JP2001343987A - Method and device for voice synthesis - Google Patents

Abstract

PROBLEM TO BE SOLVED: To resolve the problem that it is impossible to perform switching between sound prolongation/non-prolongation in accordance with the case that the same word should be spoken by sound prolongation as normal speaking or should be clearly articulated when using conventionally known sound prolongation processing in text voice conversion which generates a synthesized voice from a text. SOLUTION: A sound prolongation setting means is provided, and information of sound prolongation candidate positions where sounds may be prolonged and language information are used to switch sound prolongation so that sounds may not be prolonged at the time of clear articulation and may be prolonged at the time of normal speaking.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech synthesizing method for generating a synthesized speech for input text information by connecting speech units, and an apparatus therefor. More specifically, the present invention relates to a setting as to whether or not a long speech unit is to be used for a specific voice position that may be prolonged.

[0002]

2. Description of the Related Art In recent years, a speech synthesizing method or apparatus has been developed in which, when text of a sentence mixed with Kanji or Kana is input, a synthesized speech for the sentence or an apparatus therefor has been developed, and attempts to improve the quality of the synthesized speech have been continued. ing.

One of the important factors influencing the quality of such synthesized speech is how to process long sounds.

For example, conventionally, in a speech synthesis method for connecting speech units corresponding to each text on the basis of text information of hiragana or katakana given as reading information of a sentence, it is determined whether or not long sounds can be omitted. Proposals have been made to improve the quality of synthesized speech based on information. Specifically, with respect to the utterance of the word "computer", information is added that the long sound "-" at the end of the word can be omitted. On the other hand, the long sound "-" located ahead of the end is provided with information that it cannot be omitted. As a result, the utterance of "computer" or "computer" by the synthesized speech is allowed, but the utterance of "computer" or "computer" is prohibited (Japanese Patent Laid-Open No. 3-156666).

Due to the characteristics of long sounds in Japanese pronunciation, unlike a conventional device in which a long sound is added or omitted, at the position of a specific sound, a long sound voice is used instead of a speech unit for the reading information. It may be preferable to employ a segment.

For example, the reading of the word "experience" is "Kakeken" in katakana notation. In practice, however, the pronunciation of "keken" in katakana notation has better voice quality and sounds natural to the listener. is there. This means that the "i" immediately after the "d"("e","ke","se", ...)
Depends on the pronunciation rule that can be changed to a long sound and pronounced.

[0007] In addition, the step sounds ("o", "ko", "so",
…)), The pronunciation rule that “U” can be changed to a long sound, for example, it is more natural for the listener to pronounce the word “Report” given “Hokukoku” as a reading as “Hokukoku”. It sounds like. These pronunciation rules are detailed in NHK's “pronunciation accent dictionary”.

Therefore, in the conventional speech synthesis method,
Normally, based on pronunciation rules, a long voice is assigned to a specific reading and speech synthesis is performed.

On the other hand, depending on the use of the speech synthesizer, for example, the word "experience" is used as a learning material for reading kanji, so that the word "experience" is "K", "I", In some cases, it may be necessary to perform speech synthesis so that the user can hear "" and "".

[0010] However, in the conventional speech synthesizer, whether or not to lengthen specific reading information permitted to be changed to a long sound is determined by analyzing input reading information (for example, text information). In the process of doing so, it was decided uniquely. For example, by using a language dictionary having a conversion table function in which all the specific reading information permitted to be changed to the long sound is changed to the long sound information, all of the specific reading sound permitted to be changed to the long sound can be used. It will be lengthened.

As described above, in the conventional speech synthesizer, the user cannot freely set whether or not to change specific reading information into long sound information (hereinafter, referred to as long sound).

On the other hand, even in the case of speech (units of phonemes, words, phrases, etc.) that can be lengthened in the same way, depending on where the speech exists in the text, or its linguistic information (whether it is a part of speech or a speech) Depending on the environment, such as whether or not the word is an autonomous word), the process of lengthening the sound may or may not be preferable. In other words, the process of allowing long sounds may make it easier to hear or may make it harder to hear.

As described above, in the conventional processing for increasing the length of a tone, even if the same word is desired to be lengthened as in the case of ordinary utterance, there is a difference between the case where the same word is desired to be made longer and the pronunciation of each sound of the reading. Even if there is a case where it is desired to do so, it has not been possible to arbitrarily switch between prolonged and non-prolonged.

