JP2009244661A - Device, method, and program for speech synthesis - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To efficiently find a phoneme of a deletion object which contributes to sound quality improvement. <P>SOLUTION: A device for speech synthesis comprises: a candidate phoneme acquiring section 3 for acquiring the phoneme which is used for synthetic speech as a candidate phoneme based on language processing of an input text; a phoneme selecting section 4 for selecting the phoneme which is most suitable for synthetic speech in the candidate phonemes, as an optimum phoneme by calculating a phoneme selection score which is an index for indicating suitability in the speech synthesis of the candidate phoneme based on language processing result; and a waveform creation section 6 for creating a synthetic speech waveform based on the optimum phoneme. When the phoneme of the deletion object is included in the optimum phoneme, the candidate phoneme acquiring section 3 acquires the candidate phonemes other than this phoneme again. The waveform creation section 6 includes an improvement perspective index calculation section 140 for calculating an improvement perspective index which indicates sound improvement possibility, when the phoneme is deleted, in a speech synthesis device for creating the synthetic speech waveform again. The improvement perspective index calculation section 140 makes an improvement perspective index large, if a phoneme selection score is high. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to speech synthesis technology, and more particularly to a speech synthesizer, a speech synthesis method, and a speech synthesis program for synthesizing speech from text.

  Conventionally, various speech synthesizers have been developed that analyze a text sentence and generate synthesized speech by rule synthesis from speech information indicated by the sentence. FIG. 11 is a block diagram showing a configuration of a general rule synthesis type speech synthesizer. Details of the configuration and operation of the speech synthesizer having such a configuration are described in Non-Patent Documents 1 to 3, and Patent Documents 1 and 2, for example.

  The speech synthesizer shown in FIG. 11 includes a language processing unit 1, a prosody generation unit 2, a candidate segment acquisition unit 3, a segment selection unit 4, a segment information storage unit 5, and a waveform generation unit 6. It has.

  The unit information storage unit 5 stores a speech unit generated for each speech synthesis unit and attribute information of each speech unit. Here, the speech segment is information used to generate a synthesized speech waveform, and is often extracted from a recorded natural speech waveform. Examples of speech segments include speech waveforms themselves cut out for each synthesis unit, linear prediction analysis parameters, cepstrum coefficients, and the like.

  The speech element attribute information refers to the phoneme environment of natural speech from which each speech element is extracted, and the phoneme information and prosodic information such as pitch frequency, amplitude, and duration information.

  As a speech synthesis unit, phonemes, CV, CVC, VCV (V is a vowel and C is a consonant) are often used. Details of the length of the speech element and the synthesis unit are described in Non-Patent Document 1, Non-Patent Document 3, and Non-Patent Document 5.

  The language processing unit 1 performs linguistic processing such as morphological analysis, syntax analysis, reading of the text sentence and reading of the accent on the input text sentence, and a symbol string representing “reading” such as a phoneme symbol; The morpheme parts of speech, utilization, accent type, and the like are output to the prosody generation unit 2, candidate segment acquisition unit 3, and segment selection unit 4 as language processing results.

  The prosody generation unit 2 is based on the linguistic processing result output from the language processing unit 1 and is generated by prosody information (pitch, time length, power, etc.) of the synthesized speech based on the strength of the sound, the length, the shortness, the height, etc. Information related to the rhythm of the word to be generated) and generated to the segment selection unit 4 and the waveform generation unit 6.

  The candidate segment acquisition unit 3 refers to the language processing result, selects a speech unit that may be used for the synthesized speech from the speech units stored in the segment information storage unit 5, and This is transmitted to the selection unit 4.

  The unit selection unit 4 selects a speech unit having a high fitness with respect to the language processing result and the generated prosodic information from the units supplied from the candidate unit acquisition unit 3, and selects the selected speech unit. It outputs to the waveform generation part 6 with attached information.

  The waveform generation unit 6 generates a waveform having a prosody close to the prosody generated by the prosody generation unit 2 from the selected speech segment, connects the waveforms, and outputs the synthesized speech. Note that waveform synthesis is described in Non-Patent Document 4.

  In the above processing, the segment selection unit 4 performs predetermined synthesis of information representing the characteristics of the target synthesized speech (hereinafter referred to as “target segment environment”) from the input language processing result and prosodic information. Calculate for each unit. The information contained in the target segment environment includes the corresponding / preceding / following phoneme name, presence / absence of stress, distance from the accent core, pitch frequency and power of synthesis unit, unit duration, cepstrum, MFCC (Mel Frequency Cepstral Coefficients), and Δ amount thereof (change amount per unit time).

  Next, when a target segment environment is given, the segment selection unit 4 selects a speech unit that matches specific information (mainly corresponding phoneme) specified by the target segment environment from the segment information storage unit 5. Select multiple pieces. The selected speech unit is a candidate speech unit used for synthesis. Then, a “score (or cost)” that is an index indicating the appropriateness of the selected candidate segment as a speech segment used for synthesis is calculated. Since the goal is to generate a high-quality synthesized speech, if the score is high (or the cost is low), that is, the degree of appropriateness is high, the quality of the synthesized speech is high. Therefore, it can be said that the score is an index for estimating the degree of deterioration of the quality of the synthesized speech. In Non-Patent Document 6, cost is used for selecting speech segments.

  Here, the score calculated by the segment selection unit 4 includes a unit score and a connection score. The unit score represents the estimated sound quality degradation level caused by using the candidate segment under the target segment environment, and is calculated based on the similarity between the segment environment of the candidate segment and the target segment environment. . On the other hand, the connection score represents the estimated sound quality degradation level caused by the discontinuity of the segment environment between connected speech segments, and is calculated based on the affinity of the segment environment between adjacent candidate segments. The

  Various methods for calculating the unit score and the connection score have been proposed so far. In general, information included in the target segment environment is used to calculate the unit score, and the connection score includes the pitch frequency, cepstrum, MFCC, short-time autocorrelation, power, and Δ of these at the segment boundary. Quantity etc. are used.

