JPH0887297A - Voice synthesis system - Google Patents

Abstract

PURPOSE: To obtain a highly natural synthesized voice having clearness and rythm close to an actual voice. CONSTITUTION: When a text or a phonetic symbol column is inputted, a voice information retrieving section 1 analyzes an actual voice, retrieves whether or not uttering contents matched with the inputted text or an inputted voice symbol column are in existence in a voice information database 2 stored with an extracted voice feature amount and the corresponding uttering contents and when matched uttering contents exist, the contents are transmitted to a synthesized voice generating section 3. The section 3 generates a synthesized voice by performing a processing corresponding to the voice information. When matched uttering contents do not exist, the inputted text or the inputted phonetic symbol column is transmitted to the section 3 as it is, which generates a synthesized voice based on a synthesized voice generating rule 4.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice synthesizing system for converting an arbitrary input text or an input phonetic symbol string into a synthetic voice and outputting it.

[0002]

2. Description of the Related Art FIG. 10 is a block diagram showing a main configuration of a conventional speech synthesis system, showing a case where an arbitrary input text is converted into synthetic speech and output. 30 in the figure
Reference numeral 4 denotes a language processing unit. The language processing unit 304 applies a word dictionary 303 to text input to determine the reading and accent positions of words in the input text, and analyzes the sentence structure. Control information for controlling the intonation is created and passed to the voice length pattern generation unit 300. Usually, reading of a word, accent position, and intonation control information are all represented by a symbol string called a phonetic symbol string.

The voice length pattern generation unit 300 applies the voice length generation rule 400 to convert a phonetic symbol string into a length information (voice length) pattern of each sound, and outputs the voice length pattern and the phonetic symbol string. It is passed to the FO pattern generation unit 301. The FO pattern generation unit 301 applies the FO generation rule 401 and based on the phonetic symbol string and the voice length pattern, a fundamental frequency (F) which is a physical quantity corresponding to accent and intonation.
O) time change pattern, so-called FO pattern is generated,
The phonetic symbol string, the voice length pattern, and the FO pattern are passed to the voice parameter pattern generation unit 310.

The voice parameter pattern generator 310 is
The voice parameter generation rule 410 is applied to generate a voice parameter pattern based on the phonetic symbol string (especially reading information), the voice length pattern, and the FO pattern, and passes this to the waveform generation unit 311. Here, the voice parameter pattern is usually the RARCOR coefficient (coefficient corresponding to the vocal tract cross-sectional area),
Alternatively, it is a temporal change pattern of a sound source signal such as a formant (vocal tract resonance) frequency and a sound source signal. Further, in the case of a so-called waveform editing method in which short speech waveform units are connected to generate a synthetic speech, the information is connection information such as the type of segment waveform, which is a short speech waveform unit, and connection timing.

The waveform generator 311 generates an actual digital voice waveform based on the passed voice parameter pattern and passes it to the DA converter 5. For example, when the voice parameter pattern is a PARCOR coefficient, the waveform generation unit 311 includes a PARCOR filter and a sound source generation unit, and performs a process of driving the filter with a sound source signal. Further, in the case of the waveform editing method, the segment waveforms are arranged at appropriate positions and a process for smoothly connecting them is performed. The DA converter 5 converts the digital voice waveform generated by the waveform generator 311 into an analog voice waveform and outputs it as a synthesized voice.

[0006]

By the way, in such a conventional speech synthesis system, all acoustic and prosodic processing is performed based on each rule prepared in advance. Since the synthesized voice is generated based on the information, the quality of the synthesized voice is unavoidably deteriorated. Also, since the voice length pattern, the FO pattern and the voice parameter pattern are generated according to the rule, unnatural accent, intonation, etc. There was a problem.

A first object of the present invention is to prepare a plurality of types of voice information databases such as an acoustic information database and a prosody information database, and when the input text is a text registered in these databases, It is to improve the quality of the reproduced voice by reconstructing the voice based on the extracted information and outputting it.

Another object of the present invention is to use a voice waveform database and / or a voice parameter database as an acoustic information database and a voice length database and / or a voice length / speech intensity database and / or voice as a prosody information database. The purpose is to enable speech synthesis corresponding to various input texts or input phonetic symbol strings by using the long / FO pattern database.

