JPH07160290A - Sound synthesizing system - Google Patents

Abstract

PURPOSE:To obtain a synthesized sound excellent in clearness and naturalness and to reduce the synthesis processing time by automatically adding stored sound waveform and attribute information corresponding to the stored sound waveform to an information storage section or updating them. CONSTITUTION:An attribute information storage section 3 and a waveform information storage section 5 control a sound waveform using phoneme information and accent information as index, it is judged whether or not the stored sound waveform and its attribute information are to be added during sound synthesis and if judged to be added, add these. Moreover, from a connected word information storage section 8 in which connected words are stored taking correspondence with their phoneme information, connection word having no correspondence with the stored sound waveform is retrieved and the sound waveform having correspondence with the connected word is generated using the stored sound waveform and the attribute information, or the sound waveform inputted from an external and its attribute information are stored as a stored sound waveform and its attribute information. Thus, a stored sound waveform and its attribute information are automatically added and updated.

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, and more particularly to a system for synthesizing voice by editing, combining and transforming voice units.

[0002]

2. Description of the Related Art In general, a speech synthesis method is a method in which phonological information and prosody information are input and a synthetic speech waveform corresponding to them is output. Here, the phonological information is a phonetic transcription (for example, pronunciation of “Keiken” is generated from a phonetic transcription of a written character (for example, “experience”) such as a word or a text (for example, “Keiken” is a pronunciation of “experience”) “Kaken”, notation is defined as information of “Keken”). The prosody information is defined as accent phrases / breaks of breath paragraphs, accent information of word / clause combination results, and intonation information. This prosody information is generated from accents, morphemes, syntactic analysis results, etc. for the above notation characters. Such speech synthesis methods can be roughly classified into a waveform editing method and a rule synthesizing method. The conventional method will be described below with reference to the drawings.

First, a conventional waveform editing method will be described. The waveform editing method uses a sentence,
This is a method of synthesizing a speech waveform by accumulating words or phrases in units and reading out and splicing the corresponding speech based on the input phonological information or written character information. The accumulated speech waveform is managed by using phonological information or written character information of a synthesis unit such as a sentence, a word, or a phrase as an index.

FIG. 5 is a basic block diagram of the waveform editing system. In FIG. 5, 51 is an input terminal, 52 is a synthesis unit decomposition unit, 53 is a voice waveform editing unit, 54 is a voice waveform storage unit, and 55 is an output terminal. The input terminal 51 outputs the input phoneme information to the synthesis unit decomposition unit 52. The synthesis unit decomposition unit 52 decomposes the phoneme information from the input terminal 51 into synthesis units, and outputs it to the speech waveform editing unit 53. The voice waveform storage unit 54 stores a voice waveform obtained by cutting out a voice uttered by a person in advance in a synthesis unit. The voice waveform editing unit 53 outputs the synthesis unit to the voice waveform storage unit 54. The voice waveform storage unit 54 outputs the voice waveform corresponding to each synthesis unit to the voice waveform editing unit 53. The voice waveform editing unit 53 stitches the voice waveforms from the voice waveform accumulating unit 54, and outputs the resultant waveform as a synthesized voice waveform via the output terminal 54. The voice waveform storage unit 54
As a means for accumulating the voice waveform in the above, there is a means for accumulating the characteristic parameters of the waveform in addition to the means for accumulating the cut-out waveform itself. In the latter case, the voice waveform editing unit 53 synthesizes the voice waveforms using the characteristic parameters and then joins the voice waveforms together. In addition, the voice waveform editing unit 53 may shape the voice waveform of the joining portion in order to improve the quality of synthesized voice.

This conventional method has an advantage that the synthesized speech quality is good because the accumulated speech waveforms are directly spliced and output. However, there is a drawback in that waveforms other than combinations of voice waveforms stored in the voice waveform storage unit 54 cannot be synthesized. Further, in order to synthesize an arbitrary voice waveform, it is necessary to store a huge amount of voice waveforms in the voice waveform storage section 54, and it is practically impossible to synthesize any voice waveform.

Next, a conventional rule combining method will be described. The rule synthesizing method stores characteristic parameters of a phoneme or syllable-like speech waveform and prosody information for connecting them in advance, and uses the rules based on the inputted phonological information and prosody information. This is a method of synthesizing a speech waveform from characteristic parameters. The accumulated characteristic parameters are managed using as an index the phoneme information of a synthesis unit such as a phoneme or a syllable.

FIG. 6 is a basic block diagram of the rule composition method. In FIG. 6, 61 and 62 are input terminals, and 6
3 is a synthesis unit decomposition unit, 64 is a characteristic parameter storage unit, 6
Reference numeral 5 is a voice waveform synthesis unit, 66 is a prosody rule storage unit, and 67 is an output terminal.