[0014]

A conventional speech synthesis method,
Alternatively, in the device, in order to solve the above-mentioned inconvenience, it is necessary to register both the reading corresponding to the prolonged sound and the reading corresponding to the non-prolonged sound in the language dictionary, thereby designating the prolonged / non-prolonged sound. It is also conceivable to make it possible to do so. However, in this method, unless the designation is changed each time, there is an inconvenience that the sound is fixed to either the long sound or the non-long sound, and the setting is not freely performed.
Further, in this case, there is a problem that the language dictionary becomes enormous as the number of registered words increases.

Further, in a speech synthesis system mounted on a small device or the like, the input text may not be a mixture of kanji and katakana text (a katakana code string). In this case, since a language dictionary for registering headings, readings, parts of speech, and the like is often omitted, it has been difficult to switch between long sound and non-long sound.

Therefore, the present invention is applicable to both cases where it is desired to lengthen a word such as a normal pronunciation for a word or the like having a candidate for lengthening, and for a case where it is desired to pronounce each sound clearly as read. It is an object of the present invention to provide a speech synthesis method and a device therefor.

[0017]

A voice synthesizing method according to the present invention performs a lengthening process for changing reading information of a voice different from a long sound into long sound information corresponding to a long sound based on the reading information of the voice. A possible lengthening candidate position is set in advance, and it is possible to set whether or not to perform the lengthening process on the lengthening candidate position. As a result, it is possible to selectively lengthen the sound of the long-sounding candidate, so that it is possible to realize an easy-to-listen speech synthesis method according to the use and purpose of the synthesized speech.

The speech synthesis method according to the present invention further comprises an analysis process for generating at least reading information by analyzing input text information; and a reading process for a voice different from a long sound based on the generated reading information. A long sound candidate detection process for detecting a long sound candidate position capable of performing a long sound process for changing information to long sound information corresponding to a long sound;
A lengthening process for changing predetermined voice reading information to the long sound information based on a rule that sets whether or not the detected reading information of each detected long sound candidate position is changed to the long sound information; And a segment editing process of editing a speech segment based on the read information thus generated to generate a synthesized speech. In this way, by setting rules for lengthening the length of sound, it is possible to generate synthesized speech that is easy to hear.

The rule in this case is based on the position information of the prolongation candidate in the sentence. As the position information, the position of the sentence including the prolongation candidate, the position of the phrase including the prolongation candidate, the prolongation candidate , The position of the mora containing the prolongation candidate, the position of the syllable containing the prolongation candidate, the position of the section between poses containing the prolongation candidate, and the accent phrase containing the prolongation candidate , Or the position of the exhalation paragraph containing the prolongation candidate. It is possible to set rules for prolongation in the Japanese language depending on the position of the prolongation candidate, and to generate natural synthesized speech.

Further, the rule in this case is based on linguistic information of the prolonged candidate in the sentence, ie, part-of-speech information, body language, body language stop, independent word, attached word, and the like. It is possible to set rules for prolongation in the Japanese language that depend on the linguistic information of the prolongation candidate, and generate a natural synthesized speech.

Further, the speech synthesizing apparatus of the present invention analyzes the input text information to generate at least reading information, and, on the basis of the generated reading information, performs processing on a speech different from a long sound. A long sound candidate detection processing unit that detects a long sound candidate position that can change the reading information to long sound information corresponding to a long sound, and a rule that sets whether or not to change the read information of each detected long sound candidate position to a long sound A long tone setting unit comprising: a long tone processing unit that changes reading information at a position defined by a rule provided in the long tone setting unit to the long tone information; and And a speech synthesizer for editing a speech unit to generate a synthesized speech. Therefore, rules for lengthening the length of sound can be set, and a speech synthesizer that generates synthesized speech that is easy to hear can be realized.

[0022]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG.
Explanation will be made based on 6.

FIG. 1 is a block diagram showing the configuration of the speech synthesizer of the present invention. In this device, while a sentence mixed with kanji or kana is input as text information 1a, a synthesized speech 7a is generated by a unit synthesis method.
It is possible to perform a process of switching between prolonged sound and non-prolonged sound at a specific portion of the synthesized speech 7a.

First, with respect to the input text information 1a (a sentence mixed with kanji and kana), "reading" and "part of speech" are determined for each word, and "accent" and "pause" are determined. For this reason, in the analysis unit 1, although not shown, for example, “reading”, “part of speech”,
A language dictionary storing information on "accents" and "pauses" is provided, and these information are assigned by the collation processing on the text information 1a. Here, the detection up to the syllable level is performed for the reading information.

The prolonged sound candidate detection unit 2 includes a text analysis unit 1
This is to detect a portion that may be prolonged in the result of the analysis. As a detection method here, for example, a long sound candidate list is used as shown in FIG. This list accumulates combinations (two syllables or three syllables) of readings that are recognized as prolonged in the Japanese pronunciation rules and readings immediately preceding the reading conditions. "I" or "u" immediately after the step sound is included as a candidate for longer sound.