[0009]

The principle of the present invention will be described. FIG. 1 is a principle diagram 1 showing the first principle of the present invention,
In the figure, reference numeral 1 indicates a voice information search unit, and 3 indicates a synthetic voice generation unit.

The voice information retrieving unit 1 extracts various voice feature amounts from the real voice that is a voice uttered by a person when a text or a phonetic symbol string is input, and utterance contents (corresponding to which voice feature) It is searched whether or not there is utterance content that matches these input texts or input phonetic symbol strings in the voice information database 2 that stores (labels indicating whether or not they are spoken).

The voice information is a voice waveform, voice parameters, FO, voice strength, voice length, etc. The voice feature amount is a physical feature amount of a voice, a so-called voice parameter, and generally refers to a feature amount in the voice frequency region, which includes a spectrum (frequency intensity) formant (vocal tract resonance frequency), LPC. (Linear prediction coefficient), PAR
There are COR coefficients (coefficients corresponding to vocal tract cross-sectional areas) and the like.

As a result of the search, if there is a matching utterance content, this is passed to the synthetic speech generation unit 3, and if there is no matching utterance content, the input text or the input phonetic symbol string is directly synthesized speech. Pass to the generation unit 3. The synthetic speech generation unit 3 generates synthetic speech from an input text or an input phonetic symbol string based on the synthetic speech generation rule 4.

A first invention is an invention based on this principle. In a speech synthesis system for converting an input text or an input phonetic symbol string into synthetic speech and outputting the synthetic speech, a speech feature quantity extracted from an actual speech is stored. Voice information database,
Search means for searching the voice feature amount stored in the voice information database corresponding to the input text or the input phonetic symbol string, and as a result of the search, the voice feature amount corresponding to the voice information database exists. Is provided with means for forming a voice based on the voice feature amount, and means for generating a synthetic voice according to a predetermined rule when the corresponding voice feature amount does not exist.

FIG. 2 is a principle diagram showing a second principle of the present invention.
In the figure, 10 is an acoustic information retrieval unit, 30 is a prosody information generation unit, and 31 is an acoustic information generation unit. Note that the acoustic information means information in the time domain such as a voice waveform of the voice information, information in the frequency domain such as a spectrum, and the acoustic feature amount is information on the speech synthesis of the information. Means significant information. The prosody information is the prosody (intonation, accent, rhythm,
Intensity), and the prosodic features are the fundamental frequency (FO), which is the physical feature of the prosody corresponding to the intonation and accent, the voice length corresponding to the rhythm, and the voice intensity corresponding to the intensity. Say.

When a text or phonetic symbol string is input, the acoustic information retrieval unit 10 inputs it into an acoustic information database 20 in which various acoustic feature quantities extracted from actual speech and corresponding utterance contents are stored. It is searched whether or not there is utterance content that matches the text or the input phonetic symbol string.
If the matched utterance content exists, this is directly output as acoustic information, and if the matched utterance content does not exist, the input text or the input phonetic symbol string is directly passed to the prosody information generation unit 30.

The prosody information generation unit 30 generates prosody information from a text or phonetic symbol string based on the prosody information generation rule 40, and passes this to the acoustic information generation unit 31. The acoustic information generation unit 31 generates acoustic information from the prosody information based on the acoustic information generation rule 41.

A second invention is based on this principle, and is characterized in that an audio information database storing acoustic feature amounts extracted from actual voices in the voice feature amounts is used as the voice information database.

FIG. 3 is a principle diagram showing a third principle of the present invention.
11 is a prosody information search unit. When a text or phonetic symbol string is input, the prosody information search unit 11 stores a prosody information database 21 in which various prosodic feature quantities extracted from actual speech and corresponding utterance contents are stored.
It is searched whether or not there is utterance content that matches the input text or the input phonetic symbol string. If there is a matching utterance content, the obtained prosody information is used as the acoustic information generation unit 3
If there is no matching utterance content,
The input text or the input phonetic symbol string is directly passed to the prosody information generation unit 30.

The acoustic information generation unit 30 generates prosody information from the input text or the input phonetic symbol string based on the prosody information generation rule 40, and passes this to the acoustic information generation unit 31. The acoustic information generation unit 31 generates acoustic information from the prosody information passed from the prosody information search unit 11 or the prosody information generation unit 30 based on the acoustic information generation rule 41.