The input terminal 61 outputs the input phoneme information to the synthesis unit decomposition unit 63. Characteristic parameter storage unit 6
4 is a syllable synthesis unit (for example, CV, VC, VC
V, or CVC; C represents a consonant, and V represents a vowel). The synthesis unit decomposition unit 63 decomposes the phoneme information from the input terminal 61 into synthesis units, and outputs it to the characteristic parameter storage unit 64.
The characteristic parameter storage unit 64 outputs the characteristic parameter corresponding to each synthesis unit to the synthesis unit decomposition unit 63. The synthesis unit decomposition 63 outputs the feature parameter from the feature parameter storage unit 64 to the speech waveform synthesis unit 65.
The prosody rule storage unit 66 stores prosody control rules in which prosody information and prosody parameters are associated with each other. Input terminal 62
Outputs the input prosody information to the speech waveform synthesizer 65. The voice waveform synthesizer 65 outputs the prosody information from the input terminal 62 to the prosody rule accumulator 66. Prosody rule storage unit 6
6 outputs a prosody control parameter corresponding to the prosody information to the speech waveform synthesizer 65. The voice waveform synthesizer 65
A synthetic voice is generated using the characteristic parameter from the synthesis unit decomposition unit 63 and the prosody control parameter from the prosody rule storage unit 66, and is output via the output terminal 67.

This conventional method has an advantage that any speech waveform can be synthesized because all the characteristic parameters and the prosody control parameters are stored with a phoneme or a syllable degree as a synthesis unit. That is, in principle, it is not necessary to update the contents of the characteristic parameter storage unit 64 once constructed. However, it has the drawbacks that the quality of synthesized speech is deteriorated and the correction processing time is increased because the number of joints in the waveform is so large that the correction of the discontinuity is incomplete.

As a conventional method other than the above two methods, a voice synthesis method using a composite voice unit has been proposed. This is to store compound speech units having various lengths and structures from phoneme level to syllables, morphemes, words, phrases and sentences, and synthesize speech waveforms using compound speech units and sub-units included in them. The method is characterized by The description of this system does not particularly mention the method of accumulating the speech waveform of the composite synthesis unit and the partial unit thereof. However, in one specific example described in the same description, those speech waveforms are managed using the phoneme information as an index.

FIG. 7 is a schematic block diagram of this conventional system. In FIG. 7, 71 is an input terminal, 72 is a voice unit selection unit, 73 is a voice unit file, 74 is a voice control knowledge file, 75 is a unit joining unit, 76 is a synthetic waveform generation unit,
Reference numeral 77 is an output terminal. The input terminal 71 outputs the input phoneme information and prosody information to the voice unit selection unit 72. The voice unit file 73 stores various voice unit strings collected in advance. In the voice knowledge file 74,
The rules and tables that give concrete synthesis conditions such as extraction environment, fundamental frequency, and phoneme duration are accumulated. The voice unit selection unit 72 selects a voice unit string suitable for synthesis from the voice unit file 73 based on these two pieces of information. The selection criterion of the voice unit is a specific configuration condition such as a desired extraction environment of the synthesis unit, a fundamental frequency, and a phoneme time from the voice control knowledge file 74 based on the phonological environment and the prosody information of the synthesizing portion. The voice unit sequence selected by the voice unit selection unit 72 is given to the unit joining unit 75, and the selected units are joined. The combined voice units are given to the synthetic waveform generation unit, and generate a synthetic waveform based on the synthetic unit series obtained by the unit combination unit 75. Then, the generated sound is output to the output terminal 77. Incidentally, in one specific example of this conventional method, as an index on the voice unit file 73 when the voice unit selection unit 72 selects a voice unit,
Only phonological information is used.

As an advantage of this conventional method, by using not only the accumulated voice waveform itself but also a part thereof, the voice quality is clearer than that of the rule synthesis method, and more arbitrary than the waveform editing method. It is possible to synthesize voice waveforms. However, if the desired voice waveform itself is not stored in the voice unit file 73, the voice waveform must be combined with the voice unit in the voice unit file 73. However, the voice unit file 73
Since many complex speech units having various lengths and structures from a phoneme level to syllables, morphemes, words, phrases, and sentences are stored in the speech unit, the speech unit selection unit 72 selects from the file 73. The number of voice unit strings
It depends on the number of voice unit combinations selected. That is, if the number of combinations becomes enormous, the time required for selecting the voice unit sequence also becomes long, and it becomes difficult to implement the system in a system that requires real-time voice synthesis processing.

[0013]

As described above, the conventional method emphasizes the synthesized voice quality, but has the drawback that it cannot synthesize an arbitrary speech waveform, and it can synthesize any speech waveform, while it does not. It has a drawback that the voice quality is felt to be deteriorated and a relatively large amount of arbitrary high-quality voice waveforms can be synthesized, but it has a drawback that a huge amount of synthesis processing time is required.

In view of the problems of the conventional method and the tendency of the cost of the memory to be low, the present invention assumes that the memory is used to some extent, and the synthesized voice quality has good clarity and naturalness, and waveform editing. It is an object of the present invention to provide a speech synthesis method capable of shortening the synthesis processing time while satisfying that more arbitrary speech waveforms can be synthesized.