  As described above, the unit score and the connection score are calculated by using a plurality of pieces of various pieces of information (pitch frequency, cepstrum, power, etc.) related to the segment. After calculating the unit score and the connection score for each segment, a speech segment that maximizes both the connection score and the unit score is uniquely obtained for each synthesis unit.

  The segment obtained by the score maximization is referred to as an optimal segment because it is selected as the most suitable segment for speech synthesis from the candidate segments. When the unit selection unit 4 obtains each optimum unit for all synthesis units, the unit selection unit 4 finally outputs an optimum unit sequence (optimum unit sequence) to the waveform generation unit 6 as a selection result.

  The segment selection unit 4 calculates the score and selects the optimum segment. However, since the calculation formulas and parameters used for selecting the optimum segment are inappropriate, the best sound quality is not necessarily obtained. The piece to be achieved is not necessarily selected.

  In addition, there is a possibility that an element containing deterioration that cannot be determined by a score prepared in advance is selected. For example, an element mixed with sudden noise is inappropriate as an optimum element and should be excluded when selecting an element. However, since sudden noise is not considered in the unit score or connection score, the unit score and the connection score do not move away from the optimum segment due to a decrease in the score, and may be selected as the optimum segment.

  Therefore, in Patent Document 3, for the purpose of solving the above-described problem, the generated synthesized speech is heard, and a segment that seems to be poor in sound quality is found and deleted, that is, a use-prohibited segment is deleted. There is a method for editing synthesized speech.

  In Patent Document 3, the designation of segment deletion is performed for each unit. For example, assuming that the synthesis unit is a syllable and performing segment deletion for synthesized speech whose utterance content is `` Konchiwa '', if it is determined that the sound quality of `` ko '' is poor, the element used in the unit of `` ko '' Designate a piece as a prohibited piece. Since the prohibited use pieces are not used for the synthesized speech, the sound quality of the places where the use prohibited pieces have been used is improved. In order to efficiently find the use-prohibited segment, Patent Document 3 proposes a method of presenting a synthesized speech editor with segments that are considered to have poor sound quality, that is, with a low segment selection score.

Further, Patent Document 4 selects an optimum speech unit for each synthesis unit for each frequency based on the evaluation value, obtains and stores a total evaluation value at that frequency, and pitches at a frequency having the best total evaluation value. The pattern is used at the time of synthesis.
Japanese Patent Laying-Open No. 2005-091551 JP 2006-084854 A JP 2006-313176 A JP 2004-138728 A Huang, Acero, Hon: “Spoken Language Processing”, Prentice Hall, pp. 689-836, 2001. Ishikawa: “Basics of Prosodic Control for Speech Synthesis”, IEICE Technical Report, Vol. 100, No. 392, pp. 27-34, 2000. Abe: “Synthetic unit basis for speech synthesis”, IEICE technical report, Vol. 100, No. 392, pp. 35-42, 2000. Moulines, Charapentier: “Pitch-Synchronous Waveform Processing Techniques For Text-To-Speech Synthesis Using Diphones”, Speech Communication 9, pp. 435-467, 1990. Segi, Takagi, Ito: “A CONCATENATIVE SPEECH SYNTHESIS METHOD USING CONTEXT DEPENDENT PHONEME SEQUENCES WITH VARIABLE LENGTH AS SEARCH UNITS”, Proceedings of 5th ISCA Speech Synthesis Workshop, pp. 115-120, 2004 Kawai, Toda, Ni, Tsuzaki, Tokuda: “XIMERA: A NEW TTS FROM ATR BASED ON CORPUS-BASED TECHNOLOGIES”, Proceedings of 5th ISCA Speech Synthesis Workshop, pp. 179-184, 2004

  However, the conventional speech synthesizer described in Patent Document 3 described above has the following problems.

  It is a problem that a location with a low segment selection score is not necessarily a location with a high possibility of improvement. Since the element having the highest score is basically selected as the optimum element, it is unlikely that the score is significantly increased by deleting the optimum element at a position having a low score. That is, if the score is low, it can be said that there is a low possibility that a segment with good sound quality will appear.

  Further, even if a segment containing sound quality degradation that is difficult to exclude by score is deleted, if the score is low and the sound quality is low, the sound quality improvement effect by deleting the segment is small. Therefore, it is not appropriate to intensively search for a low-score part as a method of finding a prohibited unit and efficiently improving the quality of synthesized speech.

  The invention described in Patent Document 4 has a problem in that the response in a situation where the number of speech segments is not sufficient is insufficient, and a portion with poor sound quality is likely to occur suddenly.

  The present invention has been made in view of the above problems, and realizes a speech synthesizer, a speech synthesis method, and a speech synthesis program that can efficiently find a speech segment to be deleted that should be prohibited from use. For the purpose.