A third invention is based on this principle, and is characterized in that a prosodic information database storing prosodic characteristic amounts extracted from real speech in the speech characteristic amounts is used as the speech information database. .

FIG. 4 is a principle diagram showing a fourth principle of the present invention.
In the figure, 10 indicates an acoustic information retrieval unit. When a text or phonetic symbol string is input, the acoustic information search unit 10 inputs an input text or an input phonetic sound into an acoustic information database 20 that stores various acoustic feature amounts extracted from actual speech and corresponding utterance contents. It is searched whether or not there is utterance content that matches the symbol string.

When the utterance contents that match each other are present, this is directly output as acoustic information. If there is no matching voice content, the input text or the input phonetic symbol string is directly passed to the prosody information searching unit 11. Prosody information search unit 1
1 indicates whether or not the utterance content that matches the input text or the input phonetic symbol string exists in the prosody information database 21 in which various prosodic feature quantities extracted from the actual speech and the utterance content corresponding thereto are stored. To search for

When the utterance contents that match each other are present, the prosody information including the prosody feature amount is passed to the acoustic information generation unit 31,
If there is no matching utterance content, the input text or the input phonetic symbol string is passed to the prosody information generation unit 30 as it is. The prosody information generation unit 30 generates prosody information from the input text or the input phonetic symbol string based on the prosody information generation rule 40, and passes this to the acoustic information generation unit 31. The acoustic information generation unit 31 generates acoustic information from the prosody information passed from the prosody information search unit 11 or the prosody information generation unit 30 based on the acoustic information generation rule 41.

A fourth invention is an invention based on this principle. As a speech information database, an acoustic information database storing acoustic characteristic amounts extracted from actual speech and a prosodic characteristic amount extracted from actual speech are stored. It is characterized by using the prosody information database.

A fifth invention is based on the above-mentioned second and fourth principles, and is characterized in that a voice waveform database storing voice waveforms is used as the acoustic information database.

A sixth invention is also an invention based on the above second and fourth principles, and as an acoustic information database,
It is characterized by using a voice parameter database that stores spectrum, vocal tract cross-sectional area, or formant frequency.

The seventh invention is also an invention based on the above second and fourth principles, and as an acoustic information database,
It is characterized by using a voice waveform database and a voice parameter database.

An eighth invention is an invention based on the above-mentioned third and fourth principles. As a prosodic information database, only the voice length of voice length, voice intensity, and fundamental frequency, or any two or more of them are used. It is characterized by using a database storing.

A ninth aspect of the present invention is also based on the third and fourth principles, and is characterized in that a voice length / speech intensity database storing voice length and voice intensity is used as a prosody information database.

A tenth aspect of the present invention is also based on the above third and fourth principles, and is characterized in that a voice length / FO database storing voice length and fundamental frequency is used as a prosody information database.

An eleventh invention is also an invention based on the above third and fourth principles, and is characterized in that a voice length database storing only voice length is used as a prosody information database.

The twelfth invention is also based on the above-mentioned third and fourth principles, and as a prosodic information database, a voice length / speech intensity / FO database, a voice length / speech intensity database, and a voice length / FO database. Alternatively, any two or more databases of the voice length database are used.

[0033]

According to the first aspect of the present invention, it is possible to output high quality synthetic speech by using the database storing the speech feature amount.

According to the second aspect of the present invention, by using the acoustic feature quantity, a synthetic speech having a high degree of intelligibility close to an actual speech can be obtained.

According to the third aspect of the invention, since the prosodic feature quantity is used, it is possible to obtain a highly natural voice close to an actual voice.

According to the fourth aspect of the present invention, a function having both the second and third functions can be obtained.

According to the fifth aspect of the invention, by using a voice waveform as the acoustic information data, a voice with a high degree of naturalness and clarity can be obtained.

According to the sixth aspect of the invention, synthetic speech having a low degree of intelligibility but a small amount of data and high naturalness can be obtained.

According to the seventh aspect of the invention, it is possible to obtain an operation which has both functions of the fifth and sixth aspects.

In the eighth invention, by using one or more of the voice length, the voice intensity, and the FO, a synthetic voice having a highly natural prosody can be obtained.

