JP3892691B2 - Speech synthesis method and apparatus, and speech synthesis program - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
  The present invention relates to a speech synthesis method and apparatus for stably obtaining high-quality synthesized speech, and a speech synthesis program.
[0002]
[Prior art]
  Conventionally, a great deal of effort has been expended in creating content such as radio, television, cable broadcasting, and the Internet. In the current multi-channel era such as cable broadcasting and the Internet, creating all content manually requires a lot of work. Cost and labor are required. For this reason, a speech synthesis system that generates synthesized speech from text data has attracted attention.
[0003]
  If a speech synthesis system is used, the cost for content creation can be reduced, and a large amount of content can be created in a short time.
[0004]
  As this type of speech synthesis system, for example, a corpus-based speech synthesis system using a text corpus is known.
[0005]
  The speech synthesis system belonging to the above-mentioned corpus-based speech synthesis searches the database for the optimal synthesis unit for a synthesized sentence based on low-order acoustic information and low-order physical information, and performs prosody modification. Synthetic speech is created by connecting without connecting. In this way, speech synthesis is performed with respect to the real voice by not changing the prosody.
[0006]
  Here, the low-order information refers to physical information close to air vibration. For example, information such as fundamental frequency and spectrum that can be directly digitized from a speech waveform.
[0007]
[Problems to be solved by the invention]
  However, in the above-mentioned conventional speech synthesis system, although the prosody is not changed, the quality of the real voice is respected, but the synthesis unit is based on low-order parameters that are far from the group of sentences, such as cepstrum and phonemes. Since it is selected, there is no consistency in consideration of the whole sentence, and the synthesized result becomes unstable, such as connection within the phrase or a voice different from the original prosodic pattern of the sentence.
[0008]
  That is, in a system in which a synthesis unit is selected based on parameters obtained from low-order information, the quality of synthesized speech greatly depends on the database search result. For this reason, the quality is good when a synthesis unit having a high degree of matching with the target sentence can be selected, otherwise the quality is degraded, and the quality of the synthesized speech is unstable.
[0009]
  For this reason, it is difficult to say that the conventional speech synthesis system can realize sufficient quality required for broadcasting such as the above-described radio, television, cable broadcasting, and Internet.
[0010]
  SUMMARY OF THE INVENTION In view of the above problems, an object of the present invention is to realize a speech synthesis system having sufficient quality for broadcasting and to enable automatic creation and update of contents, a speech synthesis method and apparatus, and a speech synthesis program. Is to provide.
[0011]
[Means for Solving the Problems]
  In order to achieve the above object, the present invention provides three types of information: information indicating the boundaries of phrases, role information of words indicating whether phrases are emphasized speech, and speech prosodic pattern information.News andA first database in which speech waveforms, phoneme string data, and data indicating phoneme boundaries are stored in association with each other;For each deterioration pattern defined by the word role information and the prosodic pattern information,A first value obtained by a plurality of listeners evaluating a plurality of speech waveforms connected at a phrase boundary and having different fundamental frequency differences before and after the connection point, and a difference between fundamental frequencies connected within the phrase and before and after the connection point A second value obtained by evaluating a plurality of listeners with sounds of a plurality of sound waveforms having different valuesTogaUsing a second database stored as information, a synthesized speech is generated using a small phoneme sequence consisting of at least a part of the phoneme sequence recorded in the database as a synthesis unit candidate. .
[0012]
  In speech synthesis, the synthesis unit candidates that match a part of the phoneme string indicated by the text data are searched from the database, and for each of the searched synthesis unit candidates, the synthesis unit candidates Three types of information indicating the boundary of the phrase, role information of the word, and prosodic pattern informationEvaluation based on informationThe value is the first value and the second valueAndThe average value of the evaluation values at each connection point is obtained as a comprehensive evaluation value, and the candidate of the composite unit having the highest comprehensive evaluation value is selected as the composite unit..
[0013]
  Furthermore, a speech waveform corresponding to the selected synthesis unit is extracted from the database, and the extracted speech waveform is connected in correspondence with the phoneme string indicated by the text data.