When such a long tone candidate list is used, the reading information obtained from the text analysis unit 1, that is, the reading character string (for example, the character string of katakana text) is added to the character string of the second syllable and thereafter. It is possible to search for a match with a character string in the list, and detect the matched portion as a location where there is a possibility of prolongation, that is, a position of a prolongation candidate.

In the prolonged candidate detection section 2, for each word of the language dictionary used in the text analysis section 1, information as to whether or not a specific portion of the word may be prolonged. Can be added, and when the input text information is analyzed, it is possible to detect information indicating the possibility of prolonging the tone added to the word matched by the language dictionary.

In the long sound setting section 3, rules for long sound / non-long sound for a specific voice portion are set in advance. The rule of lengthening / de-prolonging is made up of a plurality of conditions as shown in FIG. In this case, rule 1 states that “the syllable of the prolongation candidate is other than the first syllable in the phrase,
If the word is a proper noun, all words are not prolonged. " Rule 2 is that "if the syllable of the prolongation candidate is the last syllable in the phrase and the syllable position is the last syllable in the sentence, all the pronouns will not be prolonged." Rule 3 states that, in the case of a non-first syllable other than the first syllable in a phrase, all syllables are not lengthened. In this case, according to the above rules 1 to 3, it can be considered that the rule for performing the prolongation is set for the syllable of the prolongation candidate which does not correspond to these rules.

The prolongation processing unit 4 determines the prolongation / non-prolongation of a portion that may be prolonged, based on the rules set in the prolongation setting unit 3. That is, the reading information of the specific candidate is changed to the reading information “−” of the long sound in accordance with the rules.

The phoneme / prosodic determination unit 5 converts the phonetic symbol string into a phoneme symbol string using the read information subjected to the lengthening determined by the lengthening processing unit 4 and the text analysis unit 1 determines the phoneme symbol string. Prosody information (eg, fundamental frequency, power, phoneme duration) based on information such as accents and poses is determined. The result obtained by this phoneme / prosodic determination unit 5 is:
For example, as shown in FIG. 4, prosodic information is given to each phoneme. In FIG. 4, a fundamental frequency corresponding to each phoneme symbol indicates information indicating a pitch of a sound, a phoneme duration indicates information indicating a length of a sound, and a power coefficient indicates a weight coefficient of power applied to a speech unit. If the power coefficient is 1.0,
This means that the power of the speech unit selected from the speech database is used as it is.

Therefore, for a syllable that has been lengthened, the phoneme of the immediately preceding vowel is extended. For example,
In the case of “e”, the phoneme duration time 102 msec is extended, and the variation of the fundamental frequency and the power can be controlled together with the extension time. Note that these pieces of information can be uniquely set by the phoneme / prosodic determination unit 5, but a setting reflecting each word or the position of a word in a sentence is made in a language dictionary provided in the text analysis unit 1. It is also possible.

The speech synthesis unit 7 selects appropriate speech units from the speech database 6 so as to match the phoneme symbol string obtained from the phoneme / prosodic determination unit 5 and connects them to form the text 1a. Correspondingly, a synthesized speech 7a that has undergone a predetermined lengthening process is generated.

In the above, the case where the setting of the prolonged / non-prolonged sound is set based on the rule regarding the prolonged sound candidate position in the phrase (rules 1, 2, and 3) has been described as an example. Is not limited to this. That is, in the present invention, the position of a sentence that includes a phoneme that is a candidate for prolongation and a word that includes the same, the position of a sentence that includes the candidate for prolongation, the position of a phrase that includes the candidate for prolongation, , The position of the mora containing the prolongation candidate, the position of the syllable containing the prolongation candidate, the position of the section between poses containing the prolongation candidate, and the accent phrase containing the prolongation candidate , Or the setting of the prolonged / non-prolonged sound based on the position of the exhalation paragraph including the prolonged sound candidate. In order to consider the positions of prolonged candidates after subdividing the unit of speech in this way, a language dictionary with information on phonemes, mora, pause sections, accent phrases, intake paragraphs, etc. By providing, the respective positional relationships can be analyzed.

Next, in the present invention, an example of the operation in the case where the setting of the non-prolonged sound is performed for a specific portion will be described with reference to the flowchart of FIG.