According to the ninth invention, a synthetic voice having a rhythm with high naturalness can be obtained.

According to the tenth aspect of the present invention, a synthetic voice having highly natural rhythm, intonation and accent can be obtained.

According to the eleventh invention, a synthetic voice having a highly natural rhythm can be obtained with a small amount of data.

According to the twelfth invention, a synthetic speech having a natural prosody can be obtained.

[0045]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention will be specifically described below with reference to the drawings showing the embodiments. (Embodiment 1) Embodiment 1 embodies the first, second, and third principles, and FIG. 5 is a block diagram showing the configuration of Embodiment 1 of the present invention. Reference numeral 100 in FIG. 5 denotes a voice waveform search unit.

When the text or phonetic symbol string is input, the voice waveform search section 100 searches the voice waveform database 200. The voice waveform database 200 includes a phonetic symbol string indicating utterance content and voice waveform data (P
If a plurality of pairs of CM data or data that has been information-compressed by an encoding technique such as ADPCM are stored, and if there is utterance content that matches the input text or the input phonetic symbol string, the corresponding voice The waveform is immediately converted to DA converter 5
If there is no matching utterance content, the input text or the input phonetic symbol string is directly passed to the voice parameter searching unit 101.

The voice parameter retrieving unit 101 extracts voice parameters from real voice, for example PARCOR and LS.
A search is made as to whether or not there is utterance content that matches the input text or input phonetic symbol string in the voice parameter database 201 that stores P, formant frequencies, and utterance content corresponding thereto.

If there is a matching utterance content, the voice parameter is passed to the waveform generating section 311, and if there is no matching utterance content, the input text or the input phonetic symbol string is passed to the prosody information searching section 11 as it is. . The prosody information retrieval unit 11 determines whether or not the utterance content that matches the input text or the phonetic symbol string exists in the prosody information database 21 that stores various prosodic feature quantities extracted from the actual speech and the corresponding utterance content. Search whether or not. If the matched utterance content exists, the prosody information is passed to the voice parameter pattern generation unit 310, and if the matched utterance content does not exist, the input text or the input phonetic symbol string is passed to the prosody information generation unit 30 as it is.

The prosody information generation unit 30 uses the prosody information generation rule 4
Based on 0, prosody information is generated from the input text or the input phonetic symbol string, and this is passed to the voice parameter pattern generation unit 310. The voice parameter pattern generation unit 310 generates a voice parameter pattern from the passed prosody information, that is, a phonetic symbol string (especially reading information), a voice length and a voice intensity pattern, and an FO pattern, based on the voice parameter generation rule 410. This is passed to the waveform generation unit 311. The generated voice parameter pattern is specifically RARCO
It is connection information such as the R coefficient or the temporal change pattern of the formant frequency and the sound source signal, and in the case of a so-called waveform editing method, the type of a segment waveform, which is a short unit of the audio waveform, the connection timing, and the like. .

The waveform generation unit 311 generates an actual voice waveform based on the voice parameter pattern passed from the voice parameter search unit 101 or the voice parameter generation unit 310, and passes it to the DA conversion unit 5. In the voice waveform generation process, when the voice parameter pattern is the RARCOR coefficient, the waveform generation unit 311 includes a PARCOR filter and a sound source generation unit. The RARCOR filter is driven by the sound source signal. The waveform is placed at an appropriate position, and the process of connecting them smoothly is performed. The DA conversion unit 5 converts the digital voice waveform generated by the waveform generation unit 311 or extracted from the voice waveform database 200 by the voice waveform search unit 100 into an analog voice waveform, and outputs the analog voice waveform.

In the first embodiment, the voice waveform database 200 and the voice parameter database 2 are used.
By using 01, high quality synthetic speech can be generated, and by using the prosody information database 21, natural synthetic speech can be generated.

(Embodiment 2) Embodiment 2 embodies the third principle, and FIG. 6 is a block diagram showing the configuration of Embodiment 2 of the present invention. Reference numeral 110 in FIG. 6 denotes a voice length / voice strength / FO pattern search unit. When a text or phonetic symbol string is input, the voice length / voice strength / FO pattern search unit 61 stores the voice length / voice strength / FO pattern extracted from the actual voice and the voice length / storing content corresponding to this. The voice strength / FO database 210 is searched for whether or not there is utterance content that matches the input text or the input phonetic symbol string.