  In order to solve the above-described problems, a speech synthesizer according to the present invention includes a language processing unit that performs language processing including input text reading and accent analysis, morphological analysis, and syntax analysis on the input text, and the language Prosody generation unit that generates prosody information related to the strength, shortness, and high / low of the sound based on the processing result, and a speech unit that may be used for synthesized speech based on the result of the language processing as candidate segments A candidate segment acquisition unit to acquire, and based on the prosodic information and the result of the linguistic processing, calculate a segment selection score which is an index indicating appropriateness in speech synthesis of the candidate segment; A candidate segment acquisition unit including: a unit selection unit that selects a speech unit optimal for synthesized speech as an optimal unit; and a waveform generation unit that generates a synthesized speech waveform based on the optimal unit. Is the optimal If the segment includes a segment to be deleted, the candidate segment is re-executed by removing the segment, and the waveform generator regenerates a synthesized speech waveform in the speech synthesizer. An improved expected index calculation unit that calculates an improved expected index that represents a high possibility that sound quality will be improved when a piece is deleted, and the improved expected index calculation unit calculates the expected improvement if the segment selection score is large It is characterized by increasing the index.

  In order to solve the above-described problems, a speech synthesis method according to the present invention includes a language processing procedure for performing language processing including input text reading and accent analysis, morphological analysis, and syntax analysis on the input text, and the language Prosody generation procedure for generating prosody information related to sound strength, short / long and high / low based on the result of processing, and speech unit that may be used for synthesized speech based on the result of language processing as candidate segments Based on the candidate segment acquisition procedure to be acquired, and the prosodic information and the result of the language processing, a segment selection score that is an index indicating the appropriateness in speech synthesis of the candidate segment is calculated. A candidate segment acquisition procedure comprising: a segment selection procedure for selecting a speech unit optimal for synthesized speech as an optimal segment; and a waveform generation procedure for generating a synthesized speech waveform based on the optimal segment. In the speech synthesis method for regenerating a synthesized speech waveform, when the optimum segment includes a segment to be deleted, the candidate segment is removed again, and the waveform generation procedure regenerates a synthesized speech waveform. An improved expected index calculation procedure for calculating an improved expected index that indicates the likelihood that sound quality will be improved when a segment to be deleted is deleted, and the improved expected index calculation procedure has a larger segment selection score. For example, the improvement expected index is increased.

  In order to solve the above-described problem, a speech synthesis program according to the present invention is based on input text, including language processing including reading and accent analysis, morphological analysis, and syntax analysis of the input text, and a result of the language processing. A prosody generation process for generating prosody information related to the strength, shortness, and height of the sound, and a candidate unit for acquiring a speech unit that may be used for synthesized speech based on the result of the language processing as a candidate unit Based on the results of the acquisition process, the prosodic information, and the language processing, a unit selection score that is an index indicating the appropriateness in speech synthesis of the candidate unit is calculated, and the best choice for synthesized speech from the candidate units A candidate segmentation process by causing a computer to execute a segment selection process for selecting a correct speech segment as an optimal segment and a waveform generation process for generating a synthesized speech waveform based on the optimal segment. In the speech synthesis program for regenerating the synthesized speech waveform, the processing is performed again when the optimal segment includes a segment to be deleted, excluding the segment, and obtaining the candidate segment. , Causing the computer to execute an improved expected index calculation process for calculating an improved expected index that indicates a high possibility that sound quality will be improved when the segment to be deleted is deleted. If the one-side selection score is large, the improvement expected index is increased.

  According to the present invention, an improvement expected index indicating the high possibility of sound quality improvement when deleting a segment to be deleted is calculated, and if the segment selection score is large when calculating the improvement expected index, the improvement is expected. By increasing the index, it is possible to realize a speech synthesizer, a speech synthesis method, and a speech synthesis program that can efficiently find a speech unit to be deleted that should not be used.

Next, the configuration of the embodiment of the present invention will be described in detail with reference to the drawings.
[First Embodiment]
FIG. 1 is a block diagram showing a configuration of a speech synthesizer according to the first embodiment of the present invention. In the configuration according to the present embodiment shown in FIG. 1, the language processing unit 1, prosody generation unit 2, candidate segment acquisition unit 3, segment selection unit 4, segment information storage unit 5, waveform generation unit 6, deletion target An unusable element information acquisition unit 11, an unusable element information storage unit 12, an optimum element information storage unit 13, and an improvement expected index calculation unit 140 to which an element deletion command for designating an element is input.

  The unit information storage unit 5 stores a speech unit generated for each speech synthesis unit and attribute information of each speech unit.

  The prohibited-use segment information storage unit 12 selects a segment that is designated as unusable as a synthesized speech among the segments registered in the segment information storage unit 5, that is, a segment to be excluded from candidate segments. It is recorded.

  In the initial state, the unusable segment information storage unit 12 has no stored segments, and if the synthesized speech editor designates the segment deletion, the record is accumulated, and the recorded contents after the synthesized speech editing is completed. Is generally erased. However, there may be a usage method in which the segments that are permanently prohibited from use are recorded from the initial stage, and a usage method in which the use prohibited segments are registered cumulatively without erasing the recorded content even after editing the synthesized speech. .

  The optimum information segment storage unit 13 stores the segment information selected by the segment selection unit 4. Accordingly, there is no recorded content in the initial state, and the recorded content is erased after editing of the synthesized speech is completed. Further, every time the segment selection is executed, the content of the optimum information segment storage unit 13 is updated.

  Next, the detailed operation of the speech synthesizer according to the first embodiment will be described with reference to the block diagram of FIG.

  FIG. 2 is a flowchart for explaining the operation of the speech synthesizer according to the first embodiment of the present invention.

  In FIG. 2, the language processing unit 1 performs analysis such as morphological analysis, syntax analysis, and reading on the input text sentence, and as a language processing result, a prosody generation unit 2, a candidate segment acquisition unit 3, It outputs to the segment selection part 4 (step S101).

  The prosody generation unit 2 generates the prosody information of the synthesized speech based on the language processing result output from the language processing unit 1, and outputs it to the segment selection unit 4 and the waveform generation unit 6 (step S102).