[0014]
  For example, the deterioration state of the quality differs depending on whether or not the connection point between the speech waveform fragments is at the phrase boundary. In addition, emphasized words are words that have important meanings in the text, often with increased power and pitch, and the state of deterioration of quality differs depending on whether the words are emphasized. Become.
[0015]
  The prosodic pattern information can be used to determine the continuity and consistency of the prosodic pattern at the connection point between speech fragments. In the continuity of prosodic patterns, for example, if a speech fragment with a high fundamental frequency at the beginning is connected after a speech fragment with a fundamental frequency gradually decreasing at the end, the quality will deteriorate because the place where the fundamental frequency is supposed to decrease increases. Cheap. Conversely, if a speech fragment having a lower fundamental frequency at the head is connected after a speech fragment whose fundamental frequency is gently lowered at the end, the quality is unlikely to deteriorate because the degree of decrease only increases.
[0016]
  As for the consistency of prosodic patterns, look at ± (plus / minus) of the direction of the difference in fundamental frequency at the connection point between speech fragments, and whether it matches the inclination direction of the continuous prosodic pattern of the sentence The quality deterioration state differs depending on
[0017]
  In this way, synthesis is selected as usual by selecting synthesis units based on evaluation values obtained using information indicating the boundaries of phrases, higher-order linguistic information such as word role information and prosodic pattern information. It is possible to solve the problem of instability in the quality of synthesized speech caused by using low-order information for unit selection.
[0018]
DETAILED DESCRIPTION OF THE INVENTION
  Hereinafter, an embodiment of the present invention will be described with reference to the drawings.
[0019]
  FIG. 1 is a block diagram showing a speech synthesizer according to an embodiment of the present invention. In this embodiment, a speech synthesizer is configured by installing a speech synthesis program in a computer.
[0020]
  In FIG. 1, reference numeral 1 denotes a computer, which is a central processing unit 11 mainly composed of a CPU, a memory 12 connected to the central processing unit 11, a display unit 13, an input unit 14, and a memory composed of semiconductor elements. 15 and a general computer including an acoustic unit 16 including a digital / analog (D / A) converter as a sound generator and a speaker.
[0021]
  The storage unit 12 stores the speech synthesis program 17, the text analysis dictionary 18 for speech synthesis, and the evaluation value information 19, and various information necessary for speech synthesis is constructed as a speech synthesis database 20. ing.
[0022]
  The input unit 14 is means for inputting text data to be synthesized, and includes, for example, an interface to an information recording medium such as a keyboard, a mouse, a porcelain floppy disk and a compact disk, a network communication interface, and the like.
[0023]
  In the speech synthesis database 20, as a database necessary for text-to-speech synthesis, a speech waveform 21, phoneme string data 22 corresponding to the utterance content, data 23 indicating phoneme boundaries, information 24 indicating phrases boundaries, and phrases are emphasized. Word role information 25 indicating whether or not the speech is present and speech prosodic pattern information 26 are stored in association with each other. As the speech waveform 21, for example, speech actually uttered by a person is recorded and stored.
[0024]
  In this embodiment, the speech synthesis program 17 operates the central processing unit 11 to perform speech synthesis processing. In this speech synthesis process, when generating synthesized speech adapted to kana-kanji mixed text data input via a keyboard, an information recording medium, or a network, the speech synthesis database (hereinafter simply referred to as database) 20 is used. A speech waveform fragment (hereinafter referred to as a speech fragment) as a synthesis unit is selected. In this selection, a synthesis unit candidate is selected top-down from higher-order linguistic information, and a speech fragment corresponding to this synthesis unit is selected.