If there is a matching utterance content, the voice length / sound intensity / FO pattern is passed to the acoustic information generating section 31, and if there is no matching utterance content, the input text or the input phonetic symbol string is input. It is passed to the voice length / FO pattern search unit 111 as it is. The speech length / speech strength / FO database contains information indicating utterance content, a speech length pattern corresponding to the synthesis unit such as a phoneme and a syllable, and a speech strength pattern which is a temporal variation pattern of the speech strength. , A plurality of sets of FO patterns, which are temporal change patterns of reference frequencies, are stored.

The voice length / FO pattern search unit 111 searches the voice length / FO database 211 for utterance contents matching the input text or the input phonetic symbol string. When the utterance contents that match each other are present, the voice length / FO pattern data is used as the voice intensity pattern generation unit 302.
If there is no matching utterance content, the input text or the input phonetic symbol string is passed as it is to the voice length / speech intensity pattern search unit 112. The voice length / speech intensity pattern search unit 112 uses the voice length / speech intensity database 2
It searches whether or not there is utterance content that matches the input text or the input phonetic symbol string in 12, and if there is utterance content that matches, the voice length / speech intensity pattern is passed to the FO pattern generation unit 301. If it does not exist, the input text or the input phonetic symbol string is directly passed to the voice length pattern search unit 113.

The voice length pattern search unit 113 stores a voice length pattern, which is a time length for each synthesis unit of phonemes, syllables, and the like, and a voice length database 2 in which utterance contents corresponding thereto are stored.
13 is searched for whether or not the input text or the input phonetic symbol string exists, and if it exists, it is passed to the FO pattern generation unit 301, and if it does not exist, the input text or the input phonetic symbol string is input. As it is, the voice length pattern generation unit 3
Pass to 00. The voice length pattern generation unit 300 converts the input text or the input phonetic symbol string into a voice length pattern based on the voice length generation rule 400, and passes this to the FO pattern generation unit 301.

Based on the FO generation rule 401, the FO pattern generation unit 301 generates a FO pattern, which is a physical quantity corresponding to an accent or intonation, from a voice length / voice intensity pattern or a voice length pattern and a phonetic symbol string. This is used together with the phonetic symbol string to generate the voice intensity pattern generation unit 3
Pass to 02. The voice strength pattern generation unit 302, based on the voice strength generation rule 402, a phonetic symbol string, a voice length / FO.
A voice intensity pattern, which is a temporal change pattern of voice intensity, is generated from the pattern or the FO pattern and is passed to the acoustic information generation unit 31. The acoustic information generation unit 31 generates actual acoustic information based on the acoustic information generation rule 41.

In the second embodiment, the voice length
Voice strength / FO database 210, voice length / FO database 211, voice length / voice strength database 212
By using the voice length database 213 and the like, it is possible to generate synthetic voice having natural intonation, accent and rhythm.

(Embodiment 3) Embodiment 3 embodies the first, second, third, and fourth principles, and FIG. 7 is a block diagram showing the configuration of Embodiment 3. Reference numeral 304 in FIG. 7 denotes a language processing unit. When a text input is made, the language processing unit 304 refers to the word dictionary 303 to analyze the reading of the words in the input text, the accent information and the structure of the sentence, and outputs a phonetic symbol string consisting of control information for controlling the intonation. It is created and passed to the voice waveform search unit 100.

The voice waveform search unit 100 uses the phonetic symbol string created by the language processing unit as a key, and the voice waveform database 2
00 to search the voice waveform data corresponding to the utterance content that matches the key, pass the voice waveform data to the DA conversion unit 5 when the utterance content that matches the key exists, and the input phonetic symbol string when it does not exist. Is directly passed to the voice parameter search unit 101. The voice parameter search unit 101 uses the input phonetic symbol string as a key to store the voice parameter database 2
The voice parameter data corresponding to the utterance content that matches the key is searched from 01, and if the utterance content that matches the key exists, the voice parameter data is passed to the waveform generation unit 311, and if they do not match, the input phonetic symbol string is output. It is passed to the voice length / FO pattern search unit 110 as it is.

