JPH0863187A - Speech synthesizer - Google Patents

Abstract

PURPOSE: To make a synthesized speech easy to catch and to give it a natural rhythm by generating time change patterns of respective basic frequencies of fixed, variable information, successively connecting them, generating the time change pattern of the basic frequency of a sentence and synthesizing the speech. CONSTITUTION: A text analysis means generates a phonogram line from an input sentence to output it to an acoustic parameter generation means 11. Related to a fixed form part, in a fixed form FO pattern/duration time length generation means 8, and related to an unfixed form part, in an unfixed form FO pattern/duration time length generation means 9, an FO pattern and a duration time length are generated respectively. These FO pattern and duration time length are connected successively in an FO pattern/duration time length connecting/editing means 10, and the FO pattern and the duration time length of the whole sentence are generated. The acoustic parameter generation means 11 generates an acoustic parameter based on the phonogram line. Further, a speech signal generation means 12 generates and outputs a speech signal from the FO pattern, etc.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech synthesizer,
Especially used for traffic information and voice services for weather conditions,
The present invention relates to a voice synthesizing device for synthesizing a voice composed of fixed information (hereinafter, referred to as a fixed part) common to all of a group of messages to be combined and variable information (hereinafter, referred to as an atypical part) that is not common to message groups. .

In recent years, the demand for labor saving and mechanization of society in general has become stronger and stronger, and the fields of various voice services are no exception. Currently, voice synthesis is used for traffic information, weather condition voice service, bank transfer inquiry service, etc. The device is in use. Therefore, it is necessary for the speech synthesizer to provide synthetic speech that is easy to hear and has a natural prosody.

[0003]

2. Description of the Related Art In a conventional speech synthesizer, a fixed part has
A recording / editing method that reproduces a voice that has been recorded in advance, or an analysis and synthesis method that accumulates a voice that has been converted into a voice parameter and synthesizes the voice from the parameter is used. In addition, atypical parts such as proper nouns and numbers use rules from character strings to generate a voice by using a rule synthesizing method, and voices synthesized by each method are connected or switched and output. It was common.

FIG. 9 shows a block diagram of a speech synthesizer according to the prior art. In the figure, 1 is a text input means, 2 is a text analysis means, 3 is a fixed part synthesis means, 4 is an atypical part synthesis means, 5 is an output voice connection means, and 6 is a voice output means. The text input means 1 analyzes the text in the text analysis means 2 while referring to the word dictionary. As a result, the fixed part is input to the fixed part synthesizing means 3 and a voice is synthesized from the accumulated fixed part voice data. The part consisting of variable information is input to the atypical part synthesizing means 4 to perform rule composition from a character string.
The voices synthesized by the respective synthesizing means are connected by the output voice connecting means 5 so as to be continued as a sentence, and output via the voice output means 6.

[0005]

In view of the voice quality, however, the voice quality of the rule synthesis system is inferior to that of the recording / editing system and the analysis / synthesis system at present.

Therefore, in a voice in which a fixed part based on the recording / editing method or the analysis / synthesis method and an atypical part based on the rule synthesis method are connected, there is a gap in the quality between the fixed part and the atypical part, and important information in the sentence is included. There was a problem that the atypical parts were difficult to hear. On the other hand, it is easier to hear if the whole sentence is generated with the same quality, and especially since the speech quality of the rule synthesis method has improved due to technological improvements in recent years, it is sufficient to synthesize all by rule synthesis. It has come to be practical. Of course, if you use the rule composition method, even if you want to change the fixed part,
You can save the trouble of re-recording the voice.

By the way, when synthesizing a voice from a kanji / kana mixture sentence that we use in our daily lives, the rule synthesizing method is different from the recording / editing method or the analysis / synthesis method in that the natural prosody is referred to while referring to the dictionary and the rules. It is necessary to generate (intonation, accent, pose, etc.). There are the following two problems in this process.