[0025]
  Higher-level linguistic information refers to information with semantic content that interprets physical phenomena perceived by humans, and is information that gives meaning by interpreting physical information such as air vibrations by human language ability. is there. For example, it is information that cannot be digitized directly from speech waveform fragments, such as phrase boundaries, word roles, and prosodic pattern consistency, and requires several levels of abstraction. In the present embodiment, as high-order linguistic information, boundary information of phrases (voiced speech segments and exhalation paragraphs sandwiched between pauses), word role (word emphasis utterances, etc.) information, prosodic patterns (connection points) The continuity and consistency of the prosodic pattern at the end of the previous speech waveform fragment and the beginning of the speech waveform fragment after the connection point, etc. The prosodic pattern is information on changes in phonemes such as pitch and accent). Note that the prosodic pattern information includes information regarding the variation pattern of the fundamental frequency. In addition to emphasizing utterance information, information obtained by judging emotional expressions such as marvel, enthusiasm and disappointment from the context of the surroundings can be used as the role information of words. .
[0026]
  With such high-order linguistic information, it is possible to stably output high-quality synthesized speech, rather than selecting a synthesis unit using only low-order information. Furthermore, by using higher-level linguistic information, it is possible to solve the problems of the prior art such as a sense of incongruity at the connection point between selected speech fragments and a prosodic pattern different from the meaning of the sentence. .
[0027]
  Therefore, according to the present embodiment, since a synthesis unit is selected using high-order linguistic information, high-quality synthesized speech can be obtained stably. As a result, it is possible to create synthesized speech having the quality required in the field of broadcasting, and it is also possible to automatically create and update content in order to make up for the lack of content in the multi-channel era of cable broadcasting and the Internet.
[0028]
  FIG. 2 is a flowchart for explaining the processing of the speech synthesis program 17 in the present embodiment, and FIG. 3 is a diagram showing the flow of speech synthesis using high-order linguistic information performed by the speech synthesis program 17. Specific operations will be described with reference to these drawings.
[0029]
  When text data 101 of a kana-kanji mixed sentence created by a word processor or the like for speech synthesis is input (S1), the text data 101 is analyzed 102 (S2). In the text analysis 102, the text data 101 is analyzed using the text analysis dictionary 18. For example, for the input text 101 of the kana-kanji mixed sentence, the reading kana, the phoneme string corresponding to the reading kana, and the continuation of each phoneme Information such as the length of time, the accent type for each phrase, how to combine each phrase, and the prosodic pattern corresponding to the accent type and how to combine is output as an analysis result.
[0030]
  After this, the composition unit search 103 is performed using the text analysis result. In the synthesis unit search 103, the phoneme sequence of the text data 101 is divided into a plurality of subphoneme sequences, and the phoneme sequence that matches the text data 101 at the phoneme level is searched from the database 20 (S3). The phoneme sequence extracted from the database 20 by the search is set as a synthesis unit candidate 104 used for speech synthesis (S4). Here, the subphoneme sequence is a part of a phoneme sequence composed of a plurality of phonemes, and is a phoneme sequence composed of one or more phonemes.
[0031]
  Further, the narrowing down of candidates 105 is performed using the above-described higher-level linguistic information (S5), and for each subphoneme sequence from among a plurality of synthesis unit candidates 104 extracted as candidates for each subphoneme sequence of the text data 101. One optimum synthesis unit is determined (S6).
[0032]
  In the candidate narrowing-down process 105, the candidate narrowing-down 105 is performed using high-order linguistic information stored in the database 20 from a plurality of composition unit candidates 104 extracted by the composition unit search 103. The narrowed synthesis unit candidate 106 is output. Further, the narrowed synthesis unit candidate 106 is again used as the synthesis unit candidate 104, the candidate narrowing down 105 is performed, and the final synthesis unit is selected by narrowing down the candidate synthesis unit to the last one.
[0033]
  In the present embodiment, as high-order linguistic information, information 24 indicating the boundary between phrases, word role information 25 indicating whether or not the phrase is emphasized speech, and speech prosodic pattern information 26, one of three pieces of information. The composition unit is narrowed down using two or more pieces of information.
[0034]
  Next, the speech fragment associated with the selected synthesis unit is acquired from the database 20 (S7), and the acquired speech fragment is sequentially connected so as to correspond to the phoneme string of the text data 101 (S8). The obtained continuous speech waveform is sent to the acoustic unit 16 and outputted from the speaker as speech (S9).