[0015]

A feature of the present invention is information storage for storing and managing a stored speech waveform previously cut out in units of words or phrases in association with attribute information consisting of phoneme information of the stored speech waveform. A part of the input phoneme information corresponding to a desired voice, or input information obtained by dividing the input phoneme information into a plurality of pieces and the synthesis unit attribute information which is all or a part of the attribute information, and corresponds to the synthesis unit attribute information. A method of synthesizing the voice by searching the stored voice waveform and joining the stored voice waveforms, wherein the stored voice waveform and the attribute information corresponding to the stored voice waveform are added during the synthesis of the voice. If it is determined to be added, the accumulated voice waveform and the attribute information corresponding to the accumulated voice waveform are added to the information storage unit, and a consecutive word is set as phoneme information of the consecutive word. The continuous word information accumulating unit that is accumulating in response to the accumulated speech waveform is searched for the consecutive words, and the speech waveform corresponding to the uncorrelated consecutive words is converted into the accumulated speech. A waveform generated using the attribute information of the waveform and the accumulated speech waveform, the generated speech waveform corresponding to the compound word and the attribute information thereof are added to the information storage unit, and the speech waveform input from the outside and the speech waveform By adding or updating the attribute information of the stored voice waveform and the stored voice waveform as the attribute information of the stored voice waveform to the stored voice waveform and the attribute information corresponding to the stored voice waveform. This is a speech synthesis method of adding or updating to the information storage unit.

[0016]

Embodiments of the present invention will be described in detail below with reference to the drawings.

FIG. 1 is a schematic block diagram of an embodiment of the speech synthesis system of the present invention. In FIG. 1, 1 is an input terminal,
2 is a synthesis unit selection unit, 3 is an attribute information storage unit, 4 is a voice synthesis unit, 5 is a waveform information storage unit, 6 is an output terminal, 7 is a stored information update unit, 8 is a continuous word information storage unit, and 9 is information. The update determination unit 10 is an input terminal. According to the present invention, the attribute information accumulating unit 3 and the waveform information accumulating unit 5 manage the speech waveform using the phoneme information and the accent information as indexes, and select all or a part of the accumulated waveform in consideration of before and after the connected portion. A characteristic is that the contents of the attribute information accumulating unit 3 and the waveform information accumulating unit 5 are automatically updated in the case of generating synthetic speech by stitching. This update of the stored contents is performed not only during voice synthesis but also during non-voice synthesis. That is, the voice waveform newly synthesized at the time of voice synthesis is registered in the attribute information storage unit 3 and the waveform information storage unit 5 together with its phoneme information and accent information. Further, when a voice waveform, phoneme information, accent information, and cut-out point information (information about a cut-out point of a waveform corresponding to phoneme information / accent information of a voice waveform) are input from the input terminal 10 during non-voice synthesis. Is searched for stored information in the attribute information storage unit 3 and the waveform information storage unit 5, and if the input information is new information, it is registered, and if it is already registered information, it is updated to the one with higher voice waveform quality. On the other hand, when it is between voice synthesis or when no information is input from the input terminal 10, the continuous word information storage unit 8 used at the time of text analysis or the like.
To synthesize and register the speech waveforms of the consecutive words that are not registered in the attribute information storage unit 3 and the waveform information storage unit 5.
Hereinafter, the present invention will be described in detail with respect to these three cases.

First, the update of stored information at the time of voice synthesis will be described.

The input terminal 1 outputs the input phonological information / prosodic information / morpheme information to the synthesis unit selecting section 2. The attribute information storage unit 3 uses the phoneme information (speech notation) / accent information (accent strength information) as an index.
Attribute information related to the accumulated speech waveform extracted in advance in units of words or phrases (that is, a composite waveform such as phonological information of the speech waveform, accent information, and original waveform, that is, which existing accumulated waveform is a synthesized waveform? And the address information of the waveform information of the waveform stored in the voice waveform storage unit 5) are stored and managed. The synthesis unit selection unit 2 first inputs phonological information / prosodic information / morpheme information from the input terminal 1 and extracts accent information from the prosodic information. Next, phonological information / accent information / morpheme information are accent phrases (such as having one main accent,
The phonetic information / accent information of the accent phrase is output to the attribute information storage unit 3. The attribute information storage unit 3 selects attribute information having all or part of the phoneme information / accent information of the input accent phrase. All or part of this accent phrase is defined as a "composite unit". This synthesis unit is the minimum unit for constructing a desired speech waveform. Then, the synthesis unit selection unit 2 sets the attribute information having each synthesis unit as the attribute information of the synthesis unit together with the extraction position information of the synthesis unit.
Output to. The synthesis unit selection unit 2 compares the phoneme information / accent information of the accent phrase with the phoneme information / accent information of the synthesis unit from the attribute information storage unit 3 and extracts the difference together with a part immediately before and after that. To do. Then, the phoneme information / accent information of the extracted portion is output to the attribute information storage unit 3 again. Again, the attribute information storage unit 3 selects the attribute information having all or a part of the input phoneme information / accent information, and outputs it to the synthesis unit selection unit 2 as the attribute information of the synthesis unit. By repeating this, the accent phrase is composed of one or more composition units. The synthesis unit selection unit 2 determines the phoneme information / accent information / morpheme information of the above-mentioned accent phrase, the attribute information and the synthesis order of the synthesis units that form the information, and the information update determination unit 9
Output to. Then, the above-mentioned exchange is repeated between the synthesis unit selecting unit 2 and the attribute information accumulating unit 3 until the synthesis unit is selected for all the phoneme information / accent information obtained from the input terminal 1. When the synthesis unit for all the phoneme information / accent information from the input terminal 1 is selected, the synthesis unit selection unit 2 displays the attribute information of the synthesis unit, the synthesis order of the synthesis unit, and the prosody information from the input terminal 1. Output to the voice waveform synthesizer 4. The waveform information storage unit 5 uses the address information, which is one of the outputs of the synthesis information storage unit 2, as an index, and the corresponding waveform information (that is, the voice waveform itself or its characteristic parameters and the cut-out point information).
Are accumulated and managed. The waveform information in the waveform information storage unit 5 has a one-to-one correspondence with the attribute information in the attribute storage unit 2. The voice waveform synthesizer 4 extracts the address information and cut-out point information of the waveform information from the attribute information from the synthesis unit selector 2 and outputs it to the waveform information accumulator 5. The waveform information storage section 5 searches for waveform information based on the address information from the speech waveform synthesis section 4, and outputs it to the speech waveform synthesis section 4. The voice waveform synthesizer 4 is based on the waveform information from the waveform information accumulator.
Generate a speech waveform for a synthesis unit. And if necessary,
The generated voice waveform is edited while inserting pauses at breaks in the exhalation paragraph, adding intonation of the whole sentence, and interpolating voice waveforms between synthesis units. The generated synthetic speech waveform is output via the output terminal 6.