  The candidate segment acquisition unit 3 refers to the language processing result supplied from the language processing unit 1 and the prohibited segment information stored in the prohibited segment information storage unit 12, and the segment information storage unit A speech unit that can be used for synthesized speech is selected from speech units registered in 5 and transmitted to the unit selection unit 4. At this time, an element registered as a prohibited element or an element whose reading is different is excluded from candidate candidates. In addition, segments that have significantly different linguistic features (such as the relative relationship between accent phrase boundaries and the distance from the accent kernel) compared to the language processing result may be excluded from the candidates (step S103).

  The unit selection unit 4 is a unit supplied from the candidate unit acquisition unit 3 with a speech unit having a high degree of fitness with respect to the language processing result supplied from the language processing unit 1 and the prosody information generated by the prosody generation unit 2. The information is selected from the pieces, and is transmitted to the optimum unit information storage unit 13 and the waveform generation unit 6 together with the attached information of the selected speech unit. Also, each candidate segment calculated when selecting the optimal segment and the segment selection score (unit score, connection score, etc.) of each segment are transmitted to the improvement expected index calculation unit 140 as segment selection information. (Step S104).

  The waveform generation unit 6 generates a waveform having a prosody close to the prosody generated by the prosody generation unit 2 from the speech unit selected by the unit selection unit 4, and connects these waveforms to generate synthesized speech. Output (step S105). The prosody of the generated waveform may faithfully reproduce the prosody generated by the prosody generation unit 2, or the generated prosody may be reproduced based on only the prosody of the selected segment while ignoring the generated prosody. Prosody may be generated.

  The optimum information segment storage unit 13 updates the segment information selected by the segment selection unit based on the segment selection information supplied from the segment selection unit 4 (step S106).

  Based on the segment selection information supplied from the segment selection unit 4, the expected improvement index calculation unit 140 estimates the possibility of sound quality improvement by segment deletion for each unit, and outputs the estimated improvement index (step) S107).

  When calculating an improvement expected index using the segment selection information, a segment selection score is mainly used. If the segment selection score is high, there is a high possibility that the sound quality will be improved when the segment is deleted. The unit selection score used for calculation of the expected improvement index mainly includes a unit score, a connection score, and a score calculated based on both (for example, the sum of both).

  In order to derive an improvement expected index based on the segment selection score, estimation of the segment selection score after the segment deletion is required. As a method considered to be ideal, there is a method of obtaining a unit selection score after deleting a unit before deleting, that is, a method of obtaining a unit selection score obtained by deleting the current optimum unit for each unit. Can be mentioned.

  However, in order to calculate, it is generally necessary to execute the deletion of the optimum segment and the reselection process of each segment for each unit, so that a large amount of calculation is often required.

  Therefore, as a method for estimating the segment selection score after the segment deletion by the improvement expected index calculation unit or the like, a method using the segment selection score of the current optimum segment as an estimated value, or a plurality of segments having a high score A promising method is to calculate an estimated value based on the score.

In the method of calculating the estimated value based on the score of multiple segments with high scores, the average value of the score, the weighted sum, etc. are calculated for each unit, and the estimated value of the segment selection score after deleting the segment is calculated. To do. For example, when the top N rank of the score is S ₁ to S _N (S ₁ > S ₂ >...> S _N > 0) and the estimated value T is calculated by the weighted sum, the calculation formula is as follows: Given in.

但し、ａ₁, ａ₂, …, ａ_N は正の実数である。素片削除後に選択される可能性が高い素片は高スコアの素片であることから、ａ₁＞ａ₂＞...＞ａ_Nという関係を満足することが望ましい。又、最適素片のスコアを除いて推定値を計算する方法も有望である。これは、最適素片のみが大幅にスコアが高く、準最適な素片のスコアが低い場合には、最適素片を削除するとスコア及び音質が大幅に低下する可能性が高いためである。

Here, a ₁ , a ₂ ,..., A _N are positive real numbers. Since a segment having a high possibility of being selected after deleting the segment is a segment having a high score, it is desirable to satisfy the relationship of a ₁ > a ₂ >...> a _N. Also, a method for calculating the estimated value by removing the score of the optimum segment is promising. This is because if only the optimal segment has a significantly high score and the semi-optimal segment has a low score, it is highly likely that the score and sound quality will be significantly reduced if the optimal segment is deleted.

  As a method for obtaining an improvement expected index from the estimated value of the segment selection score after deleting the segment, for example, a method of calculating the improvement expected index using a linear function of the estimated score as shown below can be cited. At this time, when the estimated score is x (> 0) and the expected improvement index is F (x), the relationship between them is given by the following equation (2).

但し、ａ、ｂは実数で、ａ＞０である。

However, a and b are real numbers, and a> 0.

  Further, as shown in FIG. 3, there is a method in which a table showing the relationship between the estimated score and the expected improvement index is prepared in advance and the index value is determined with reference to the table. In addition, there are a method of using an exponential function, a quadratic or higher-order multidimensional function and a polynomial function as the function F (x), and a method of outputting an estimated value of the unit selection score itself. In any method, if the estimated score is high, the improvement probability tends to be high, which is a condition that should be satisfied by the calculation of the improvement expected index.

  In step S108, the prohibited-use element information acquisition unit 11 should refer to the input element deletion command and the optimum element information stored in the optimum-element information storage unit 13 to be deleted, ie, prohibited to use. When the information on the element is obtained and the element deletion instruction is acquired (YES), the information on the element to be prohibited is transmitted to the use prohibited element information storage unit 12, and the use prohibited element information storage unit 12 is also transmitted. Is transmitted to the candidate segment acquisition unit 3 (step S109). Further, the prohibited-use element information storage unit 12 updates the prohibited-use element information based on the transmitted information on the prohibited-use element (step S110).