The first problem is in the process of analyzing a kanji / kana mixed sentence and generating a phonetic character string. Here, the phonetic character string is a phoneme (in Japanese, it is almost equal to the Roman character notation) or syllable (in Japanese, it is almost equal to the kana character notation) column, and the notation indicating the position of the pause and the accent. It is the included character string. Since Japanese is not written in words and there are many ways to read kanji, when trying to generate phonetic strings from dictionaries and rules, misreading, incorrect accents, insertion of unnatural poses, etc. It happens frequently. The first problem is solved by synthesizing a character string rule extracted from a voice conversion input string file as a storage means for storing a previously created input character string containing prosody information (see Japanese Patent Laid-Open No. 4-107598). However, it is required to reduce the construction cost.

The second problem is in the process of generating acoustic (physical) parameters from a phonetic character string. For example, intonation is a change in voice pitch and is generally controlled using a time change pattern (hereinafter, referred to as F0 pattern) of a fundamental frequency which is the lowest frequency included in voiced sound. . This is represented by a time series of fundamental frequencies every few milliseconds (msec). The famous rules for generating this F0 pattern from the phonetic character strings are the Fujisaki model and the point-pitch model, but there are subtle changes depending on the human complex vocalization mechanism, content, and meaning. It is difficult to find the F0 pattern to be performed by a simple rule. Further, the time length of each phoneme or syllable is set to an appropriate value so that the utterance becomes natural without being caught or delayed. However, this time length is not uniquely determined by the type of phoneme or syllable, but is complicatedly influenced by the position in the sentence in which this phoneme or syllable is placed and the surrounding phonological environment. It is something that cannot be obtained by such rules.

[0010]

FIG. 2 is a conceptual diagram of the present invention. In the following, this figure and the example sentence "Tonight's [Tokyo] region's weather will be [clear]" will be explained.

[0011] The text consists of a typical part "Tonight ... the weather in the region ..." and an atypical part "Tokyo" and "sunny". Each atypical part is "Kanagawa Prefecture". It is possible to replace with a word like "rain". When synthesizing such a sentence, with respect to the fixed form part, the F0 pattern and the duration length of the fixed form part are extracted from the voice uttered by a person, and for example, in the case of the F0 pattern, the fundamental frequency value every several msec. As a time series of, if the duration is long, it is accumulated as a series of the length of each phoneme. Regarding the atypical parts, the F0 patterns of all combinations of the number of syllables such as words or syllables expected to be input to the atypical parts and the accent type are accumulated, and the input sentence or the phonetic character that is analyzed is stored. The F0 pattern having the same combination of the syllable number and the accent type is read from the column. Since this F0 pattern is determined not only by the number of syllables and accent types but also by the F0 pattern of the whole sentence, the F0 patterns have different ones depending on where they are inserted in the fixed part, and the selection is made. There is a need to. For example, since the word "Tokyo" is a 4-mora 0 type, a 4-mora 0 type F0 pattern is selected from the patterns to be inserted at the position "Tonight ... The duration of the atypical part is generated by the rule. The F0 pattern of the entire sentence is created by connecting the searched (or generated) F0 pattern and the duration length separately to the fixed part and the non-fixed part. The F0 pattern is connected continuously throughout the sentence.

Further, even if the F0 pattern is not stored in the atypical part and is generated by the rule, the F
Higher quality voice can be obtained than when all 0 patterns are generated by the rule.

[0013]