On the other hand, the information update determination unit 9 uses the phoneme information / accent information / accent information / of the accent phrase from the synthesis unit selection unit 2.
Based on the morpheme information, the attribute information (including the waveform generation information) of the synthesis unit that constitutes the morpheme information, and the synthesis order, it is determined whether or not the contents of the attribute information storage unit 3 and the waveform information storage unit 5 are updated. It should be noted that examples of updating criteria include a case where the accent phrase is composed only of a synthesis unit extracted from the original waveform, and a case where the word excluding the particle and auxiliary verb from the accent phrase is not registered. Here, for convenience, a unit of registration / update composed of an accent phrase or a part of the accent phrase is defined as a “combined word”. In the case of updating, the phoneme information / accent information of the continuous word, the attribute information of the synthesis unit forming the same, and the synthesis order are output to the accumulated information updating unit 7 to request the update. Incidentally, if it is not updated, nothing is done. Similar to the speech waveform synthesizer 4, the accumulated information updater 7 accesses the waveform information accumulator 5 based on various information from the information update determiner 9, synthesizes the speech waveform for the consecutive word, and synthesizes it. The waveform is output to the waveform information storage unit 5, and a request for storage and return of address information is made. The waveform information storage unit 5 stores the combined waveform as a new waveform according to a predetermined format and adds address information. Then, the address information is output to the accumulated information updating unit 7.
Thereafter, the accumulated information updating unit 7 outputs the phoneme information / accent information of the continuous word and the waveform generation information indicating the complex waveform to the attribute information accumulating unit 3 together with the received address information, and accumulates it as the attribute information of the continuous word. Request to do so.

Next, during non-voice synthesis, the input terminal 10
Update of accumulated information when a voice waveform, phoneme information, accent information, and cut-out point information are input will be described. It should be noted that explanations that overlap with the above explanation will be omitted, and only the different operations will be explained in detail.

When the voice waveform, the phoneme information, the accent information, and the cut-out point information are input from the input terminal 10, the accumulated information updating section 7 first outputs the phoneme information / accent information to the attribute information accumulating section 3, Ask if there is attribute information to update. The attribute information storage unit 3 searches for attribute information to be updated, that is, whether there is attribute information that completely matches the phoneme information / accent information from the storage information update unit 7 and the waveform generation information is a composite waveform. To do. When such attribute information exists, the waveform generation information is changed to the original waveform and the address information is output to the accumulated information updating unit 7. The accumulated information updating unit 7 outputs the address information, the voice waveform from the input terminal 10 (or the characteristic parameter obtained by analyzing the waveform) and the cut-out point information to the waveform information accumulating unit 5, and requests the waveform information to be updated. To do. The waveform information storage unit 5 replaces the stored waveform and cut-out point information indicated by the input address information with the new input voice waveform and cut-out point information.

On the other hand, if the attribute information accumulating section 3 completely matches the phonological information / accent information from the accumulated information updating section 7 and there is no attribute information in which the waveform generation information is a composite waveform, that fact is indicated. To the stored information updating section 7. As a result, the accumulated information updating unit 7 outputs the voice waveform (or the characteristic parameter obtained by analyzing the waveform) and the cut-out point information to the waveform information accumulating unit 5 as waveform information, and requests accumulation and return of address information. The waveform information storage unit 5 stores this waveform information as new information,
The address information is output to the accumulated information updating unit 7. The accumulated information updating section 7 outputs the phoneme information / accent information from the input terminal 10 as attribute information to the attribute information accumulating section 3 together with the received address information, and requests registration. The attribute information storage unit 3 registers this attribute information as new information whose waveform generation information is the original waveform.