  A segment deletion command is acquired for each synthesis unit. For example, if the synthesis unit is syllable and the utterance content of the synthesized speech is "Konchiwa", the unit deletion is specified from "ko", "n", "ni", "chi", "wa" Is done. The designated place may be only one of “ko” or two places of “ko” and “chi”. The segment at the designated location is uniquely identified from the optimum segment information.

  When the segment deletion is performed, it is necessary to regenerate the synthesized speech excluding the segment to be deleted, so that a signal for notifying that the prohibited-unit information storage unit 12 has been updated is sent to the candidate segment acquisition unit 3. , The process from acquisition of candidate segments to waveform generation of synthesized speech is executed again.

  When the candidate element acquisition unit 3 receives a signal notifying that the prohibited element information storage unit 12 has been updated, the candidate element acquisition unit 3 again uses the language processing result and the prohibited element information stored in the prohibited element information storage unit 12. With reference to the piece information, a speech unit that can be used for the synthesized speech is selected and transmitted to the unit selection unit 4 (step S103).

  Then, the procedure after step S104 is executed again, and if a segment deletion command is not acquired in step S108 (NO), a series of processing according to the present embodiment ends.

  As described above, according to the present embodiment, the speech synthesizer uses the property that if the segment selection score is high, the sound quality improvement expected by segment deletion is high, and based on the segment selection score. Calculate the expected improvement index. At that time, the unit selection score after deleting the unit is estimated from the score of the optimal unit or a plurality of high-scoring units for each unit, and reflected in the calculation of the expected improvement index of each unit. For this reason, it becomes possible to estimate the score and sound quality that change due to the segment deletion, and it is possible to obtain an improved expected index with higher reliability than in the past.

  The reason for this is that it is more efficient to search for places with high likelihood of improvement over places with low segment selection scores, and places with high chance of improvement are not places with low scores, but places with high scores but poor sound quality That's why.

In places where the score is high, a segment having a high level score is often selected again even if the score drops slightly due to segment deletion. Therefore, since it is highly likely that an alternative segment having a high score and sound quality is selected when the segment is deleted, the quality of the synthesized speech can be improved efficiently.
[Second Embodiment]
In the first embodiment, the expected improvement index is calculated using the segment selection score. However, it is difficult to completely trust the segment selection score as a sound quality index. If the segment selection score is a sufficiently reliable indicator for finding a prohibited unit to be deleted, a high-quality segment should be selected as the optimal segment when selecting a segment. Therefore, it can be said that the work itself of finding and specifying prohibited pieces is not necessary.

  Therefore, if information other than the segment selection score is also used, it is possible to obtain a more reliable improvement expected index. Therefore, in the second and third embodiments, an example in which an improvement expected index is calculated using information other than the unit selection score will be described.

  Here, FIG. 4 is a block diagram showing the configuration of the speech synthesizer according to the second embodiment of the present invention.

  The configuration according to the present embodiment shown in FIG. 4 is characterized by including a candidate segment acquisition unit 31 and an improvement expected index calculation unit 141 as compared with the first embodiment shown in FIG. To do. The detailed operation of the speech synthesizer according to the second embodiment will be described below with reference to the block diagram of FIG.

  FIG. 5 is a flowchart for explaining the operation of the speech synthesizer according to the second embodiment of the present invention. FIG. 5 is a flowchart for explaining the operation of the first embodiment. Compared with FIG. 2, step S101, step S102, step S104, step S105, step S106, step S108, step S109, and step S110 is common, and includes step S203 instead of step S103 in FIG. 2, and step S207 instead of step S107 in FIG.

  In step S <b> 203, the candidate segment acquisition unit 31 refers to the language processing result supplied from the language processing unit 1 and the prohibited segment information stored in the prohibited segment information storage unit 12. A speech unit that may be used for synthesized speech is selected from speech units registered in the information storage unit 5 and transmitted to the unit selection unit 4. Also, the number of candidate segments for each unit is transmitted to the expected improvement index calculation unit 141.

  Based on the number of candidate segments supplied from the candidate segment acquisition unit 31, the estimated improvement index calculation unit 141 estimates the possibility of sound quality improvement by segment deletion for each unit, and outputs it as an expected improvement index ( Step S207).

  When calculating the expected improvement index using the number of candidate segments, if the number of candidate segments is small, it is unlikely that the alternative segment will achieve high sound quality even if the segment deletion is specified. Reduce the value of the expected improvement index. The number of segments registered in the segment information storage unit 5 and the number of segments acquired by the candidate segment acquisition unit 31 is often different for each unit. That is, the number of pieces that can be replaced differs depending on the unit type. For example, when the synthesis unit is a syllable, the number of segments such as “Wa” and “O” is large, but the number of segments of “Vu” may be small.

  Further, the large number of candidate segments means that there are many segments having various feature quantities (pitch frequency, duration length, cepstrum, etc.). For this reason, there is a high possibility that an optimum segment having a high segment selection score will appear at a location where the number of candidate segments is large. Therefore, if the number of candidate segments is large, there is a high possibility that an alternative segment with a high segment selection score will appear.

  As a method of calculating the improvement expected index from the number of candidate segments, the method used for calculating the improvement expected index from the estimated score is similarly used when obtaining the improvement expected index using the segment selection score. In any method, if the number of candidates is large, the improvement probability tends to be large, which is a condition to be satisfied in the calculation of the improvement expected index.