The principle of the present invention is shown in FIG. In the figure, 1 is a text input means, 7 is a text analysis means, and 8 is a fixed part F.
0 pattern / duration generation means, 9 atypical part F0 pattern / duration generation means, 10 F0 pattern / duration connection editing means, 11 acoustic parameter generation means, 12 audio signal generation means, 6 Indicates audio output means, respectively. The text to be combined is input to the text input means 1. The text analysis unit 7 separates the input text into a non-standard part and a standard part. If the input text is a normal kana-kana mixed sentence, it is necessary to perform text analysis such as that used for rule composition of arbitrary sentences in order to separate it into a fixed part and an unfixed part. And the atypical part can be separately input, it is sufficient to simply output the atypical part and the atypical part to the respective F0 pattern / duration generation means. Further, the text analysis means 7 generates a phonetic character string (phoneme string or syllable string) from the input sentence and outputs it to the acoustic parameter generation means 11. For the standard part, the standard part F0 pattern / duration generation means 8 is used. For the non-standard part, the non-standard part F0 is used.
In the pattern / duration generation means 9,
Generate F0 pattern and duration. These F
The 0 pattern and the duration length are sequentially connected in the F0 pattern / duration duration connection editing means 10 to generate the F0 pattern and duration length of the entire sentence. The acoustic parameter generation means 11 generates acoustic parameters such as formants based on phonetic character strings such as phoneme strings or syllable strings. The acoustic parameters are determined by the synthesizing method used in the audio signal generating means 12. In addition, there is a waveform editing method that directly edits the waveform as a synthesizing method, and when this method is used, the waveform connection information is generated as an equivalent parameter instead of an acoustic parameter. Handle it by including it in the parameter. Audio signal generation means 1
In 2, the audio signal is generated from the F0 pattern, the duration, and the acoustic parameter, and is output from the audio output means 6.

[0014]

EXAMPLE Three levels can be considered for the F0 pattern generation method. The first level is F0 extracted from natural speech.
This is a method of accumulating the patterns as they are in the time-series format of the fundamental frequency and reading them at the time of synthesis, and is expected to produce the most natural speech synthesis. The second level is a method in which the F0 pattern of natural speech is approximated by a model, the parameters of the model are accumulated, and the parameters are converted into the time series form of the fundamental frequency at the time of synthesis. Third
Is a method of regularly generating model parameters from text analysis results and generating a time series of fundamental frequencies from the parameters.

There are two possible levels of the duration generating method. The first level is a method of accumulating the durations extracted from the natural speech as they are as a sequence of durations and reading them at the time of synthesis. The second level is a method of regularly generating the above time length from the text analysis result. As the F0 pattern of the non-standard part and the standard part and the duration length generation method, combinations of the above levels can be considered. These will be described below as examples.

A block diagram of the first embodiment of the present invention is shown in FIG. This embodiment corresponds to claims 2, 4, 8 and 9 of the patent. In the figure, 011 is a text input part, 71 is a text analysis part, 72 is a fixed / atypical judgment part, 73 is an output switching part, 74 is a word dictionary, 75 is a fixed part sentence example storage part, 8
1 is a fixed part duration reading unit, 82 is a fixed part F0 pattern reading unit, 83 is a fixed part duration accumulation unit, 8
4 is a fixed part F0 pattern storage part, 91 is an atypical part duration generation part, 92 is an atypical part F0 pattern reading part,
93 is an accent dictionary, 94 is an atypical part F0 pattern storage part, 101 is a duration connection editing part, 102 is a F0 pattern connection editing part, 111 is an acoustic parameter generation part, 1
Reference numeral 12 is an acoustic parameter storage unit, 121 is an audio signal generation unit, and 61 is an audio output unit.

The fixed part F0 pattern extracted from the natural voice of the fixed part is stored in advance in the fixed part F0 pattern accumulating portion 84, and the fixed part F0 patterns of all combinations of the syllable number and the accent type are stored for the fixed part. Is stored in the non-standardized section F0 pattern storage section 94, and the standardized section duration length extracted from the natural voice of the standardized section is stored in the standardized section duration storage section 83. The text to be combined is input to the text input unit 011. When the input is a mixed kanji and kana notation, in the text analysis unit 71,
The text is analyzed with reference to the word dictionary 74. The fixed form / atypical form determination unit 72 refers to the fixed form example stored in the fixed form sentence example storage unit 75, and separates the analysis result into a fixed form part and an atypical form part. The output switching unit 73 outputs the fixed-form portion and the non-fixed-form portion to the F0 pattern generation unit for their respective durations. At this time, as a result of analyzing the text, a phonetic character string (a phoneme string or a syllable string) of the input text is output to the acoustic parameter generation unit 111.