Lastly, the update of the stored information will be described when it is between voice synthesis or when no information is input from the input terminal 10. It should be noted that explanations that overlap with the above explanation will be omitted, and only the different operations will be explained in detail. In the continuous word information accumulating unit 8, phoneme information, accent information and the like of each continuous word are registered in advance. This is also a dictionary often used for word identification and morphological analysis during text analysis. The accumulated information updating unit 7 accesses the consecutive word information accumulating unit 8 based on an arbitrary rule, and obtains the phoneme information / accent information of a certain consecutive word registered therein. Next, the accumulated information updating unit 7 outputs the phoneme information / accent information from the consecutive word information accumulating unit 8 to the attribute information accumulating unit 3 and asks whether the consecutive word has been registered. The attribute information accumulating unit 3 checks whether the attribute information corresponding to them is registered, and outputs the result to the accumulated information updating unit 7. The above process is repeated until the unstored attribute information is obtained. When a compound word that is not registered in the attribute storage unit 3 is obtained, the stored information update unit 7 accesses the attribute storage unit 3 and, as with the synthesis unit selection unit 2, one or more that compose the compound word. Select and determine the synthesis unit of. Then, similar to the voice waveform synthesizer 4, the accumulated information updater 7 accesses the waveform information accumulator 5 on the basis of the attribute information of the synthesis unit and the synthesis order from the attribute information accumulator 3 and handles the continuous word. Synthesize voice waveforms. After that, the synthesized waveform is output to the waveform information storage unit 5 to request storage and return of address information. Waveform information storage unit 5
Accumulates the synthesized waveform and gives address information.
Then, the address information is output to the accumulated information updating unit 7. The accumulated information updating unit 7 includes, together with the address information from the waveform information accumulating unit 5, phoneme information / accent information of consecutive words,
Also, the waveform generation information indicating that it is a composite waveform is output to the attribute information storage unit 3 and requested to be stored as the attribute information of the continuous word.

Focusing on the difference from the above-described one embodiment, other embodiments of the present invention and matters not defining the present invention will be described below.

In the above embodiment, in order to clarify that the synthesis unit selection unit 2 does not capture the information of the speech waveform itself when selecting the synthesis unit, the phonological information / accent information and the waveform information (speech waveform). It has been described that (including its own information) is separately stored and managed in the attribute storage unit 3 and the waveform information storage unit 5, but the phonological information / accent information and the waveform information are stored and managed in one storage unit. It is also possible to do it. In this case, the storage unit is the synthesis unit selection unit 2
Attribute information for the phoneme information / accent information from
The waveform information is output in response to the address information from the voice waveform synthesizer 4.

In the above-described embodiment, the information of the voice waveform itself stored in the waveform information storage section 5 is not a matter defining the present invention and is not limited to the voice waveform itself, but may be a characteristic parameter thereof. Conceivable.

In the above embodiment, the synthesizing unit selecting section 2 divides the phoneme information / accent information / morpheme information into accent phrases, but this is not a requirement for the present invention and is not limited to accent phrases. It is also possible to divide into larger units and conversely smaller units.

In the above-described embodiment, the phoneme information and the accent information are used as the index of the attribute information in the attribute information storage unit 3. However, prosody information is used instead of the accent information, or one of them is used. It is also possible to use the information of the department. The same applies to the information output from the stored information update unit 7 to the attribute information storage unit 3, the information input to the input terminal 10, and the information stored in the combined word information storage unit 8.

In the above embodiment, the morpheme information is
It is input from the input terminal 1 because it is used as an example of the determination criterion in the information update determination unit 9, but it is necessary to input from the input terminal 1 when it is not used as the determination criterion in the information update determination unit 9. There is no.

In the above embodiment, the synthesizing unit selecting section 2 and the accumulated information updating section 7 compare the phonological information / accent information of the accent phrase with the phonological information / accent information of the synthesizing unit from the attribute information accumulating section 3. Then, the difference is extracted together with the part immediately before and after that, but a method that does not consider the part immediately before and after is also conceivable.

The process of updating the above three stored information will be specifically described below with reference to the drawings.

First, the process of updating the stored information during voice synthesis will be described in detail.

FIG. 2 is a diagram showing a concrete example of a process of updating the stored information at the time of voice synthesis in one embodiment of the present invention.
Further, as an example of a criterion for determining the accumulated information update, it is used that the accent phrase is composed only of the synthesis unit extracted from the original waveform. In FIG. 2, (a) indicates that the notation character of the unregistered continuous word is "speech synthesis", (b) indicates that the reading kana of (a) is "onsei gosei", (c).
Is (on) phonological information (phonetic notation) is "onseegoo"
Seed, (d) indicates that the accent information in (a) is "LHHHHLLL", and (e) indicates the attribute information before update stored in the attribute information storage unit 3.
(F) shows the phoneme information / accent information of the word including the selected synthesis unit, and (g) shows the updated attribute information stored in the attribute information storage unit 3. In the phoneme information of (c), "ee" means making the "e" longer. The accent information in (d) is expressed in two steps, high and low, for simplicity, and "L" means a phoneme with a low pitch and "H" means a phoneme with a high pitch. (E) and (g) 1st row, 2nd row, 3rd
The line, the fourth line, and the fifth line show the attribute information ID, the phoneme information, the accent information, the address information, and the waveform generation information, respectively. In (f), the underlined portion is the composition unit.