[0035]
  Note that the continuous speech waveform obtained by connecting speech fragments in the process of S8 may be temporarily stored as speech data in a memory or information storage medium. Needless to say, it is also possible to easily input synthesized speech created by another apparatus as speech data and output it as speech.
[0036]
  Next, the main part of the speech synthesis in this embodiment will be described in more detail.
[0037]
  FIG. 4 is a diagram for explaining a main part of the speech synthesis in the present embodiment, and narrows down the optimum synthesis unit from the plurality of searched synthesis unit candidates 104 based on higher-order linguistic information. The flow is shown.
[0038]
  The synthesis unit candidate 104 retrieved by the process of S3 is obtained by extracting a subphoneme sequence that matches the subphoneme sequence constituting the input text data 101 from the database 20, and the phonemes stored in the database 20 are extracted. A plurality of consonant sequence candidates 104 are usually selected as synthesis unit candidates 104.
[0039]
  For example, when speech is to be synthesized from the text data 101 “Tomorrow will be cloudy then rain”, the speech waveform corresponding to the text content “ASITAWAKUMORIDESU” is stored in the database 20 Of these, the part “ASHITAWAKUMORI” is extracted as a candidate for the synthesis unit.
[0040]
  In addition, the “ASITAWAK” portion of the above can be extracted as a candidate for the synthesis unit. That is, the synthesis unit in the present embodiment is extracted regardless of the delimiters such as mora, syllables, and phrases. In other words, the synthesis unit may be any one obtained by matching and extracting the phoneme string that can be used to synthesize the target text from the database 20 described above.
[0041]
  Here, when selecting a composition unit, it can be used as “ASITAWAKUMORI” and “NOTIAMEDESYOU”, and can also be used as “ASITAWA” and “KUMORINOTIAMEDESYOU”. For this reason, it is necessary to decide how to select a variable-length synthesis unit. Conventionally, the cepstrum distance in the input text and the database is obtained, and the phoneme environment and the fundamental frequency are compared. However, the present invention is different from the prior art in that the phrase boundary described above is selected when selecting the synthesis unit. Is to use higher-level linguistic information such as However, since a conventional text analysis technique or the like is used as a technique for extracting high-order linguistic information when the database 20 is constructed, there is no particular difference from the conventional technique. The present invention is not a part for extracting information, but a part for using extracted high-order linguistic information for selection of a synthesis unit is a significant difference from the prior art.
[0042]
  Next, a method for determining how to use variable-length synthesis units will be described.
[0043]
  In the present embodiment, in the above-described synthesis unit narrowing-down process 105, three pieces of information are used: information 24 indicating the boundaries of phrases, word role information 25 indicating whether the phrases are emphasized speech, and speech prosodic pattern information 26. ing. Note that the synthesis unit may be determined using one of the phrase boundary information 24, the word role information 25, the prosodic pattern information 26, or any two pieces of information.
[0044]
  In other words, in the present embodiment, these three types of information are used as parameters for quality degradation and quality evaluation values (hereinafter simply referred to as evaluation values) with respect to a difference (gap) between fundamental frequencies at connection points between speech fragments selected as synthesis units. Used to determine).
[0045]
  As an example, a graph of a quality deterioration pattern is shown in FIG. In this graph, the horizontal axis indicates the difference in the basic frequency of the synthesis unit at the connection point between the audio fragments, and the vertical axis indicates the quality when the audio fragments are connected and synthesized. The value indicating the quality evaluation is calculated based on the subjective evaluation value when the person listens to the synthesized speech.
[0046]
  In FIG. 5, the evaluation value for the difference in fundamental frequency at the connection point when the connection point between speech fragments is at the phrase boundary, and the difference in fundamental frequency at the connection point when the connection point between speech fragments is in the phrase. The evaluation value is expressed.