Regarding the standard part, the standard part duration reading part 81 reads the duration from the standard part duration storage part 83, and the standard part F0 pattern reading part 82 reads the standard part F0 pattern storage part. The F0 pattern is read from 84 and output to the F0 pattern connection editing unit 102 via the duration connection editing unit 101, respectively. For the non-standard part, the non-standard part duration generation unit 91 generates the duration length according to a rule. In general, the duration generation according to a rule is performed by searching a duration table for each phoneme or syllable in an atypical part and correcting the duration table. Next, in the non-standard part F0 pattern reading unit 92,
The accent of the word of the atypical part is acquired from the accent dictionary 93, and the read F0 pattern is referenced from the atypical part F0 pattern accumulating unit 94 based on the number of syllables and the accent type, and the read F0 pattern is connected to the duration connection editing unit 101 and the F0 pattern connection. Editorial department 1
Output to 02. In the duration connection editing unit 101, the phoneme time lengths of the fixed part and the atypical part are connected in order,
Create a sequence of durations for the entire sentence. In the F0 pattern connection editing unit 102, the F and F
The 0 patterns are connected in order, and the F0 pattern of the whole sentence is created. Since the F0 pattern is continuous during utterance, if there is discontinuity in each of the F0 patterns read by the two fixed parts and the non-fixed part, editing such as appropriate smoothing must be performed.

On the other hand, in the acoustic parameter generator 111,
Acoustic parameters are generated based on the input phonetic character string.
The acoustic parameter storage unit 112 stores acoustic parameters. The acoustic parameter mentioned here is a value obtained by digitizing audio data by using an audio generation model in order to compress the data volume, such as formant and PARCO.
There are types such as R and LSP. The synthesis methods using these acoustic parameters are the formant synthesis and the PAR, respectively.
Called COR synthesis and LSP synthesis, the audio signal generation unit 121
Is realized by In addition, there is a waveform editing method that directly edits the waveform as a synthesizing method.When this method is used, the waveform connection information is generated as an equivalent to the acoustic parameter, but here,
It is included in the acoustic parameters. The acoustic parameters are accumulated for each phonetic character or in a unit that is subdivided according to the preceding and following phoneme environments. By reading this according to the phonetic character string and connecting them, the acoustic parameter string of the synthetic sentence is generated. The voice signal generation unit 121 generates a voice signal from the duration time of the synthetic sentence, the F0 pattern, and the acoustic parameter sequence generated as described above. The voice output unit 61 DA-converts the voice signal to output it as synthesized voice.

A block diagram of the second embodiment of the present invention is shown in FIG. This embodiment corresponds to claims 3 and 5 of the patent. In this embodiment, the fixed part F0 pattern reading part 82 and the fixed part F0 pattern storage part 84 of the first embodiment are replaced by the fixed part F0.
In the parameter reading unit 85, the fixed part F0 pattern generation unit 86, and the fixed part F0 parameter storage unit 87, the non-fixed part F0 pattern reading unit 92 and the non-fixed part F
The 0 pattern accumulating section 94 is replaced with an atypical section F0 parameter reading section 95, an atypical section F0 pattern generating section 96, and an atypical section F0 parameter accumulating section 97.

In this embodiment, the F0 pattern extracted from the natural voice is approximated by a model in advance, and its parameters are stored in the fixed part F0 parameter storage part 87 and the non-fixed part F0 parameter storage part 97. When synthesizing a voice, regarding the fixed form part, in the fixed form part F0 parameter reading part 85, the F0 parameter of the fixed form part is read from the fixed form part F0 parameter storage part 87, and the fixed form part F0 is read.
The pattern generation unit 86 generates a time series of basic frequencies (F0 pattern) from the parameters. Similarly, regarding the non-standard part, the non-standard part F0 parameter reading unit 9
5, the accent of the word of the atypical part is acquired from the accent dictionary 93, and the appropriate F0 is obtained from the atypical part F0 parameter accumulating part 97 according to the number of syllables and the accent type.
The parameters are read and the non-standard part F0 pattern generation part 9
6, the time series (F0
Pattern).