The process of updating the stored information will be described below step by step. 1) The synthesis unit selecting unit 2 accesses the attribute information accumulating unit 3 and sets the synthesis unit of the accent phrase “speech synthesis” having the phoneme information “onseegoosee” / accent information “LHHHHLLLL” as shown in (f) of FIG. As shown, "syllabic""on" / "LH", "venus""se"
e "/" HH ", and" goosee "of" photosynthesis "
/ Select "HLLL". 2) The synthesis unit selection unit 2 sends the phoneme information / accent information of the synthesized accent phrase, the attribute information of each synthesis unit, and the synthesis order to the information update determination unit 9. The accent phrase "speech synthesis" in FIG. 2A is equivalent to a continuous word. 3) The information update determination unit 9 compares the attribute information ID of the attribute information of each synthesis unit with the waveform generation information in (e) of FIG. 2 and has the same content in all synthesis units (that is, all synthesis units). Since the unit has been extracted from the original waveform), it is determined that updating is possible. 4) The information update determination unit 9 sends the phoneme information / accent information of the continuous word “speech synthesis”, the attribute information of each synthesis unit, and the synthesis order to the stored information update unit 7. 5) The accumulated information updating unit 7 accesses the waveform information accumulating unit 5 based on the various information from the information update judging unit 9 and synthesizes the speech waveform of the continuous word "speech synthesis". 6) The accumulated information updating unit 7 sends the speech waveform of the continuous word “speech synthesis” to the waveform information accumulating unit 5. 7) The waveform information storage unit 5 registers the speech waveform of the continuous word “speech synthesis” and adds address information. 8) The waveform information storage unit 5 sends the address information to the storage information update unit 7. 9) The accumulated information updating unit 7 sends to the attribute information accumulating unit 3, together with the received address information, the phoneme information / accent information of the continuous word “speech synthesis”, the attribute information of each synthesis unit, and the synthesis order. 10) The attribute information storage unit 3, as shown in (g) of FIG.
From the received various information, the phoneme information “onseegoosee” and the accent information “LH” of the continuous word “speech synthesis”
HHHLLL ", address information" 0014 ", and waveform generation information" 0001 (1-2), 0002 (3-4),
0003 (3-6) "is generated and registered, and the attribute information ID" 0004 "is given to the attribute information.

Next, at the time of non-voice synthesis, the input terminal 10
The process of updating the stored information when the voice waveform, the phoneme information, the accent information, and the cut-out point information are input will be specifically described. In this update, the phoneme information / accent information of the input speech waveform is stored in the attribute information storage unit 3
Or not, the process different from the update process 1 or the update process 2 is performed.

FIG. 3 shows that in the embodiment of the present invention, a voice waveform, phoneme information, accent information, and cut-out point information are input from the input terminal 10 at the time of non-voice synthesis, and waveform information corresponding thereto is already accumulated. FIG. 6 is a diagram showing a specific example of an update process (update process 1) of stored information in the case where the information is stored. In FIG. 3, (a) shows that the notation characters of the voice waveform input from the input terminal 10 are “voice synthesis”,
(B) shows that the phonetic transcription of (a) is "Onsei Gosei", and (c) shows that the phoneme information (phonetic notation) of (a) is "
"onseegoose", and (d) indicates that the accent information in (a) is "LHHHHLLLL". The information in (c) and (d), and the voice waveform and cut-out point information are It is artificially input from the input terminal 10. Further, in Fig. 3, (e) shows the pre-update attribute information stored in the attribute information storage unit 3,
(F) shows the phoneme information / accent information of the word including the selected synthesis unit, and (g) shows the updated attribute information accumulated in the attribute information accumulating unit 3 through the updating process 1. Other expressive assumptions in FIG.
be equivalent to.

The updating process 1 will be described below step by step. 1) The accumulated information updating unit 7 uses the phoneme information “onseegoosee” / accent information “LH from the input terminal 10.
HHHLLLL "is output to the attribute information storage unit 3. 2) Since the attribute information storage unit 3 has the received phoneme information / accent information, the waveform generation information of the attribute information is assigned to the attribute information ID (that is, the original waveform). The address information “0014” is sent to the accumulated information updating unit 7. 3) The accumulated information updating unit 7 receives the received address information “00”.
14 ", and outputs the voice waveform and the cut-out point information to the waveform information storage unit 5. 4) The waveform information storage unit 5 receives the received address information" 00 ".
The waveform information of 14 "is updated to a new one.