As described above, according to the present embodiment, using the property that the probability that a segment with a high segment selection score will appear increases as the number of candidates increases, the improvement index is increased as the number of candidates increases. To do. In particular, in comparison with the first embodiment, since the improvement expected index is calculated without using the segment selection score, it is more effective than the first embodiment in a situation where the segment selection score is not sufficiently reliable. It is possible to obtain a measure for improvement.
[Third Embodiment]
Next, a speech synthesizer according to the third embodiment of the present invention will be described. Here, FIG. 6 is a block diagram showing a configuration of a speech synthesizer according to the third embodiment of the present invention.

  In the configuration according to the present embodiment shown in FIG. 6, compared to the first embodiment shown in FIG. 1, the use prohibited piece information acquisition unit 112, the improvement expected index calculation unit 142, and the use prohibited piece acquisition. A number-of-times calculation unit 152 is provided.

  The detailed operation of the speech synthesizer according to the third embodiment will be described below with reference to the block diagram of FIG.

  FIG. 7 is a flowchart for explaining the operation of the speech synthesizer according to the third embodiment of the present invention. FIG. 7 is a flowchart for explaining the operation of the first embodiment, compared with FIG. 2, step S101, step S102, step S103, step S104, step S105, step S106, step S108 and step. S110 is common, and has step S307 instead of step S107 in FIG. 2 and step S309 instead of step S109 in FIG. 2, and includes step S311 separately.

  In this step S307, the expected improvement index calculation unit 142 determines that the sound quality may be improved by deleting the unit for each unit based on the number of the prohibited unit acquisition obtained from the prohibited unit acquisition number calculating unit 152. Estimate and output as an expected improvement index.

  When calculating the expected improvement index using the number of unusable element acquisitions, if the number of unusable element acquisitions is large, it is unlikely that the alternative element will achieve high sound quality even if you specify deletion of the element. Basically, decrease the value of the expected improvement index. When performing segment deletion a plurality of times for the same unit, the segments with the highest score are deleted in order. Therefore, at a place where many deletions have been made, it waits for an appearance of a high-quality sound element from among the low-scoring elements among the candidate elements at that position, so that the sound quality improvement at that place is Prospect decreases.

  As a method for calculating an improvement expected index from the number of use-prohibited segment acquisitions, the method used for calculating the improvement expected index from the estimated score is similarly used when obtaining the improvement expected index using the segment selection score. . In any method, if the number of use-prohibited pieces is small, the improvement expectation tends to increase. This is a condition that should be satisfied by the calculation of the improvement expected index. Therefore, when using equation (2), it is a condition that a is a negative real number.

  In step S108, the prohibited-use element information acquisition unit 112 refers to the input element deletion command and the optimum element information stored in the optimum element information storage unit 13, and should be deleted, that is, prohibited to use. When the element information is obtained and the element deletion command is acquired (YES), the information on the element to be prohibited is transmitted to the use prohibition element information storage unit 12 and the use prohibition element acquisition frequency calculation unit 152. At the same time, a signal notifying that the use prohibition element information storage unit 12 has been updated is transmitted to the candidate element acquisition unit 3 and the use prohibition element acquisition count calculation unit 152 (step S309).

  In step S <b> 311, the use prohibition element acquisition count calculation unit 152 receives the signal indicating that the use prohibition element information storage unit 12 has been updated from the use prohibition element information acquisition unit 112 each time. The number of times the piece information is acquired is updated and transmitted to the expected improvement index calculation unit 142.

  Since the initial value of the use prohibition element acquisition count calculation unit 152 is set to zero, the use prohibition element is not received unless a signal notifying that the use prohibition element information storage unit 12 has been updated is not received. The acquisition count calculation unit 152 outputs zero.

  The number of use-prohibited segment acquisitions is counted for each unit (for each syllable if the synthesis unit is a syllable).

  As described above, according to the present embodiment, the number of use-prohibited segment acquisitions is obtained by using the property that the probability that a unit with a high unit selection score will appear increases if the use-prohibited unit acquisition count is small. If there is little, raise the expected improvement index.

  Compared to the first embodiment, since the expected improvement index is calculated without using the segment selection score, the improvement that is more effective than the first embodiment in a situation where the segment selection score is not sufficiently reliable. An expected index can be obtained.

Also, it is effective when you want to preferentially consider parts that are not deleted in order to improve the overall sound quality of the synthesized speech in a situation where deletion is repeated multiple times at multiple places. Is the method.
[Fourth Embodiment]
Next, a speech synthesizer according to the fourth embodiment of the present invention will be described. Here, FIG. 8 is a block diagram showing a configuration of a speech synthesizer according to the fourth embodiment of the present invention.

  In the configuration according to the present embodiment shown in FIG. 8, compared to the first embodiment shown in FIG. 1, the candidate segment acquisition unit 31, the prohibited use segment information acquisition unit 112, and the expected improvement index calculation unit 143 and a use prohibition element acquisition frequency calculation unit 152 are provided.

  The candidate element acquisition unit 31 is equivalent to the candidate element acquisition unit 31 in the block diagram of the second embodiment of FIG. 4, and the prohibited element information acquisition unit 112 and the prohibited element acquisition frequency calculation unit 152 is equivalent to the unusable element information acquisition unit 112 and the unusable element acquisition frequency calculation unit 152 in the block diagram of the third embodiment of FIG.

  The detailed operation of the speech synthesizer according to the fourth embodiment will be described below with reference to the block diagram of FIG.

  FIG. 9 is a flowchart for explaining the operation of the speech synthesizer according to the fourth embodiment of the present invention. FIG. 9 is common to step S101, step S102, step S104, step S105, step S106, step S108, and step S110 as compared to FIG. 2 which is a flowchart for explaining the operation of the first embodiment. It is.

