JP3235747B2 - Voice synthesis device and voice synthesis method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

[Table of Contents] The present invention will be described in the following order. BACKGROUND OF THE INVENTION Problems to be Solved by the Invention Means for Solving the Problems (FIG. 1) Function (FIG. 1) Embodiment (FIGS. 1 to 4) Effects of the Invention

[0002]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech synthesizing apparatus and a speech synthesizing method, and more particularly to a speech synthesizing apparatus and a speech synthesizing method suitable for generating a synthesized speech in accordance with a regular speech synthesizing method.

[0003]

2. Description of the Related Art Conventionally, a speech synthesizer using a regular speech synthesis system analyzes a sequence of input characters and synthesizes parameters in accordance with a predetermined rule, thereby synthesizing any words. Have been made to gain. That is, the speech synthesizer based on the rule speech synthesis method analyzes an input character sequence, detects an accent for each phrase in accordance with a predetermined rule, and inverts the entire character sequence from the sequence of the phrases. And pitch parameters expressing the pose and the like.

Further, the speech synthesizer similarly divides each phrase into speech units such as CV units in accordance with a predetermined rule, and then generates a synthesis parameter representing the spectrum. As a result, a synthesized sound is produced based on the pitch parameter and the synthesis parameter.

In order to synthesize higher quality speech,
For a voiced part having periodicity as a voice unit, the real voice is analyzed and processed, and the voice waveform data corresponding to one period is held, and for a voiceless part having no periodicity, the real voice is held as voice waveform data as it is. At the time of synthesis, these voice waveform data are superimposed on the waveform based on the pitch parameter to generate a synthesized sound.

[0006]

By the way, in the conventional waveform superimposition technique, the voice is synthesized by adjusting the number of frames by repeating or thinning out the voice waveform data in the voice unit based on the pitch parameter. ing. Each speech unit used here is influenced by the phonemic environment before and after it in the actual speech from which it is extracted, and the effect appears in the synthesized speech.

That is, in a certain voice unit, the phoneme environment at the time of synthesis may differ from the phoneme environment in the extracted real voice. As a result, an unnatural speech waveform is generated in each speech unit connection part of the synthesized speech as compared with the real speech, and abnormal hearing occurs due to discontinuity in the frequency domain, and the quality of the synthesized speech is reduced. Had a problem when it was apt to deteriorate.

The present invention has been made in view of the above points, and proposes a voice synthesizing apparatus and a voice synthesizing method capable of producing a synthesized sound with less deterioration in quality as compared with actual human voice and without a sense of incongruity. What you want to do.

[0009]

According to the present invention, in order to solve the above-mentioned problems, for each voice unit, a voiced portion having a periodicity within the voice unit is obtained by an actual voice analysis process. Speech waveform data corresponding to each one pitch cycle is stored in memory as speech units for the required number of frames, and at the same time, the front end frame and the rear end frame of the speech unit are envelopes obtained by analysis processing together with the speech waveform data. Information and fine structure information are stored together in a memory, and at the same time, for a voiceless part having no periodicity in a voice unit, a voice unit storage unit 2 which stores the real voice in the memory as voice waveform data as it is, an input phonological symbol and a prosody A speech synthesis rule unit 4 for generating a pitch pattern in accordance with a predetermined phoneme rule and a prosody rule based on symbols;
In the interpolation frame at the time of synthesis of the voiced part, the envelope information of the rear end frame of the preceding speech unit and the front end frame of the rear speech unit are interpolated, and a time waveform is obtained along with the fine structure information of the preceding speech unit. Is provided as the speech waveform data of the interpolation frame, and a speech synthesis unit 5 for generating a synthesized sound based on the speech unit and the pitch pattern.

Further, in the present invention, a sentence analyzing unit 3 for analyzing a sequence of inputted characters to detect a word, a boundary of a phrase and a basic accent, and for each voice unit, a period within the voice unit. Voiced part
The audio waveform data corresponding to each one-pitch cycle obtained by the analysis processing of the actual audio is stored in the memory as the audio unit for the required number of frames, and at the same time, the audio waveform data is stored in the front end frame and the rear end frame of the audio unit. Together with the envelope information and the fine structure information obtained by the analysis process, and simultaneously stores in a memory an unvoiced portion having no periodicity within the voice unit, and stores the real voice as it is as voice waveform data in the memory. And sentence analysis unit 3
According to a predetermined phoneme rule and a prosody rule based on the analysis result, a speech synthesis rule unit 4 for generating a pitch pattern, and an interpolation frame at the time of synthesis of a voiced portion, a rear end frame and a rear speech unit of a preceding speech unit A time waveform is obtained by interpolating the envelope information of the front end frame and the fine structure information of the preceding voice unit, and the time waveform is used as the voice waveform data of the interpolation frame to generate a synthesized sound based on the voice unit and the pitch pattern. The voice synthesizing unit 5 is provided.