A block diagram of the third embodiment of the present invention is shown in FIG. This embodiment corresponds to claim 6 of the patent. In this embodiment, the atypical portion F0 pattern reading unit 92 of the first embodiment is used.
The non-standard part F0 pattern accumulator 94 is replaced with the non-standard part F0 pattern generator 98. Since the processing of the other parts is the same as that of the first embodiment, only the non-standard part F0 pattern generation part 98 will be described.

The non-standard part F0 pattern generator 98 acquires the accent of the non-standard part word from the accent dictionary 93, and generates the F0 pattern according to the rule in consideration of the position in the sentence. As a method of generating the F0 pattern by a rule, a method using a model such as a Fujisaki model or a point pitch model is generally used, and these methods can also be applied in this case.

A block diagram of the fourth embodiment of the present invention is shown in FIG. This embodiment corresponds to claims 10 and 11. In the present embodiment, the text analysis is made more accurate by replacing the user interface of the text input unit of the first embodiment. Input interface section 01
In 2, the standard part is read from the standard part sentence example accumulating part 013 and displayed as a user interface as shown in FIG. 7 or 8. In FIG. 7, a column having only a display-only function is provided in the fixed part, and a column having an edit function in which a word can be freely input / edited is prepared in the fixed part so that the user can input the fixed part. Urge. If you input with such an interface, it is unnecessary to judge the fixed part and the atypical part,
A text analysis can be performed by searching only the fixed part in the word dictionary 74.

In FIG. 8, input candidates for the non-standard parts are stored in the standard sentence storage unit 13, and when the column of the non-standard parts is designated, the input candidates that should be included in the column are displayed, and the candidate selection means is used. , It has an interface that allows you to specify which is the input. Similarly, it is unnecessary to determine the fixed part and the unfixed part, and the text analysis can be performed by searching only the fixed part in the word dictionary 74. Subsequent processing is the same as in the other embodiments.

[0026]

As described above, according to the present invention, in a voice synthesizer for synthesizing fixed-form sentence voices, which is used for a voice service of traffic information or weather conditions, it is easy to hear and has a natural prosody. Speech can be synthesized.

[Brief description of drawings]

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

FIG. 2 is a conceptual diagram showing the basic idea of the present invention.

FIG. 3 is a first embodiment of the present invention.

FIG. 4 is a second embodiment of the present invention.

FIG. 5 is a third embodiment of the present invention.

FIG. 6 is a fourth embodiment of the present invention.

FIG. 7 is a first example of a user interface of the present invention.

FIG. 8 is a second example of the user interface of the present invention.

FIG. 9 is a conventional example.

[Explanation of symbols]

1 Text Input Means 2, 7 Text Analyzing Means 3 Fixed Form Synthesizing Means 4 Atypical Form Composing Means 5 Output Voice Connecting Means 6 Audio Output Means 8 Fixed Form F0 Pattern / Duration Length Generation Means 9 Atypical Form F0 Patterns / Duration Long generation means 10 F0 pattern / duration duration connection editing means (abbreviated as editing means) 11 Acoustic parameter generation means 12 Speech signal generation means 61 Speech output section 71, 71 'Text analysis section 72 Fixed / atypical decision section 73 Output Switching unit 74 Word dictionary 75, 013 Fixed part sentence example accumulation unit 81 Fixed part duration reading unit 82 Fixed part F0 pattern reading unit 83 Fixed part duration length accumulation unit 84 Fixed part F0 pattern accumulation unit 85 Fixed part F0 parameter reading unit 86 standard part F0 pattern generation part 87 standard part F0 parameter storage part 91 non-standard part persistence Length generator 92 Atypical part F0 pattern reading part 93 Accent dictionary 94 Atypical part F0 pattern accumulating part 95 Atypical part F0 parameter reading part 96, 98 Atypical part F0 pattern generating part 97 Atypical part F0 parameter accumulating part 011 Text input section 012 Input interface section 101 Duration connection editing section 102 F0 pattern connection editing section 111 Acoustic parameter generating section 112 Acoustic parameter accumulating section 121 Voice signal generating section

