i "or" a / ena / i "in the front of the voice unit / ena / of the connection part acoustic parameter data CH is extracted to obtain a distance value, and candidates are narrowed down to those having a small distance value. , Last voice unit / a
Since the final phoneme "e" in re / is a transition from the consonant, the beginning phoneme "e" of the voice unit / ena /
"Katajik", which is a voice unit with an extraction environment in which the front of the
"/ Ena / i" has a smaller distance value and remains as a final candidate.

Speech unit with extraction environment "katajik / e"
As for na / i ", it is assumed that there was no problem with the validity.
Therefore, the articulated and combined natural sound "katajikena"
The voice element data with extraction environment 14A for the voice unit / ena / when i "is uttered is adopted.

Next, the ending voice unit type [ai] is targeted. In this case, although detailed description is omitted, only the extraction environment-added voice unit “katajiken / ai /” is narrowed down to candidates from the plurality of extraction environment-attached voice units shown in FIG. Articulated natural sound "kat"
ajikenai "voice unit / ai /
The voice element data 14A with the extraction environment is used.

The voice unit data 14A stored in the voice unit data storage unit 14 is generated by cutting out a utterance of a word or the like, and is based on the VCV unit, but the voice unit dictionary 16 By referring to, select a speech unit that is continuously extracted and generated from a series of utterances such as words, that is, a group of phonemes that are phonologically concatenated, and save time and effort for calculating combinations. It is possible to suppress the connection distortion of.
That is, in this embodiment, since such a selection rule is provided as described above, the input phoneme sequence "iraren"
For the voice unit types [ena] and [ai] of "ai", the voice units / ena / and / ai / extracted from one word utterance "katajikenai" are applied, and the connection distortion between them can be eliminated.

Therefore, according to the above-described embodiment, since the property of articulatory coupling of voices is emphasized and synthesis is performed in consideration of the input sentence (phoneme sequence) and the phoneme environment of each voice unit, a natural voice with a feeling of real voice is obtained. It is possible to obtain a synthetic voice signal close to a voice.

Further, even when there is no voice unit in an appropriate phonological environment, since a standard voice unit is prepared, an inappropriate voice unit that causes large distortion due to connection may not be used for synthesis. Absent.

Further, since standard voice units are prepared, it is not necessary to prepare a considerably large number of voice units in consideration of the phonological environment, and the memory capacity and the processing amount can be reduced.

Furthermore, in order to obtain a synthesized voice signal having a real voice and closer to a natural voice, the search process for each voice unit is the center, and the distance value may be calculated slightly. From this point as well. The processing amount can be reduced. For example, even if the number of speech unit data with extraction environment (henceforth, the voice unit) is doubled, the processing required for searching the voice unit dictionary is doubled, and the total processing amount is never doubled. As a result, in order to improve the sound quality, it is easy to deal with the expansion of the voice unit data, and it is possible to perform voice synthesis of a quality that maximizes the ability of the device.

Further, when the voice unit is selected, the voice unit extracted and generated is preferentially selected from the same utterance as the immediately preceding voice unit, so that the connection distortion between voice units can be reduced. .

In the above embodiment, the similarity of the connection portion between voice units (thus between voice segment data) is judged by the distance value, but it may be judged by other parameters. good. Also, the distance value may be calculated at the point of confirmation when replacing the candidate based on the degree of approximation of the phoneme environment, and is not limited to the timing of the embodiment.

Further, it goes without saying that the input sentence targeted by the present invention is not limited to the Japanese sentence.

[0070]

As described above, according to the present invention, environment-independent speech segment data generated by analysis and generation from a speech signal in which one note is clearly uttered without having a phonological environment,
A speech element data with an extraction environment, which is analyzed and generated from a speech signal uttered so as to have a phonological environment, is prepared, and selection information of these speech element data is prepared for each type of speech unit, The selection information is referred to for each voice unit in the phoneme sequence into which the input character information is converted, and the phoneme environment is approximated to the voice unit in this phoneme sequence, and the connection with the immediately preceding voice unit data is extracted well. If there is speech unit data with environment, select it, and if not, select speech unit data unrelated to the environment. A voice synthesizing apparatus and a voice synthesizing method that can be obtained with a small processing amount can be realized.