According to the present invention, for each voice unit, voice waveform data corresponding to each one-pitch period obtained by analysis processing of a real voice is obtained for a voiced portion having periodicity in the voice unit. Are stored in the memory for the required number of frames as a voice unit, and at the same time, the envelope information and the fine structure information obtained by the analysis processing together with the voice waveform data are stored in the memory in the front end frame and the rear end frame of the voice unit. Stored in memory, and at the same time, for unvoiced portions with no periodicity in the voice unit, stored the real voice in the memory as voice waveform data as it is,
In accordance with predetermined phonological rules and prosody rules based on the input phonological symbols and prosodic symbols, a pitch pattern is generated, and in the interpolation frame at the time of synthesis of the voiced portion, the rear end frame of the preceding voice unit and the front end of the rear voice unit. A time waveform is obtained by interpolating the envelope information of the frame and the fine structure information of the preceding voice unit, and using the time waveform as the voice waveform data of the interpolated frame, a synthesized sound is generated based on the voice unit and the pitch pattern. I made it.

Further, in the present invention, a sequence of inputted characters is analyzed to detect a word, a boundary of a phrase and a basic accent, and for each voice unit, a voiced portion having a periodicity within the voice unit. For each of the above, the required number of frames of audio waveform data corresponding to each one-pitch cycle obtained by the analysis processing of the actual audio are stored in the memory as the audio unit, and at the same time, the audio waveform data is stored in the front end frame and rear end frame of the audio unit Envelope information and fine structure information obtained by the analysis process together with the data are stored together in a memory, and at the same time, for unvoiced parts without periodicity in the voice unit,
The actual voice is stored as it is in the memory as voice waveform data,
A pitch pattern is generated in accordance with a predetermined phonological rule and a prosodic rule based on the analysis result of the character sequence, and in the interpolation frame at the time of synthesis of the voiced portion, the pitch of the rear end frame of the preceding voice unit and the front end frame of the rear voice unit are calculated. A time waveform is obtained by interpolating the envelope information together with the fine structure information of the preceding voice unit, and using the time waveform as the voice waveform data of the interpolation frame based on the voice unit and the pitch pattern to generate a synthesized sound. did.

In the present invention, a voice based on Japanese is used as the voice.

[0014]

The voice waveform data corresponding to each one-pitch period obtained by the analysis processing of the real voice is used for the voiced portion having periodicity, and the real voice is used for the voiceless portion having no periodicity as it is. In the audio unit stored in the memory for the required number of frames as data, the front end frame and the rear end frame of the audio unit together with the audio waveform data,
By the analysis process, the envelope information of the audio waveform data and the fine structure information are provided together. Further, a pitch pattern of the synthesized voice is generated according to a predetermined prosody rule, and voice waveform data necessary for the synthesized voice is read from a memory according to the phoneme rule, and a synthesized voice is generated based on the voice waveform data and the pitch pattern. By doing so, it is possible to utter a synthesized sound with less deterioration in quality as compared to actual human voice and without a sense of incongruity.

[0015]

BRIEF DESCRIPTION OF THE DRAWINGS FIG.

In FIG. 1, reference numeral 1 denotes a schematic configuration of a speech synthesis apparatus including an arithmetic processing unit as a whole, which is divided into a speech unit storage section 2, a sentence analysis section 3, a speech synthesis rule section 4, and a speech synthesis section 5. Is done.

The sentence analysis unit 3 analyzes a text input (consisting of a sentence or the like represented by a series of characters) input from a predetermined input device with reference to a predetermined dictionary and converts it into a kana character string. Later, it is broken down into words and phrases. In other words, since words are not separated in Japanese as in English, there are two types of words such as “US / industry / industry” and “US / industry / industry”, for example. Can be sectioned.

For this reason, the sentence analyzing unit 3 is configured to decompose a text input into words and phrases by using the continuous relation of words and the statistical properties of words while referring to the dictionary. It is designed to detect clause boundaries. Further, the sentence analysis unit 3 detects a basic accent for each word, and outputs it to the speech synthesis rule unit 4.

The speech synthesizing rule unit 4 follows a predetermined phonetic rule set based on Japanese features and outputs the sentence analyzing unit 3.
And a text input. That is, natural Japanese speech can be divided into about 100 utterance units when distinguished based on linguistic characteristics. For example, if the word "Sakura" is divided into utterance units, "sa" + "ak" + "ku" + "ur" +
It can be divided into five CV / VC units of “ra”.