First, when it is detected in step 10 that a text has been input, the text analyzer 1 proceeds to text analysis in step 11. For example, if the input text is “Route 1 near Kyobashi, traffic congestion.” In step 11, kanji reading, accent, pose information, and the like are determined. When a kanji-kana sentence is input, a language dictionary describing linguistic information such as headings, headlines, parts of speech, readings, and phoneme symbol strings is referred to in the text analysis step.

In step 12, the output of the analysis in step 11 is referred to by the prolonged sound candidate list (FIG. 2) by the prolonged sound candidate detecting section 2, so that there is a portion that may be prolonged. Search for or not. Here, if there is a prolongation candidate, the process proceeds to step 13, and referring to the prolongation / non-prolongation rule (FIG. 3) of the prolongation setting section 3.
It is determined whether or not the setting of the lengthening / deprolonging is performed. If the setting is performed, the process proceeds to step S15. If it is determined in step S13 that the lengthening process has not been performed, the process proceeds to step S14. In the case of FIG. 3, a rule that does not correspond to any of the rules 1, 2, and 3 is subjected to normal lengthening.

In step 15, the extracted phoneme symbol at the prolonged candidate position is compared with the setting conditions of prolonged / non-prolonged, and the phoneme symbol at the prolonged candidate position that matches the set condition is compared. Extract whether or not exists. If this exists, in step 16, the corresponding portion is subjected to the process of lengthening or non-prolonging the length in the lengthening processing unit 4 in accordance with the set contents, and the process proceeds to step 17. If not, the process proceeds directly to step S17.

Steps 17 to 20 are the same as the speech synthesis processing of the ordinary unit editing method. First,
In step 17, the phoneme / prosodic determination unit 5 converts the input text including the prolonged portion or the non-prolonged portion processed in the above steps into phoneme information by conversion into a phoneme time series, Alternatively, prosody information for each short time (mainly, fundamental frequency, power, phoneme duration) is determined. In step 18, an optimal speech unit is selected from the speech database 6 based on the phoneme information and the prosody information. In step 19, connection of each speech unit and prosody control based on the prosody information are performed. In step 20, a synthesized speech is generated.

This makes it possible to set the lengthening / deprolongation not only of the entire text but also of a specific portion, and it is possible to flexibly respond to the case where it is desired to pronounce each sound clearly by a word or the like. Therefore, a desired synthesized speech can be generated.

Next, as an example to which the present invention is applied, the content of the input text is "Route 1 near Kyobashi Station, congestion." The rules of the long sound setting section 3 are as shown in FIG. In the following, a description will be given of a case where “only the place name is not lengthened”.

First, the text analysis unit 1 determines information such as kanji reading, accent, and pause. When the accent information is indicated by an upward arrow “↑” and the pose information is indicated by a triangle mark “△”, in the case of the above text, “Kokudo Ichigo Sensou Kyoba Ushibashi ↓ Is generated as the main information.

At this time, “Kyobashi” is a proper noun that is a place name, “National highway”, “Line 1”, “nearby” and “traffic jam” are ordinary nouns.
Part-of-speech information such as "is" an auxiliary verb is also determined.

The prolongation candidate detection unit 2 refers to the prolongation candidate list in FIG. 2 or refers to information indicating the possibility of prolongation registered in the language dictionary, and determines whether it is “Kyobashi”, “Kokudo”. The three words "" and "Ichigo Usen" are extracted as long target words including phonemes of long target candidates that may be prolonged.

The lengthening processing section 4 analyzes the content of the lengthening / non-prolonging and the content of the word to be lengthened. In this case, it is determined that only the place name "Kyoubashi" matches the setting, "Kyoubashi" is not lengthened, and "Kokudo" and "Ichigo Usen" are lengthened and converted to "Kokudo" and "Ichigo Sen", respectively.

Therefore, the intermediate information generated at this stage is "commercial information". Based on the intermediate information and the part-of-speech information, phoneme information and prosody information (mainly, fundamental frequency, power, and phoneme duration) are determined in the phoneme / prosodic determination unit 5.

Thereafter, based on the phoneme information and the prosody information obtained from the phoneme / prosodic determination unit 5, the speech synthesis unit 7 selects the optimum speech unit from the speech database and connects the speech units. Prosody control processing and the like are performed, and a desired synthesized speech is generated.

In the present embodiment, the process of lengthening / de-prolonging the sound is performed before the conversion into the phoneme information. However, the process can be performed after the conversion into the phoneme information.

In the above, the rule based on the part of speech information that the linguistic information on the phrase including the prolonged candidate in the sentence corresponding to the reading information is "place name" has been exemplified. Rules based on whether or not it is a noun, a noun, an independent word, or an adjunct (stored as dictionary information in a language dictionary) can be used.