[0047]
  When the connection point is at a phrase boundary, the evaluation values when the fundamental frequency differences are 0, 30, 60, 90, 120, and 150 (Hz) are 98, 100, 93, 85, 93, and 87, respectively. (%). When the connection point is in the phrase, the evaluation values when the fundamental frequency difference is 0, 30, 60, 90, 120 (Hz) are 98, 80, 58, 48, 18 (%), respectively. ).
[0048]
  Thus, when the connection point is at the phrase boundary, a good evaluation value can be obtained even if the difference in fundamental frequency at the connection point increases. When the connection point is in the phrase, the evaluation value gradually decreases as the fundamental frequency difference increases.
[0049]
  As can be seen from the graph of FIG. 5, the quality deteriorates as the difference (gap) between the fundamental frequencies at the connection points of the audio fragments increases, and the degree of deterioration is the condition of the connection points (in this case, whether it is a phrase boundary or not). ) Will vary. In the present embodiment, this is referred to as “quality degradation pattern determined from the condition of the connection point”.
[0050]
  As a method of narrowing down the synthesis unit from the quality deterioration pattern (hereinafter referred to as the deterioration pattern), in this embodiment, based on the deterioration pattern determined from the difference of the fundamental frequency at the connection point between the audio fragments and the condition of the connection point. Thus, the synthesis unit is narrowed down by calculating the quality evaluation value for the difference between the fundamental frequencies.
[0051]
  That is, in the present embodiment, a plurality of synthesized speech waveforms are prepared as evaluation values, and values set based on values obtained by the listener evaluating the speech based on these speech waveforms are used. For example, the value evaluated by the listener for a plurality of speech waveforms connected at a phrase boundary and having different fundamental frequency differences before and after the connection point, and a plurality of differences between the fundamental frequencies connected within the phrase and before and after the connection point are different. To the value evaluated by the listener and the value evaluated by the listener for the sound of multiple sound waveforms with different fundamental frequency fluctuation patterns before and after the connection point and with different fundamental frequency differences before and after the connection point The value set based on this is used. As these evaluation values, for example, values (for example, average values) obtained by standardizing values evaluated by a plurality of listeners for one sound obtained by connecting two synthesis units are used. This evaluation value is also included in the evaluation value information 19 and stored in the storage unit 12. Here, a percentage is used as an evaluation value, and a scale is used in which the quality deteriorates as it approaches 0% without changing to 100%.The
[0052]
  Furthermore, in the present embodiment, a total evaluation value is calculated for all combinations of synthesis units that can synthesize speech corresponding to the target text, and a combination of synthesis units that maximizes the evaluation value is obtained.
[0053]
  That is, the number of synthesis units for synthesizing the target text is different depending on the selection of the synthesis unit. As the number of synthesis units increases, the degradation also increases. For example, the average value of the evaluation values at each connection point is set as the overall evaluation value, and the synthesis unit with the highest overall evaluation value is selected.The
[0054]
  On the other hand, as the phrase boundary information 24 used as higher-order linguistic information, information indicating whether the connection point between speech fragments is a phrase boundary or a phrase (within a phrase) is used. The deterioration pattern shown in FIG. 5 differs depending on whether or not the connection point between the voice fragments is at the boundary of the phrase. For example, when the connection point exists at a phrase boundary, the slope of the graph becomes gentle and the quality does not deteriorate so much. However, when the connection point exists in the phrase, it easily deteriorates.
[0055]
  As a method of narrowing down synthesis unit candidates using phrase boundaries, for example, when a synthesis unit of “ASITAWAKUMORI” and “NOTIAMEDESYOU” is extracted, the fundamental frequency difference between the end of the former and the beginning of the latter is 30 Hz. Suppose that the quality evaluation value is 80%. Since this connection point is connected not in the phrase boundary but in the middle of the phrase, in order to evaluate the quality deterioration at the time of connection, the deterioration pattern when connected inside the phrase is used.