Further, in Japanese, when words are consecutive, the initial syllable of the word after the consecutive words becomes muddy (that is, it becomes turbid), and the syllabic sounds other than the initial of the word become nasal, so that the word alone is There is a feature that the utterance changes with the case. Accordingly, the speech synthesis rule section 4 sets a phoneme rule according to these Japanese features, and converts a text input into a phoneme symbol string (that is, the above-mentioned “sa” + “ak” + "K
u "+" ur "+" ra "). Further, the speech synthesis rule unit 4 loads data of each speech unit from the speech unit storage unit 2 based on the phoneme symbol string.

Here, the speech synthesizer 1 produces a synthesized sound using a waveform editing technique, and data loaded from the speech unit storage unit 2 is synthesized in units of CV / VC. This is waveform data used when generating a sound. The voice unit data used for this waveform synthesis is configured as follows. In other words, the voiced part of the voice unit data is composed of voice waveform data corresponding to one pitch in the voiced part of the real voice in the required number of frames in the memory. The waveform of the unvoiced portion is cut out and stored in the memory as it is.

As shown in FIG. 2, the front end frame and the rear end frame of the voiced portion of the voice unit data are obtained together with the voice waveform data (FIG. 2A) by analysis processing such as cepstrum analysis. The envelope information (FIG. 2B) and the fine structure information (FIG. 2C) of the audio waveform data are simultaneously stored in the memory. Therefore, the voice unit data is CV /
In the case of the VC unit, when the consonant part C of one voice unit CV is an unvoiced consonant, a cutout waveform of an unvoiced part;
In a plurality of frames each composed of a one-pitch voice waveform, the trailing end frame also includes envelope information and fine structure information of the corresponding voice waveform.

As a result, one voice unit data is formed, and when the consonant part C of one voice unit CV is a voiced consonant, in a plurality of frames consisting of one pitch voice waveform, the leading frame and the trailing frame are included. The envelope information and the fine structure information of the corresponding audio waveforms are included, thereby forming one audio unit data.

The speech synthesis rule unit 4 synthesizes the speech unit data loaded from the speech unit storage unit 2 in an order according to the text input (hereinafter, this data is referred to as synthesized waveform data), and thus, in a state without inflection. Thus, a synthesized speech waveform obtained by reading a text input can be obtained. Further, in the connection of voice units in a voiced part in the synthesized waveform data, the following processing is performed.

In a certain phoneme chain C'VC "to be synthesized by the voice synthesis unit 5, the waveform data groups in the data C'V and VC" loaded from the voice unit storage unit 2 are arranged in order, and At the connection portion within the same phoneme V, speech waveform data obtained by the following interpolation processing is used. This is, as shown in FIG.
Interpolation processing such as linear interpolation is performed using the envelope information in the end frame of the end speech frame and the envelope information in the front frame of the rear speech unit VC ″, and the envelope information obtained by this interpolation processing and the preceding speech unit C ′ The sum of the microstructure information in the terminal frame of V is used as the frequency information of the interpolation frame at the connection portion.

This frequency information is converted from the frequency domain to the time domain, and is used as speech waveform data of the interpolated frame at the connection section, and is used for composite waveform data. The connection of the speech units by the interpolation process is also performed within the same phoneme C of the speech unit V'C and the speech unit CV "in another phoneme chain V'CV" in which C is a voiced consonant.

Further, the speech synthesis rule section 4 divides the text input into an appropriate length based on a predetermined prosody rule and detects a break (ie, a pause). Thus,
For example, as shown in FIG. 4 (A), when the sentence “I received a beautiful flower from Mr. Yamada” is input as a text input, the text input is “pretty”, as shown in FIG. 4 (B). "Hanawo", "From Yamada",
After being decomposed into "I got it", a pose is detected between "Hanao" and "Yamada-san".

Further, the speech synthesis rule section 4 detects the accent of each phrase based on the prosodic rules and the basic accent of each word. In other words, the accent of a Japanese phrase alone can be intuitively expressed in two levels, high and low, in units of kana characters (hereinafter referred to as mora). At this time, the accent position of the phrase can be distinguished according to the content of the phrase and the like.

For example, edges, chopsticks, and bridges are two-mora words, each of which is classified into a type 0 with no accent, a type 1 with an accent position in the first mora, and a type 2 with an accent position in the second mora. can do. Thus, in the case of the present embodiment, the speech synthesis rule unit 4 sequentially converts each phrase of the text input into type 1 type, type 2 and 0, as shown in FIG.
The type and the type are classified, and the accent and the pause are detected in units of a phrase.