  However, in place of step S103 in FIG. 2, step S203 equivalent to step S203 in FIG. 5, step S309 equivalent to step S309 in FIG. 7 instead of step S109 in FIG. 2, and step S107 in FIG. Step S407 is provided, and Step S311 equivalent to Step S311 of FIG. 7 is separately provided.

  In this step S407, the improvement expectation index calculation unit 143 determines the unit selection information supplied from the unit selection unit 4, the number of candidate units supplied from the candidate unit acquisition unit 31, and the number of prohibited use unit acquisition times. Based on the number of use-prohibited segment acquisitions supplied from the calculation unit 152, the possibility of sound quality improvement by segment deletion is estimated for each unit, and output as an expected improvement index.

  Examples in which the element selection information, the number of candidate elements, and the number of use-prohibited element acquisitions are used independently will be described in the first embodiment, the second embodiment, and the third embodiment. In this embodiment, a method of using these in combination will be described.

In this embodiment, after obtaining the estimated score from the segment selection information, if the estimated score is high, if the number of candidate segments is large, if the number of use-prohibited segments is small, the value of the expected improvement index is increased. . As a method of calculating the expected improvement index, there is a method of calculating a weighted sum of the estimated score, the number of candidate segments, and the number of use-prohibited segments obtained. In this case, the expected improvement index T is given by the following equation (3) with respect to the estimated score S ₁ , the number of candidate segments S ₂ , and the number of use-prohibited segment acquisition times S ₃ .

但し、ａ₁, ａ₂, ａ₃, ｂは実数であり、ａ₁＞０, ａ₂＞０, ａ₃＜０ を満たす。又、図１０に示すように、推定スコア、候補素片数、使用禁止素片取得回数と改善見込指標の関係を示す表を事前に用意し、その表を参照して指標値を決定する方法もある。

However, a ₁ , a ₂ , a ₃ , and b are real numbers and satisfy a ₁ > 0, a ₂ > 0, and a ₃ <0. Also, as shown in FIG. 10, a method of preparing a table showing the relationship between the estimated score, the number of candidate segments, the number of use-prohibited segment acquisitions, and an expected improvement index in advance, and determining an index value by referring to the table There is also.

  In addition, there are a method using an exponential function, a quadratic or higher-order multidimensional function, and a polynomial function instead of the expression (3), and a method for outputting the estimated score, the number of candidate segments, and the number of prohibited use unit acquisitions as they are. .

  The above is a method for directly calculating an improvement expected index from the estimated score, the number of candidate segments, and the number of prohibited use units, but the first embodiment, the second embodiment, and the third embodiment. A method of calculating an improvement expected index individually by the method described in the embodiment and calculating one improvement expected index from each improvement expected index may be used.

  As described above, according to the present embodiment, the expected improvement index is calculated using the segment selection score, the number of candidate segments, and the number of use-prohibited segment acquisitions. For this reason, it is possible to obtain an improvement expected index with higher reliability than in the first embodiment.

  In particular, when certain information (e.g., segment selection score) is equivalent, it is possible to correct the expected improvement based on other information (number of candidate segments or number of prohibited use), so each information Improvement of the expected improvement index can be expected compared to the case of using alone.

  The present invention is not limited to the speech synthesizer described in each embodiment, and the configuration and operation thereof can be changed as appropriate without departing from the spirit of the invention.

  The present invention can be realized by hardware, software, or a combination thereof.

  The present invention can be used in a speech synthesizer, a speech synthesis method, and a speech synthesis program that convert text into speech with high sound quality.

It is a block diagram which shows the structure of the speech synthesizer which concerns on the 1st Embodiment of this invention. It is a flowchart for demonstrating operation | movement of the speech synthesizer which concerns on the 1st Embodiment of this invention. It is an example of the table | surface which the improvement expected index calculation part of FIG. 1 uses for calculation of improvement expected index calculation. It is a block diagram which shows the structure of the speech synthesizer which concerns on the 2nd Embodiment of this invention. It is a flowchart for demonstrating operation | movement of the speech synthesizer which concerns on the 2nd Embodiment of this invention. It is a block diagram which shows the structure of the speech synthesizer which concerns on the 3rd Embodiment of this invention. It is a flowchart for demonstrating operation | movement of the speech synthesizer which concerns on the 3rd Embodiment of this invention. It is a block diagram which shows the structure of the speech synthesizer which concerns on the 4th Embodiment of this invention. It is a flowchart for demonstrating operation | movement of the speech synthesizer which concerns on the 4th Embodiment of this invention. It is an example of the table | surface which the improvement expected index calculation part of FIG. 8 uses for improvement expected index calculation. It is a block diagram which shows one structural example of a general speech synthesizer.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Language processing part 2 Prosody generation part 3, 31 Candidate element acquisition part 4 Element selection part 5 Element information storage part 6 Waveform generation part 11, 112 Use prohibition element information acquisition part 12 Use prohibition element information storage part 13 Optimal segment information storage unit 140, 141, 142, 143 Improvement expected index calculation unit 152 Unusable segment acquisition count calculation unit

Claims (18)