[0056]
  As another example, when “ASITAWA” and “KUMORINOTIAMEDESYOU” are connected as a synthesis unit, the connection point is determined as a phrase boundary. Phrase boundary information 24 stored in the database 20 is used to determine whether the phrase boundary is present (inside the phrase).
[0057]
  The deterioration pattern when connected at the boundary of a phrase is more gradual than the deterioration pattern when connected inside a phrase. For example, if the difference in fundamental frequency between two synthesis units at this phrase boundary is 90 Hz or less, the quality The evaluation value is larger than 85%, and therefore the quality evaluation value is higher than that of connecting inside the above phrase. Also, if the difference in fundamental frequency between two synthesis units when connected within a phrase is 90 Hz or more, the quality evaluation value will be 48% or less, and the quality will be better when connected at a phrase boundary (Fig. 5). For example, if the fundamental frequency difference at the connection point between the two synthesis units of “ASITAWA” and “KUMORINOTIAMEDESYOU” is 90 Hz, this evaluation value is 100% (phrase boundary), so the above-mentioned “ASITAWAKUMORI” Since the evaluation value of “NOTIAMEDESYOU” is 80% (in the phrase), “ASITAWA” and “KUMORINOTIAMEDESYOU” are selected as synthesis units.
[0058]
  As the word role information 25, information indicating whether the word is emphasized is used. Here, the word “emphasis” is also called “prominence” or “contrast emphasis”, and the strength and weakness in the sentence sound are determined by relative strength to other parts in the sentence. It is to make it stand out relatively with respect to many parts. Also, emphasized words are words that have an important meaning in the text, and often have increased power and pitch. For this reason, depending on whether or not the word is emphasized, the difference in fundamental frequency between the two synthesis units is different as in the case of the phrase boundary described above, so that the deterioration patterns are different.
[0059]
  The prosodic pattern information is used to determine the continuity and consistency of the prosodic pattern at the connection point between speech fragments and the variation pattern of the fundamental frequency.
[0060]
  In continuity of the prosodic pattern, for example, if a speech fragment with a high fundamental frequency at the beginning is connected after a speech fragment with a fundamental frequency slowly decreasing at the end, the fundamental frequency will increase, so the fluctuation of the fundamental frequency will increase. Since the pattern changes greatly, the quality tends to deteriorate. Conversely, if a speech fragment having a lower fundamental frequency at the head is connected after a speech fragment whose fundamental frequency is gently lowered at the end, the quality is unlikely to deteriorate because the degree of decrease only increases.
[0061]
  As for the consistency of prosodic patterns, look at ± (plus / minus) of the direction of the difference in fundamental frequency at the connection point between speech fragments, and match it with the inclination direction (change tendency) of the continuous prosodic pattern that the sentence has The deterioration pattern varies depending on whether or not it is performed.
[0062]
  One or more of these three types of information is used to narrow down the optimum synthesis unit. When one candidate is finally narrowed down from a plurality of candidates, it becomes the selected composition unit.
[0063]
  Also, when combining unit candidates are selected using all three types of information, the phrase boundary information 24, the word role information 25, and the prosodic pattern information 26, for example, two composition units “ASITAWA” and “KUMORINOTIAMEDESYOU” In the connection at the phrase boundary, the evaluation value is calculated using the deterioration pattern under the condition that the directionality of the prosodic pattern is not emphasized at the end of the synthesis unit before the connection point and the head after the connection point. Calculated.
[0064]
  In this embodiment, the case where there is only one connection point has been described as an example. However, even when there are a plurality of connection points, the same selection process as in the case where there is one connection point is repeated, The synthesis unit that finally gives the maximum evaluation value is selected.
[0065]
  The present embodiment can also be applied to a case where, for example, a speech waveform corresponding to each phoneme is connected to generate synthesized speech. However, as the number of connection points increases, quality evaluation for synthesized speech is possible. Needless to say, the quality is lowered because the value is relatively low. Therefore, in this embodiment, if there is a synthesis unit that requires fewer connection points, and it is extracted, it is selected as the synthesis unit.