[Brief description of drawings]

FIG. 1 is a functional block diagram of an embodiment.

FIG. 2 is a conventional functional block diagram.

FIG. 3 is an explanatory diagram showing a configuration of a voice unit dictionary according to the embodiment.

FIG. 4 is a flowchart showing a voice synthesizing operation of the embodiment.

FIG. 5 is an explanatory diagram of an operation for a specific input sentence according to the embodiment.

FIG. 6 is a conceptual diagram of a distance value calculation method according to an embodiment.

[Explanation of symbols]

13 ... Synthesis parameter generation unit, 14 ... Speech unit data storage unit, 14A ... Speech unit data with extraction environment, 14B ... Environment unrelated speech unit data, 15 ... Speech synthesis unit, 16 ... Speech unit dictionary, 17 ... Distance Value calculation unit, 18 ... Connection unit acoustic parameter data storage unit, 19 ... Voice unit selection check unit.

Claims (4)

[Claims] 1. A speech synthesizer for converting input character information into a speech signal, wherein an environment-independent speech segment is generated by analysis from a speech signal in which one note is clearly pronounced so as not to have a phonological environment. An environment-independent speech unit data storage unit that stores data and a speech unit with an extraction environment that stores speech unit data with an extraction environment that is analyzed and generated from a speech signal that is uttered to have a phonological environment. A piece data storage unit, a selection information storage unit that stores selection information of these speech unit data for each type of voice unit, and the above selection information for each voice unit in a phoneme string into which input character information is converted. With reference to the storage means, if there is speech segment data with an extraction environment that is close to the phoneme environment of this speech unit in the phoneme sequence and that has a good connection with the immediately preceding speech segment data, select it. Speech synthesis apparatus characterized by comprising a composite parameter generating means for selecting a boundary extraneous speech unit data. 2. The selection information storage means stores a phonetic unit dictionary in which the phoneme environment of the voice unit data with the extraction environment is stored for each voice unit, and connection part information of the voice unit data with the extraction environment. The synthesizing parameter generating means refers to the voice unit dictionary for each voice unit in the phoneme sequence into which the input character information is converted, A voice unit selection check unit that narrows down candidates based on the degree of approximation to the phoneme environment of a voice unit in a column, and a similarity between the narrowed voice unit data with an extraction environment and the immediately preceding voice unit data is the connection unit information storage unit. Based on the similarity between the phoneme environment and the connection part, and the similarity calculation unit obtained from the stored content of the phoneme data is narrowed down to one, and the voice unit data with the extraction environment is selected. Speech synthesis apparatus according to claim 1, characterized in that it consists of a synthetic parameter generation unit for determining whether to select the Luke environment extraneous speech unit data. 3. A speech synthesis method for converting input character information into a speech signal, wherein an environment-independent speech element is generated by analysis from a speech signal in which one note is clearly pronounced so as not to have a phonological environment. An environment-independent speech unit data storage unit that stores data and a speech unit with an extraction environment that stores speech unit data with an extraction environment that is analyzed and generated from a speech signal that is uttered to have a phonological environment. A piece data storage unit and selection information storage means for storing the selection information of these speech unit data for each type of voice unit are provided. For each voice unit in the phoneme sequence in which the input character information is converted, Referring to the selection information storage means, if there is a speech unit data with an extraction environment that is close to the phoneme environment of the voice unit in this phoneme sequence and has a good connection with the immediately preceding speech unit data, select it. Na Speech synthesis method and selects the environmental extraneous speech unit data if Re. 4. The selection information storage means stores a voice unit dictionary storing the phoneme environment of the voice unit data with the extraction environment for each voice unit, and connection part information of the voice unit data with the extraction environment. A phoneme unit of a phonetic unit in this phoneme string is configured by referring to the phonetic unit dictionary for each phonetic unit in the phoneme string in which the input character information is converted The candidates are narrowed down based on the degree of approximation to the environment, and the similarity between the narrowed down speech unit data with the extraction environment and the immediately preceding speech unit data is obtained from the stored contents of the connection unit information storage unit. After narrowing down the candidates of the speech unit data with the extraction environment to one on the basis of the degree of similarity, determining whether to select the speech unit data with the extraction environment or the speech unit data unrelated to the environment. Characterized by Claim 3
The speech synthesis method described in.