The voice length / FO pattern search unit 110 uses the input phonetic symbol string as a key to set the voice length / FO database 2
The voice length corresponding to the utterance content from 10 to F, F
The O data is searched, and if there is a matching utterance content, the voice length / FO data is passed to the voice parameter pattern generation unit 310, and if it does not exist, the input phonetic symbol string is directly searched for the voice length pattern. Hand it over to section 111.

Voice length / voice intensity / FO database 21
In 0, a plurality of pairs of a phonetic symbol string indicating a voice content, a corresponding voice length pattern which is a time length for each synthesis unit such as a phoneme and a syllable, and a FO pattern which is a time change pattern of a fundamental frequency are stored. Has been done. Voice length pattern search unit 1
Reference numeral 11 is an input phonetic character string used as a key to search the voice length database 211 for voice length data corresponding to the utterance content that matches the key.
It is passed to the FO pattern generation unit 301, and if it does not exist, the input phonetic symbol string is passed as it is to the voice length pattern generation unit 300. The voice length pattern generation unit 300 converts the input phonetic symbol string into a voice length pattern based on the voice length generation rule 400, and passes this to the FO pattern generation unit 301 together with the input phonetic symbol string.

Based on the FO generation rule 401, the FO pattern generation unit 301 generates an FO pattern, which is a physical quantity corresponding to an accent or intonation, from an input phonetic symbol string and a voice length pattern, and this is generated as an input phonetic symbol string. Along with it, it is passed to the voice parameter pattern generation unit 310. The voice parameter pattern generation unit 310 generates a voice parameter pattern from the input phonetic symbol string and the voice length / FO pattern or the FO pattern based on the voice parameter generation rule 410, and passes this to the waveform generation unit 311.

The waveform generation unit 311 digitally generates a voice waveform from the voice parameter pattern obtained from the voice parameter search unit 101 or the voice parameter pattern generation unit 310, and passes it to the DA conversion unit 5. The DA conversion unit 5 is generated by the waveform generation unit 311 or the voice waveform search unit 100.
The digital voice waveform searched in is converted into an analog voice waveform and output as a synthetic voice.

In the third embodiment, the language processing unit 304 first performs language processing to read the input text and convert it into a phonetic symbol string indicating an accent, and then searches the voice waveform database. Even if the same word is written in various notations such as kanji, hiragana, katakana, and different syllabary, there is an effect that the capacity of the database can be reduced.

(Embodiment 4) Embodiment 4 embodies the first, second, third, and fourth principles, and FIG. 8 is a block diagram showing the configuration of Embodiment 4. In the fourth embodiment, the language processing unit 304 in the third embodiment shown in FIG. 7 is used.
Is substantially the same as that of the voice length pattern search unit 111 and the voice length pattern generation unit 300. For such an embodiment 4 is now "Hello.", "Today is March 3 days.", Specifically to the process described for the case where 3 statement of "Thank you." Has been input .

[0066] It should be noted that the speech waveform data of "Hello" is stored in the speech waveform database 200, voice length · FO pattern voice length of "Thank you". It is assumed that it is stored in the FO database 210, and "Today is March 3rd." Is not stored in any database. If 3 statement now above is input, the speech waveform retrieval unit 100 when the statement is entered in "Hello" searches the speech waveform database 200, searches the speech waveform data that is stored here, the audio The waveform data is sent directly to the DA converter 5, and the synthesized voice is output as "Konichiwa".

In the case of "Thank you", the voice waveform search unit 100 and the voice parameter search unit 10
1 is not searched, but voice length / FO pattern search unit 1
10 detects this by searching the voice length / FO database 210, passes it to the voice parameter pattern generation unit 310, and outputs “Arigatogogo zaimashita” as a synthesized voice via the waveform generation unit 311. On the other hand, in the case of "Today is March 3," it is not stored in any database and reaches the language processing unit 304, where the reading and accent position of the word is determined, the structure of the sentence is analyzed, and the intonation is determined. The control information for controlling is passed to the voice pattern generation unit 300. Voice information pattern generation unit 300
Generates a voice length pattern based on the voice length generation rule 400, and passes this to the FO pattern generation unit 301.