Further, the speech synthesis rule section 4 generates a basic pitch pattern representing the intonation of the entire text input based on the detection result of the accent and the pose. That is, while the accent of a phrase in Japanese can be intuitively expressed at two levels, the actual intonation is characterized by a gradual decrease from the position of the accent (FIG. 4D). Further, in Japanese, when a phrase is continuously formed into one sentence, there is a characteristic that the intonation gradually decreases toward the pause following the pause (FIG. 4E).

Accordingly, the speech synthesis rule unit 4 generates a parameter representing the inflection of the entire text input for each mora based on such Japanese features, and then the inflection is smooth as in the case where a human utters. Is set by interpolation between moras. Thus, the speech synthesis rule section 4 synthesizes the parameters of each mora and the interpolated parameters in the order corresponding to the text input (hereinafter referred to as pitch pattern), and thereby, as shown in FIG. A pitch pattern (FIG. 4 (F)) indicating the intonation of the read voice is obtained.

The voice synthesizer 5 performs a waveform synthesis process based on the synthesized waveform data and the pitch pattern to generate a synthesized sound. This waveform synthesizing process performs the following. That is, in the voiced portion of the synthesized voice, the waveform data corresponding to one pitch in the synthesized waveform data is arranged and superimposed on the basis of the pitch pattern. In the unvoiced portion of the synthesized voice, the cut-out waveform in the synthesized waveform data is used as a desired synthesized voice waveform.

As a result, it is possible to obtain a synthesized sound whose inflection changes according to the change of the pitch pattern. Therefore, in a speech synthesis system of a waveform superposition method capable of obtaining a high-quality synthesized speech, connection distortion at a voice unit connection part in a voiced portion can be reduced, and a synthesis method without interpolation or a time axis The connection unit can be connected more smoothly than in the synthesis method based on simple waveform interpolation, and a high-quality arbitrary synthesized sound close to human voice can be obtained.

In the above configuration, a text input input from a predetermined input device is analyzed by the text analysis unit 2 with reference to a predetermined dictionary, and words, boundaries between phrases and basic accents are detected. The detection results of the words, the boundaries of phrases and the basic accents are processed by the speech synthesis rule unit 4 in accordance with a predetermined phonological rule, and synthetic waveform data representing speech read out of the text input without inflection is generated. .

Further, the detection results of the words, the boundaries of phrases and the basic accents are processed by the speech synthesis rule unit 4 in accordance with a predetermined prosody rule, and a pitch pattern representing the inflection of the entire text input is generated. The pitch pattern is output to the speech synthesizer 5 together with the synthesized waveform data, where a synthesized sound is generated based on the pitch pattern and the synthesized waveform data.

According to the above configuration, in a speech synthesis system of a waveform superposition method capable of obtaining a high-quality synthesized speech, connection distortion at a speech unit connection part in a voiced portion can be reduced, and a speech unit can be more smoothly processed. By connecting, the speech synthesizer 1 that can arbitrarily generate high-quality synthesized speech close to human speech can be realized.

In the above-described embodiment, in a certain phoneme chain C'VC "to be synthesized by the voice synthesizer 5, the envelope information in the end frame of the preceding voice unit C'V and the rear voice unit VC" For the envelope information obtained by interpolating the envelope information in the leading frame, the fine structure in the leading frame of the rear voice unit VC ″ is replaced with the fine structure information in the trailing frame of the preceding voice unit C′V. The information to which the information is added may be used as the frequency information of the interpolation frame in the connection unit.

[0038]

As described above, according to the present invention, in a voice synthesis system of a waveform superposition method capable of obtaining a high-quality synthesized voice, connection distortion at a voice unit connection part in a voiced portion can be reduced. A speech synthesizer capable of arbitrarily synthesizing a high-quality synthesized sound similar to human speech can be obtained.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a speech synthesizer according to one embodiment of the present invention.

FIG. 2 is a signal waveform diagram for describing data of a leading frame and a trailing frame in audio unit data.

FIG. 3 is a schematic diagram for explaining an interpolation process;

FIG. 4 is a schematic diagram used to explain the operation of the speech synthesizer.

[Explanation of symbols]

1 ... Speech synthesis device, 2 ... Speech unit storage unit, 3 ... Sentence analysis unit, 4 ... Speech synthesis rule unit, 5 ... Speech synthesis unit.