[0066]
  However, a speech waveform is stored for each phoneme in the database 20, and even in a situation where these must be connected, the speech synthesis method of the present invention can be used as much as possible compared to the conventional example. A good synthesis unit can be selected.
[0067]
  Further, the voice synthesis program 17 can be easily spread to many users by distributing it through an information recording medium such as an optical disk, a magnetic disk, a magneto-optical disk, and a semiconductor memory, a network, and other communication networks. Needless to say.
[0068]
【The invention's effect】
  As described above, the present inventionClaim 1According to the described speech synthesis method, since a synthesis unit is selected using higher-order linguistic information, a high-quality synthesized speech can be obtained stably. As a result, it is possible to create synthesized speech with the quality required in the field of broadcasting, and it is also possible to automatically create and update content to compensate for the lack of content in the multi-channel era of cable broadcasting and the Internet. .
[0069]
  Also,Claim 2According to the described speech synthesizer, since a synthesis unit is selected using high-order linguistic information, high-quality synthesized speech can be obtained stably. As a result, it is possible to create synthesized speech having the quality required in the field of broadcasting, and it is also possible to automatically create and update content in order to make up for the lack of content in the multi-channel era of cable broadcasting and the Internet.
[0070]
  Also,Claim 3According to the described speech synthesis program, the speech synthesis apparatus can be easily configured by being installed in a computer via an information recording medium, a network, or another communication network. Furthermore, since the synthesis unit is selected using higher-order linguistic information, high-quality synthesized speech can be obtained stably. As a result, it is possible to create synthesized speech having the quality required in the field of broadcasting, and it is also possible to automatically create and update content in order to make up for the lack of content in the multi-channel era of cable broadcasting and the Internet.
[Brief description of the drawings]
FIG. 1 is a configuration diagram showing a speech synthesizer according to an embodiment of the present invention.
FIG. 2 is a flowchart illustrating processing of a speech synthesis program according to an embodiment of the present invention.
FIG. 3 is a diagram showing a flow of speech synthesis using high-order linguistic information performed by a speech synthesis program according to an embodiment of the present invention.
FIG. 4 is a diagram for explaining a main part of speech synthesis in one embodiment of the present invention.
FIG. 5 is a diagram showing an example of a graph of a quality deterioration pattern in one embodiment of the present invention.
[Explanation of symbols]
  DESCRIPTION OF SYMBOLS 1 ... Computer, 11 ... Central processing part, 12 ... Memory | storage part, 13 ... Display part, 14 ... Input part, 15 ... Memory, 16 ... Sound part, 17 ... Speech synthesis program, 18 ... Text analysis dictionary, 19 ... Evaluation value Information, 20 ... Speech synthesis database, 21 ... Speech waveform, 22 ... Phoneme sequence, 23 ... Phoneme boundary information, 24 ... Phrase boundary information, 25 ... Word role information, 26 ... Prosodic pattern information, 101 ... Text, 102 ... text analysis, 103 ... search for synthesis unit, 104 ... candidate for synthesis unit, 105 ... narrowing down candidates, 106 ... candidate for synthesis unit, 107 ... connection, 108 ... synthesized speech.