Thereafter, in the same manner as in the third embodiment, the FO parameter pattern generating section 301, the voice parameter pattern generating section 310, and the waveform generating section 311 are passed to the DA converting section 5, and the synthesized voice of "Kyoha Sangatsu Mikades" is synthesized. Is output.

In the fourth embodiment, the input text is stored in the database by first searching the voice waveform database 200, the voice parameter database 201, the voice length / FO database 210, and the voice length database 211. When it exists, it is not necessary to perform the subsequent search processing, and the synthesized speech can be generated at high speed.

(Embodiment 5) Embodiment 5 embodies the first, second, third, and fourth principles, and FIG. 9 is a block diagram showing the configuration of Embodiment 5. 10a, 1 in FIG.
Reference numeral 0b indicates an acoustic information search unit, and reference numerals 11a and 11b indicate prosody information search units. When the text information is input, the acoustic information search unit 10a searches the acoustic information database 20 using the input text as a key, and if there is utterance content that matches the input text, directly outputs it as acoustic information. If there is no matching utterance content, the input text is directly passed to the prosody information search unit 11a.

The prosody information processing 11a searches the prosody information database 21 using the input text as a key, and if there is utterance content that matches the input text, passes it to the acoustic information generation unit 31 and the matching utterance content If it does not exist, it is passed to the language processing unit 304 as it is. Language processing unit 3
Reference numeral 04 refers to the word dictionary 303, analyzes the input text, converts it into a phonetic symbol string, and converts this into a phonetic information retrieval unit 10
pass to b.

The acoustic information retrieving unit 10b searches the acoustic information database 20 using the phonetic symbol string as a key. If there is utterance content that matches the phonetic symbol string, the prosody information database is used to generate a search, prosody information. Without generating
The acoustic information extracted from the acoustic information database 20 is directly output. If there is no matching utterance content, the acoustic information is directly passed to the prosody information searching unit 11b.

The prosody information searching unit 11b searches the prosody information database 21 using the phonetic symbol string as a key, and if there is utterance content that matches the phonetic symbol string, this is directly used as acoustic information in the acoustic information generating unit. 30 to the prosody information generation unit 30 if it does not exist. The prosody information generation unit 30 generates prosody information from the input phonetic symbol sequence based on the prosody information generation rule 40, and the prosody information generation unit 31 generates the prosody information.
Hand over to. The prosody information generation unit 31 generates acoustic information from the prosody information using the acoustic information generation rule 41, and outputs this.

In the fifth embodiment, since the acoustic information database 20 and the prosody information database 21 are searched first, if the input text exists in the acoustic information database 20 or the prosody information database 21, The subsequent processing can be omitted, and after the database is searched, linguistic processing is performed, the input text is converted into a phonetic symbol string, and then the acoustic information database and the prosody information database are searched. Even if it is written, matching with the utterance content stored in the database is possible.

[0075]

According to the first aspect of the present invention, it is of course possible to output a high-quality synthesized voice by using a voice information database that stores voice feature amounts obtained by analyzing and extracting real voice. If it does not exist in the database, it is possible to generate synthetic speech from a wide range of input texts and input phonetic symbol strings by performing speech synthesis according to rules.

In the second invention, the actual voice is analyzed,
By using the extracted acoustic feature amount, a synthesized voice with high intelligibility close to real voice can be obtained.

According to the third aspect of the invention, by using the prosodic feature quantity obtained by analyzing the real voice, a synthetic voice with high naturalness close to the real voice can be obtained.

The fourth invention has both effects of the second and third inventions.

According to the fifth aspect of the invention, by using a voice waveform as the acoustic database, a synthetic voice with high naturalness and clarity can be obtained.

According to the sixth aspect of the invention, although the clarity is lower than that of the fifth aspect, a synthetic voice having a high degree of clarity can be obtained with a small amount of data.

In the seventh invention, the effects of the fifth and sixth inventions can be combined.

According to the eighth aspect of the present invention, by using the voice length, voice intensity, and FO as the prosody information, it is possible to obtain a synthetic voice having prosody information with high naturalness.

According to the ninth aspect of the invention, by using the voice length and voice intensity as the prosody information, a synthetic voice having a highly natural rhythm can be obtained.

According to the tenth aspect of the invention, by using the voice length / FO pattern as the prosody information data, a synthesized voice with highly natural rhythm, intonation and accent can be obtained.