Continuation of front page (56) References JP-A-3-48300 (JP, A) JP-A-4-98298 (JP, A) JP-A-63-208099 (JP, A) JP-A-63-136099 (JP, A) JP-A-63-118800 (JP, A) JP-A-62-133490 (JP, A) JP-A-4-199196 (JP, A) JP-A-4-281499 (JP, A) 5-94196 (JP, A) JP-A-5-108085 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G10L 13/06

Claims (5)

(57) [Claims] For each voice unit, for a voiced part having periodicity within the voice unit, voice waveform data corresponding to each one-pitch cycle obtained by analysis processing of a real voice is stored in the voice unit. At the same time, the front end frame and the rear end frame of the audio unit are stored in the memory together with the audio waveform data together with the envelope information and the fine structure information obtained by the analysis process. At the same time, for a voiceless part having no periodicity in the voice unit, a voice unit storage unit that stores the real voice as it is as the voice waveform data in the memory, and a predetermined phonological rule and prosody based on the input phonological symbols and prosodic symbols. According to a rule, a speech synthesis rule section for generating a pitch pattern, and an interpolation frame for synthesizing the voiced portion, By interpolating the envelope information of the rear end frame of the line audio unit and the front end frame of the rear audio unit, a time waveform is obtained along with the fine structure information of the preceding audio unit. A speech synthesizer comprising: a speech synthesizer that generates a synthesized sound based on the speech unit and the pitch pattern as the speech waveform data. 2. A system for analyzing a sequence of inputted characters to find a word,
A sentence analysis unit that detects the boundaries of phrases and basic accents; and, for each voice unit, a voiced portion having periodicity within the voice unit, for each one-pitch period obtained by analysis processing of real voice. Corresponding audio waveform data is stored in the memory for the required number of frames as the audio unit, and at the same time, in the front end frame and the rear end frame of the audio unit, the envelope information and the fine structure obtained by the analysis processing together with the audio waveform data are stored. A voice unit storage unit that stores information together with the information in the memory, and simultaneously stores the actual voice in the memory as the voice waveform data without any periodicity in the voice unit, and an analysis result of the text analysis unit. A voice synthesis rule section for generating a pitch pattern in accordance with a predetermined phoneme rule and a prosody rule based on In the interpolation frame at the time of synthesizing the minute, the envelope information of the rear end frame of the preceding voice unit and the front end frame of the rear voice unit is interpolated, and a time waveform is obtained along with the fine structure information of the preceding voice unit. A voice synthesizing unit that generates a synthesized voice based on the voice unit and the pitch pattern using the time waveform as the voice waveform data of the interpolation frame. 3. For each voice unit, for a voiced portion having periodicity within the voice unit, voice waveform data corresponding to each one-pitch cycle obtained by analyzing the real voice is converted into the voice unit. As many as the required number of frames are stored in the memory, and at the same time, the front end frame and the rear end frame of the audio unit are combined with the audio waveform data together with the envelope information and the fine structure information obtained by the analysis process, and are stored in the memory. For the unvoiced part having no periodicity in the voice unit, the real voice is stored as it is as the voice waveform data in the memory, and the predetermined phoneme rule based on the input phoneme symbol and the prosody symbol is stored. A pitch pattern is generated in accordance with the prosody rules, and in the interpolation frame when the voiced portion is synthesized, By interpolating the envelope information of the rear end frame and the front end frame of the rear audio unit, a time waveform is obtained in combination with the fine structure information of the preceding audio unit, and the time waveform is converted into the audio waveform data of the interpolation frame. Wherein a synthesized voice is generated based on the voice unit and the pitch pattern. 4. Analyzing a sequence of inputted characters, and analyzing words,
Detects the boundaries of phrases and basic accents, and for each voice unit, the voice waveform corresponding to each one-pitch period obtained by the analysis processing of the real voice for the voiced part having periodicity within the voice unit Data is stored in the memory for the required number of frames as the voice unit, and at the same time, the front end frame and the rear end frame of the voice unit are combined with the envelope information and the fine structure information obtained by the analysis processing together with the voice waveform data. For the unvoiced part having no periodicity within the voice unit, the real voice is stored as it is as the voice waveform data in the memory, and the voice data is stored in the memory. Therefore, a pitch pattern is generated, and in the interpolation frame when the voiced part is synthesized, By interpolating the envelope information of the rear end frame and the front end frame of the rear audio unit, a time waveform is obtained in combination with the fine structure information of the preceding audio unit, and the time waveform is used as the audio waveform data of the interpolation frame. , Based on the voice unit and the pitch pattern,
A speech synthesis method characterized in that a synthesized speech is generated. 5. The voice synthesizing apparatus and the voice synthesizing method according to claim 1, wherein a voice based on Japanese is used as said voice.