A language processing unit that performs language processing including input text reading and accent analysis, morphological analysis, and syntax analysis on the input text, and prosodic information related to sound intensity, length, shortness, and height based on the results of the language processing A prosody generation unit that generates a speech unit that may be used for synthesized speech based on the result of the language processing, a candidate unit acquisition unit that acquires a speech unit as a candidate unit, and the prosody information and the language processing Based on the result, a unit selection score, which is an index indicating the appropriateness of the candidate unit in speech synthesis, is calculated, and a speech unit optimal for the synthesized speech is selected as an optimal unit from the candidate units. A segment selection unit and a waveform generation unit that generates a synthesized speech waveform based on the optimum segment, and the candidate segment acquisition unit includes a segment to be deleted in the optimum segment, Candidates excluding the fragment Executed the acquisition pieces again, the waveform generating unit in the speech synthesizer to regenerate a synthesized speech waveform,
An improved expected index calculating unit that calculates an improved expected index that represents a high possibility of sound quality improvement when the deletion target segment is deleted;
The speech synthesis apparatus according to claim 1, wherein the improvement expected index calculation unit increases the improvement expected index if the segment selection score is large. A deletion designation number calculation unit for calculating the number of deletion designations for designating the element to be deleted;
The speech synthesis apparatus according to claim 1, wherein the expected improvement index calculation unit increases the expected improvement index if the number of times of deletion designation is small.   The speech synthesis apparatus according to claim 1 or 2, wherein the improvement expected index calculation unit increases the improvement expected index if the number of candidate segments acquired by the candidate segment acquisition unit is large.   The improvement expected index calculation unit estimates a segment selection score after deleting the segment to be deleted, and calculates the improvement expected index based on the estimated segment selection score The speech synthesizer according to any one of claims 1 to 3.   The improvement expectation index calculation unit estimates a weighted sum of those showing a high numerical value among the unit selection scores as a unit selection score after deleting the unit to be deleted. Item 5. The speech synthesizer according to Item 4.   The improvement expectation index calculation unit estimates an average value of those showing a high numerical value among the unit selection scores as a unit selection score after deleting the unit to be deleted. 4. The speech synthesizer according to 4. Language processing procedure for performing language processing including input text reading and accent analysis, morphological analysis and syntax analysis on the input text, and prosodic information related to sound intensity, length, short and high based on the result of the language processing A prosody generation procedure for generating speech units, a candidate segment acquisition procedure for acquiring speech units that may be used for synthesized speech based on the results of the language processing as candidate segments, the prosodic information and the language processing Based on the result, a unit selection score, which is an index indicating the appropriateness of the candidate unit in speech synthesis, is calculated, and a speech unit optimal for the synthesized speech is selected as an optimal unit from the candidate units. A segment selection procedure and a waveform generation procedure for generating a synthesized speech waveform based on the optimum segment, and the candidate segment acquisition procedure includes a segment to be deleted in the optimum segment, The fragment Except executed the acquisition of candidate segments again, the waveform generation procedure in the speech synthesis method for regenerating a synthesized speech waveform,
An improved expected index calculation procedure for calculating an improved expected index that represents a high possibility of sound quality improvement when the deletion target segment is deleted;
In the speech synthesis method, the improvement expected index calculation procedure increases the improvement expected index if the segment selection score is large. A deletion designation number calculation procedure for calculating the number of times of deletion designation for designating the fragment to be deleted;
The speech synthesis method according to claim 7, wherein in the improvement expected index calculation procedure, the improvement expected index is increased when the number of times of deletion designation is small.   The speech synthesis method according to claim 7 or 8, wherein the improvement expected index calculation procedure increases the expected improvement index if the number of candidate segments acquired in the candidate segment acquisition procedure is large.   The improved expected index calculation procedure estimates a segment selection score after deleting the segment to be deleted, and calculates the improved expected index based on the estimated segment selection score The speech synthesis method according to any one of claims 7 to 9.   The improvement expectation index calculation procedure estimates a weighted sum of those showing a high numerical value among the unit selection scores as a unit selection score after deleting the unit to be deleted. Item 11. The speech synthesis method according to Item 10.   The improvement expectation index calculation procedure is to estimate an average value of those showing a high numerical value among the unit selection scores as a unit selection score after deleting the unit to be deleted. 10. The speech synthesis method according to 10. Prosody generation that generates prosodic information related to sound intensity, length, short and high based on the results of language processing including input text reading and accent analysis, morphological analysis, and syntactic analysis for the input text Processing, candidate segment acquisition processing for acquiring speech units that may be used for synthesized speech based on the results of the language processing as candidate segments, and the prosody information and the language processing results based on the results of the language processing A unit selection process for calculating a unit selection score that is an index indicating the appropriateness in speech synthesis of candidate units, and selecting a speech unit optimal for synthesized speech from the candidate units as an optimal unit; A waveform generation process for generating a synthesized speech waveform based on the optimal segment, and when the candidate segment acquisition process includes a segment to be deleted in the optimal segment, In the exception executed the acquisition of candidate segments again, the speech synthesis program the waveform generation processing to regenerate synthesized speech waveform,
Causing the computer to execute an improved expected index calculation process for calculating an improved expected index that represents a high possibility of sound quality improvement when the deletion target segment is deleted;
The speech synthesis program characterized in that the improvement expected index calculation processing increases the improvement expected index if the segment selection score is large. Further causing the computer to execute a deletion designation number calculation process for calculating the number of deletion designations for designating the element to be deleted,
14. The speech synthesis program according to claim 13, wherein in the improvement expected index calculation process, the improvement expected index is increased if the number of times of deletion designation is small.   The speech synthesis program according to claim 13 or 14, wherein the improvement expected index calculation process increases the improvement expected index if the number of candidate segments acquired in the candidate segment acquisition process is large.   The improvement expected index calculation process estimates a segment selection score after deleting the segment to be deleted, and calculates the improvement expected index based on the estimated segment selection score The speech synthesis program according to any one of claims 13 to 15.   The improvement expectation index calculation process estimates a weighted sum of those showing a high numerical value among the unit selection scores as a unit selection score after deleting the unit to be deleted. Item 17. The speech synthesis program according to Item 16.   The improvement expectation index calculation process estimates an average value of those showing a high numerical value among the segment selection scores as a segment selection score after deleting the segment to be deleted. 16. The speech synthesis program according to 16.

Images