In the eleventh aspect of the invention, by using the voice length as the prosody information data, a synthetic voice having a highly natural rhythm can be obtained with a small amount of data.

According to the twelfth invention, a synthetic speech having a more natural prosody can be obtained.

[Brief description of drawings]

FIG. 1 is a principle diagram showing the principle of the present invention.

FIG. 2 is a principle diagram showing another principle of the present invention.

FIG. 3 is a principle diagram showing still another principle of the present invention.

FIG. 4 is a principle diagram showing still another principle of the present invention.

FIG. 5 is a block diagram showing the configuration of the first embodiment of the present invention.

FIG. 6 is a block diagram showing a configuration of a second exemplary embodiment of the present invention.

FIG. 7 is a block diagram showing a configuration of a third exemplary embodiment of the present invention.

FIG. 8 is a block diagram showing a configuration of a fourth embodiment of the present invention.

FIG. 9 is a block diagram showing a configuration of a fifth embodiment of the present invention.

FIG. 10 is a block diagram showing a configuration of a conventional speech synthesis system.

[Explanation of symbols]

 1 Speech Information Retrieval Section 2 Speech Information Database 3 Synthetic Speech Generation Section 4 Synthetic Speech Generation Rule 5 DA Conversion Section 10 Speech Information Retrieval Section 11 Prosody Information Retrieval Section 20 Speech Information Database 21 Prosody Information Database 30 Prosody Information Generation Section 31 Acoustic Information Generation Part 40 Prosody Information Generation Rule 41 Acoustic Information Generation Rule 100 Speech Waveform Search Unit 101 Speech Parameter Search Unit 110 Speech Length / Voice Strength / FO Pattern Search Unit 111 Speech Length / FO Pattern Search Unit 112 Speech Length / Speech Strength Pattern Search Unit 113 Voice length pattern search unit 200 Voice waveform database 201 Voice parameter database 210 Voice length / voice intensity / FO database 301 FO pattern generation unit 302 Voice intensity pattern generation unit 310 Voice parameter pattern generation unit 311 Waveform generation unit

Claims (12)

[Claims] 1. A speech synthesis system for converting an input text or an input phonetic symbol string into synthetic speech and outputting the synthesized speech, a speech information database storing speech features extracted from actual speech, and an input text or an input phonetic symbol. Search means for searching a voice feature amount stored in the voice information database corresponding to a column, and based on the voice feature amount if a voice feature amount corresponding to the voice information database exists as a result of the search. A voice synthesizing system comprising: a means for forming a voice and a means for generating a synthesized voice according to a predetermined rule when a corresponding voice feature amount does not exist. 2. The voice synthesis system according to claim 1, wherein an acoustic information database storing acoustic feature amounts extracted from actual voices in the voice feature amounts is used as the voice information database. 3. The prosody information database that stores prosody features extracted from real voices in the voice features is used as the voice information database.
The voice synthesis system described. 4. The voice information database is characterized by using an acoustic information database storing acoustic feature amounts extracted from real voice and a prosody information database storing prosodic feature amounts extracted from real voice. Item 1
The voice synthesis system described. 5. The voice synthesis system according to claim 2, wherein a voice waveform database storing voice waveforms is used as the acoustic information database. 6. The voice synthesis system according to claim 2, wherein a voice parameter database that stores a spectrum, vocal tract cross-sectional area, or formant frequency is used as the acoustic information database. 7. The voice synthesis system according to claim 2, wherein a voice waveform database and a voice parameter database are used as the acoustic information database. 8. The prosodic information database is a voice length,
5. The voice synthesis system according to claim 3, wherein only a voice length of the voice intensity and the fundamental frequency, or a database storing any two or more is used. 9. The speech synthesis system according to claim 3, wherein a speech length / speech intensity database storing speech length and speech intensity is used as the prosody information database. 10. The voice synthesis system according to claim 3, wherein a voice length / FO database storing a voice length and a fundamental frequency is used as the prosody information database. 11. The voice synthesis system according to claim 3, wherein a voice length database storing only voice length is used as the prosody information database. 12. The prosodic information database includes at least two databases of a voice length / voice strength / FO database, a voice length / voice strength database, a voice length / FO database, and a voice length database. The speech synthesis system according to Item 3 or 4.