In FIG. 4, in one embodiment of the present invention, a speech waveform, phonological information, accent information, and cut-out point information are input from the input terminal 10 during non-speech synthesis, and corresponding waveform information is accumulated. It is a figure which shows a specific example of the update process (update process 2) of the stored information in the case of not being. In FIG. 4, (a) indicates that the notation characters of the voice waveform input from the input terminal 10 are “voice synthesis”,
(B) shows that the phonetic transcription of (a) is "Onsei Gosei", and (c) shows that the phoneme information (phonetic notation) of (a) is "
"onseegoose", and (d) indicates that the accent information in (a) is "LHHHHLLLL". The information in (c) and (d), and the voice waveform and cut-out point information are It is artificially input from the input terminal 10. Further, in Fig. 3, (e) shows the pre-update attribute information stored in the attribute information storage unit 3,
(F) shows the phoneme information / accent information of the word including the selected synthesis unit, and (g) shows the updated attribute information stored in the attribute information storage unit 3 through the update process 2. Note that other expressive assumptions in FIG.
be equivalent to.

The updating process 2 will be described below step by step. 1) The accumulated information updating unit 7 uses the phoneme information “onseegoosee” / accent information “LH from the input terminal 10.
"HHHLLLL" is output to the attribute information accumulating unit 3. 2) Since the attribute information accumulating unit 3 does not have the received phoneme information / accent information, the attribute information accumulating unit 3 notifies the accumulated information updating unit 7 of that fact. Upon reception of that fact, the unit 7 outputs the input speech waveform of the word "voice synthesis" to the waveform information storage unit 5.
Send to. 4) The waveform information storage unit 5 registers the speech waveform of the continuous word "speech synthesis" and adds the address information "0014". 5) The waveform information accumulating unit 5 receives the address information “001
4 ”is sent to the stored information updating unit 7. 6) The stored information updating unit 7 receives the received address information“ 00 ”.
14 ”and the phoneme information / accent information of the continuous word“ speech synthesis ”are sent to the attribute information storage unit 3. 7) As shown in FIG. 4G, the attribute information storage unit 3 calculates the phoneme information of the continuous word “speech synthesis” from the received various information.
onseegoosee ”, accent information“ LHHH
HLLL ", address information" 0014 ", and waveform generation information" 0004 (that is, showing the original waveform) "are generated and registered, and attribute information ID" 0004 "is added to the attribute information.

Finally, the process of updating the stored information will be described in detail when it is in the interval of voice synthesis or when the voice waveform, phoneme information, accent information, and cut-out point information are not input from the input terminal 10. 2A is stored in the continuous word information accumulating unit 8 and the notation character of the continuous word which is not registered in the attribute information accumulating unit 3 is “speech synthesis”. Except what it means
Since it is exactly the same as FIG. 2, the same figure is used to avoid duplication.

The process of updating the stored information will be described below step by step. 1) The accumulated information updating unit 7 accesses the combined word information accumulating unit 8 and the attribute information accumulating unit 3 and searches for the combined word “speech synthesis” that is not registered in the attribute information accumulating unit 3. 2) At the same time, the accumulated information updating unit 7 causes the continuous word “speech synthesis”.
The phoneme information “onseegoosee” and the accent information “LHHHHLLL” are obtained from the continuous word information storage unit 8. 3) The accumulated information updating unit 7, like the composition unit selecting unit 2,
Access the attribute information storage unit 3 to access the continuous word "speech synthesis"
As a synthesis unit of “on” / “LH” of “syllable” and “see” / ”of“ Venus ”as shown in (f) of FIG.
"HH", and "goosee" / "HLL" of "photosynthesis"
4) The accumulated information updating unit 7 accesses the waveform information accumulating unit 5 based on the phoneme information and accent information of the continuous word “speech synthesis”, the attribute information of each synthesis unit, and the synthesis order. , Synthesizes the speech waveform of the continuous word "speech synthesis". 5) The accumulated information updating unit 7 sends the speech waveform of the continuous word “speech synthesis” to the waveform information accumulating unit 5. 6) The waveform information storage unit 5 registers the speech waveform of the continuous word "speech synthesis" and adds the address information "0014". 7) The waveform information accumulator 5 stores the address information “001
4 ”is sent to the stored information updating unit 7. 8) The stored information updating unit 7 receives the received address information“ 00 ”.
14 ”together with the phoneme information / accent information of the continuous word“ speech synthesis ”, the attribute information of each synthesis unit, and the synthesis order are sent to the attribute information storage unit 3. 9) As shown in FIG. 2G, the attribute information storage unit 3 calculates the phoneme information of the continuous word “speech synthesis” from the received various information.
onseegoosee ”, accent information“ LHHH
HLLL ", address information" 0014 ", and waveform generation information" 0001 (1-2), 0002 (3-4), 00.
03 (3-6) "is generated and registered, and the attribute information ID" 0004 "is given to the attribute information.

In the specific example of the above-mentioned three updating processes, the accent information is expressed in two stages, but this is not a limitation of the present invention, and the present invention is expressed in three or more stages. It can also be adapted.

In the above-mentioned one specific example, the fact that all the synthesis units are extracted from the original waveform is taken as the update determination standard in the information update determining section 9, but this standard is merely one specific example. The present invention is not limited to this. For example, it is possible that a word obtained by removing a particle or auxiliary verb from an accent phrase is not registered in the attribute information storage unit 3.