Claims (11)

[Claims] 1. A voice synthesizing apparatus for synthesizing a group of messages by connecting fixed information common to a group of messages to be synthesized and variable information different for each group of messages, to generate a time-varying pattern of a fundamental frequency. , First generating means for generating a temporal change pattern of the fundamental frequency of the fixed information, second generating means for generating a temporal change pattern of the fundamental frequency of the variable information, and time of the fundamental frequency generated by each of the generating means. And an edit unit for sequentially connecting the change patterns to generate a time change pattern of the fundamental frequency of the sentence, and a synthesizing unit for synthesizing a voice signal by using the time change pattern of the fundamental frequency generated by the edit unit. Characteristic speech synthesizer. 2. The first generating means according to claim 1, wherein the time-varying pattern of the fundamental frequency of the fixed information extracted from the natural voice is stored by using a time-series format of the fundamental frequency. A speech synthesizer characterized by generating a time-varying pattern of a fundamental frequency by including a time series of a fundamental frequency suitable for a sentence from the storage means. 3. The first generation means according to claim 1, wherein the time variation pattern of the fundamental frequency of the fixed information extracted from the natural voice is in the form of a model parameter that approximates the time variation pattern of the fundamental frequency. Means for storing and using,
Speech synthesis characterized by generating a time-varying pattern of the fundamental frequency by providing means for reading an appropriate parameter from the means for storing the input sentence and means for generating a time series of the fundamental frequency from the parameter apparatus. 4. The second generation means according to claim 1, wherein the temporal change pattern of the fundamental frequency extracted from the natural voice with respect to the combination of the syllable number of variable information and the accent type is converted into a time series format of the fundamental frequency. A speech synthesizer characterized by generating a temporal change pattern of a fundamental frequency by including a means for storing by using it and a means for selecting and reading a time series of a fundamental frequency suitable for an input sentence from the storage means. 5. The second generation means according to claim 1, wherein the time change pattern of the fundamental frequency is the time change pattern of the fundamental frequency extracted from the natural voice for all combinations of the syllable number of the variable information and the accent type. By including means for storing the pattern using the model parameter format that approximates the pattern, means for selecting and reading an appropriate parameter from the storage means, and means for generating a time series of the fundamental frequency from the parameter, A speech synthesizer characterized by generating a temporal change pattern of a fundamental frequency. 6. The speech synthesis apparatus according to claim 1, wherein the second generation means has means for generating a time change pattern of the fundamental frequency of the variable information according to a rule. 7. When generating a duration length that is a sequence of time lengths of a synthesis unit, a first generation unit that generates a duration length of fixed information and a second generation unit that generates a duration length of variable information. The generation means and the duration length generated by each generation means are sequentially connected,
A voice synthesizing apparatus comprising: an editing unit for generating a sentence duration and a unit for synthesizing a voice signal using the duration. 8. The first generating means according to claim 7, wherein the duration of fixed information extracted from natural voice is stored, and the duration suitable for an input sentence is read from the storage. A speech synthesizer characterized by generating a duration length by comprising: 9. The speech synthesis apparatus according to claim 7, further comprising a generation unit that generates a duration of variable information. 10. The voice synthesizer according to claim 1, wherein the voice synthesizer presents fixed information, and a user inputs a synthetic sentence using a user interface for inputting and editing variable information. By doing so, a voice synthesizing device comprising a text input means that enables separation of fixed information and variable information. 11. The speech synthesizing apparatus according to claim 1, wherein the speech synthesizing apparatus presents fixed information and input candidates of variable information, and selects variable information of the candidate. And a text input means for separating fixed information and variable information from each other.