For example, in the case of speech stop, in order to emphasize the word of the speech, even if the word includes a prolongation candidate, the prolongation of the prolongation is prohibited, and the sound is clearly pronounced one by one. Can be pronounced. Depending on the application of the speech synthesizer of the present invention (such as a language learning teaching material), it is possible to make the speech of a word whose pronunciation is to be clearly uttered non-prolonged, to make the speech of the word whose emphasis is to be uttered non-prolonged, or It is possible to set rules for prolonging the speech of words that you want to pronounce clearly in this case. In this case, the part-of-speech information of the word, whether the word is a noun, or whether the word is an independent word It is possible to make a rule based on whether or not it is an adjunct.

Although the prolonged candidate list in FIG. 2 shows candidates for Japanese syllables, there is a possibility that more candidates may be added to correspond to foreign words.

It is generally known from linguistics that a phoneme or a syllable that may be prolonged is the second syllable or later in a syllable. In the rule that can be set, the setting of the partial rule of “other than the first syllable in a phrase” may be omitted.

[0052]

As is apparent from the above description, according to the present invention, it is possible to flexibly lengthen / de-prolong the sound in the speech synthesis processing. It is possible to cope with a case where it is desired to make the pronunciation of each word clearly pronounced one by one, and it is possible to produce an effect that a desired synthesized speech can be generated.

[Brief description of the drawings]

FIG. 1 is a block diagram of an embodiment.

FIG. 2 is a schematic diagram showing the contents of a long tone candidate list according to the embodiment;

FIG. 3 is a schematic diagram showing the content of a rule for lengthening / de-prolonging a tone in an embodiment.

FIG. 4 is a schematic diagram showing output contents of a phoneme / prosodic determination unit according to the embodiment;

FIG. 5 is a diagram showing a flowchart in the embodiment.

FIG. 6 is a diagram showing a rule of lengthening / de-prolonging in another embodiment.

[Explanation of symbols]

 1a Text 1 Text analysis unit 2 Prolonged sound part detection unit 3 Prolongation setting unit 4 Prolongation processing unit 5 Phoneme / prosodic determination unit 6 Voice database 7 Voice synthesis unit 7a Synthesized voice

Claims (7)

[Claims] 1. A long tone candidate position capable of performing a long tone process for changing read information of a voice different from a long tone into long tone information corresponding to a long tone based on the read information of a voice is set in advance. And a speech synthesis method that enables setting of whether or not to execute the lengthening process for the lengthening candidate position. 2. An analysis process for generating at least reading information by analyzing input text information, and converting reading information about a voice different from a long sound into long sound information corresponding to the long sound based on the generated reading information. A long sound candidate detection process for detecting a long sound candidate position capable of performing a long sound process to be changed to, and setting whether or not to change the read information of each detected long sound candidate position to the long sound information. Based on the rules, a lengthening process of changing predetermined voice reading information to the long sound information, a segment editing process of editing a speech unit based on the lengthened reading information to generate a synthesized voice, A speech synthesis method consisting of 3. The rule set in the lengthening process is as follows:
3. The speech synthesis method according to claim 2, wherein the speech synthesis method is based on position information on the prolonged sound candidate position in speech reading information. 4. The position information on the prolonged candidate position includes a sentence position including the prolonged candidate, a sentence position including the prolonged candidate, a phrase position including the prolonged candidate, and a prolonged candidate. The position of the included word, the position of the mora containing the prolonged candidate, the position of the syllable containing the prolonged candidate, the position of the section between the poses containing the prolonged candidate, and the position of the accent phrase containing the prolonged candidate 4. The speech synthesis method according to claim 3, wherein the position is a position or a position of an exhalation paragraph including a prolongation candidate. 5. The rule set in the lengthening process is as follows:
3. The method according to claim 2, wherein the utterance information is based on linguistic information about a phrase including a prolonged candidate in a sentence corresponding to the reading information.
Described speech synthesis method. 6. The speech synthesis method according to claim 5, wherein the linguistic information is part-of-speech information, a noun, a noun, an independent word, or an adjunct word. 7. An analysis processing unit that generates at least reading information by analyzing input text information;
Based on the generated reading information, a long sound candidate detection processing unit that detects a long sound candidate position that can change the reading information about the voice different from the long sound to the long sound information corresponding to the long sound,
A prolonged setting section having a rule setting whether or not to change the detected reading information of each detected prolonged candidate position to a prolonged sound; and a reading information of a position determined by a rule provided in the prolonged setting section. A speech synthesis device comprising: a long-sound processing unit that changes the word into the long-sound information; and a speech synthesis unit that edits a speech unit to generate a synthesized speech based on the long-sound reading information.