From the above description, the present invention uses phonological information and accent information to synthesize a waveform by gluing together the speech waveform itself such as a word or a phrase or a part cut out from the speech waveform. It is possible to register a new synthesized waveform during speech synthesis or to input a new speech during non-speech synthesis, while satisfying that it can generate more arbitrary speech waveforms than the waveform editing method with good voice quality with high clarity. In order to automatically update and add information used for waveform synthesis by registering a waveform or synthesizing and registering a speech waveform of a complex word in the complex word information storage unit 8,
It is possible to shorten the speech synthesis processing time.

[0046]

As described in detail above, the present invention is
The divided input information obtained by managing the accumulated voice waveform cut out in units of words or phrases in advance in association with the attribute information including the phoneme information, and dividing the input phoneme information or the like corresponding to the desired voice into one or more, In a voice synthesis method for synthesizing a desired voice by collating the synthesis unit attribute information, which is all or a part of the attribute information, and searching the stored voice waveforms corresponding to the synthesis unit attribute information and joining them. It is determined whether to add the accumulated speech waveform and its attribute information in it. If it is determined to be added, they are added, or the compound word is accumulated by correlating the compound word with its phoneme information etc. Section, search for consecutive words that are not associated with the accumulated speech waveform, generate a speech waveform corresponding to that consecutive word using the accumulated speech waveform and its attribute information, or input from outside. By accumulating the voice waveform and its attribute information as the accumulated voice waveform and its attribute information, the accumulated voice waveform and its attribute information can be automatically added and updated, thus shortening the voice synthesis processing time. It can be realized and its effect is extremely large.

[Brief description of drawings]

FIG. 1 is a schematic block diagram of an embodiment of a speech synthesis system of the present invention.

FIG. 2 is a specific example of a process of updating stored information at the time of voice synthesis according to an embodiment of the present invention, and whether it is between voice synthesis, or a voice waveform, phoneme information, accent information, and cutout from the input terminal 10. It is a figure which shows one specific example of the update process of the accumulated information when point information is not input.

FIG. 3 shows a case where a speech waveform, phoneme information, accent information, and cut-out point information are input from the input terminal 10 during non-speech synthesis in one embodiment of the present invention, and waveform information corresponding thereto is already accumulated. FIG. 6 is a diagram showing a specific example of an update process (update process 1) of stored information in FIG.

FIG. 4 shows a case where a voice waveform, phoneme information, accent information, and cut-out point information are input from the input terminal 10 during non-voice synthesis in one embodiment of the present invention, and waveform information corresponding thereto is not accumulated. It is a figure which shows a specific example of the update process (update process 2) of the accumulation information of.

FIG. 5 is a basic block diagram of a waveform editing method which is one of conventional methods.

FIG. 6 is a basic block diagram of a rule combining method which is one of conventional methods.

FIG. 7 is a schematic block diagram of a speech synthesis method using a complex speech unit, which is one of conventional methods.

[Explanation of symbols]

 1 Input Terminal 2 Synthesis Unit Selection Section 3 Attribute Information Storage Section 4 Voice Waveform Synthesis Section 5 Waveform Information Storage Section 6 Output Terminal 7 Stored Information Update Section 8 Conjunction Word Information Storage Section 9 Information Update Judgment Section 10 Input Terminal 51 Input Terminal 52 Synthesis Unit decomposition unit 53 Speech waveform editing unit 54 Speech waveform accumulation unit 55 Output terminal 61 Input terminal 62 Input terminal 63 Synthesis unit decomposition unit 64 Characteristic parameter accumulation unit 65 Speech waveform synthesis unit 66 Prosody rule accumulation unit 67 Output terminal 71 Input terminal 72 Speech Unit selection unit 73 Voice unit file 74 Voice control knowledge file 75 Unit combination unit 76 Synthetic waveform generation unit 77 Output terminal

Claims (1)

[Claims] 1. An information storage unit for storing and managing a stored voice waveform, which is cut out in advance in units of words or phrases, in association with attribute information consisting of phonological information of the stored voice waveform, and which corresponds to a desired voice. The input information obtained by dividing the input phoneme information as it is or into a plurality of pieces is collated with synthesis unit attribute information which is all or a part of the attribute information, and the accumulated speech waveform corresponding to the synthesis unit attribute information is searched for, and the accumulation is performed. A method of synthesizing the voice by joining the voice waveforms, wherein it is determined whether the accumulated voice waveform and the attribute information corresponding to the accumulated voice waveform are added during the synthesis of the voice, and it is determined that the attribute is added. In this case, the accumulated speech waveform and the attribute information corresponding to the accumulated speech waveform are added to the information accumulating section, and the consecutive words are accumulated in association with the phoneme information of the consecutive words. The information accumulating section is searched for the consecutive words that are not associated with the accumulated speech waveform, and the speech waveform corresponding to the uncombined consecutive words is defined as the accumulated speech waveform and the attributes of the accumulated speech waveform. Generated using information, the generated voice waveform corresponding to the continuous word and its attribute information are added to the information storage unit, and the voice waveform input from the outside and the attribute information of the voice waveform are stored in the stored voice waveform and Automatically adding or updating the stored voice waveform and the attribute information corresponding to the stored voice waveform to the information storage unit by adding or updating to the information storage unit as the attribute information of the stored voice waveform. A voice synthesis method characterized by.