Claims (3)

In a speech synthesis method in which a computer converts text to speech based on text data,
Information indicating the phrase boundaries, words role information that indicates whether the phrase is emphasized sound, and three types of information of the prosodic pattern information of the speech, the speech waveform, a phoneme sequence data, the corresponding data indicating a boundary of phonemes A first database attached and accumulated;
For each deterioration pattern determined by the role information of the word and the prosodic pattern information , a plurality of listeners evaluated sounds by a plurality of speech waveforms connected at phrase boundaries and having different fundamental frequency differences before and after the connection point. A second value in which a value of 1 and a second value obtained by evaluating a plurality of speech waveforms connected by a plurality of speech waveforms with different fundamental frequency differences before and after the connection point are stored as information. Using a database,
A small phoneme sequence consisting of at least a part of the phoneme sequence recorded in the first database is set as a synthesis unit candidate,
Searching the first database for a candidate for the synthesis unit that matches a part of the phoneme string indicated by the text data;
For each of the searched synthesis unit candidates, information indicating the boundary of the phrase, role information of the word, prosodic pattern information with respect to the difference in the fundamental frequency of the speech waveform at the connection point between the synthesis unit candidates. more seek criticism value based on broadcast on the first and second values stored in the second database,
An average value of the evaluation values at each connection point is obtained as a comprehensive evaluation value, and a candidate of a composite unit having the highest comprehensive evaluation value is selected as a composite unit,
Extracting a speech waveform corresponding to the selected synthesis unit from the first database;
The speech synthesis method, wherein the extracted speech waveform is connected in correspondence with the phoneme string indicated by the text data. In a speech synthesizer that converts text into speech based on text data,
Information indicating the phrase boundaries, words role information that indicates whether the phrase is emphasized sound, and three types of information of the prosodic pattern information of the speech, the speech waveform, a phoneme sequence data, the corresponding data indicating a boundary of phonemes A first database attached and accumulated;
For each deterioration pattern determined by the role information of the word and the prosodic pattern information , a plurality of listeners evaluated sounds by a plurality of speech waveforms connected at phrase boundaries and having different fundamental frequency differences before and after the connection point. A second value in which a value of 1 and a second value obtained by evaluating a plurality of speech waveforms connected by a plurality of speech waveforms with different fundamental frequency differences before and after the connection point are stored as information. A database,
A subphoneme sequence consisting of at least a part of the phoneme sequence recorded in the first database is a synthesis unit candidate, and the synthesis unit candidate that matches a part of the phoneme sequence indicated by the text data is selected. Extraction means for searching and extracting from said first database;
For each synthesis unit candidate extracted by the extraction means, information indicating the boundary of the phrase with respect to the difference in the fundamental frequency of the speech waveform at the connection point between the synthesis unit candidates, role information of the word, determined by the first and second values stored leopards value based on prosodic pattern information in said second database, obtaining an average value of the evaluation value at each connection point as a comprehensive evaluation value, the total evaluation A selecting means for selecting the candidate of the synthesis unit with the highest value as the synthesis unit;
Connection means for extracting a speech waveform corresponding to the synthesis unit selected by the selection means from the first database and connecting the extracted speech waveform in correspondence with the phoneme string indicated by the text data. A speech synthesizer characterized by the above. Information indicating the phrase boundaries, words role information that indicates whether the phrase is emphasized sound, and three types of information of the prosodic pattern information of the speech, the speech waveform, a phoneme sequence data, the corresponding data indicating a boundary of phonemes a plurality of the first database that is attached are accumulated, each degradation pattern defined by the role information and the prosody pattern information of the words, the difference between the connected and the front and rear connection points the fundamental frequency phrase boundary different The first value evaluated by a plurality of listeners for the sound generated by the plurality of sound waveforms and the first value evaluated by the plurality of listeners for the sounds generated by the plurality of sound waveforms connected in the phrase and having different fundamental frequencies before and after the connection point. to two values and it can be connected to a second database that is stored as information computer, the speech synthesis process of converting text to speech based on the text data In the speech synthesis program to I,
A subphoneme sequence consisting of at least a part of the phoneme sequence recorded in the first database is a synthesis unit candidate, and the synthesis unit candidate that matches a part of the phoneme sequence indicated by the text data is selected. Retrieving from the first database;
For each of the searched synthesis unit candidates, information indicating the boundary of the phrase, role information of the word, prosodic pattern information with respect to the difference in the fundamental frequency of the speech waveform at the connection point between the synthesis unit candidates. a step of obtaining more hail value based on broadcast on the first and second values stored in the second database,
Obtaining an average value of the evaluation values at each connection point as a comprehensive evaluation value, and selecting a composite unit candidate having the highest comprehensive evaluation value as a composite unit;
Extracting a speech waveform corresponding to the selected synthesis unit from the first database;
And connecting the extracted speech waveform in correspondence with the phoneme string indicated by the text data.

Images