JP2007025323A - Speech synthesizing method, device, program, and recording medium - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To convert text data into a synthesized speech of high quality. <P>SOLUTION: The number of accent phrases, and metrical patterns for the accent phrases are found from an input text. An elementary speech unit series is determined through collation with a database based on the metrical patterns, and quality is evaluated as to the elementary speech unit series. The quality is evaluated by a statistical method, such as, SVM using a plurality of evaluation scales, and elementary speech unit series are determined for the accent phrases, when all evaluations scales are satisfied or when a search for an elementary speech unit series is made as many times as a preset reselection permissible frequency, thus generating the synthesized speech of the best quality. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a speech synthesis technique in which an arbitrary text is input to generate and output speech that matches the text, that is, media conversion from text to speech is performed.

Speech synthesis technology that can generate the desired speech simply by inputting text is used for reading out e-mails and Web articles, reading out information for telephone voice information such as contact centers, etc. Contributes to the reduction of work costs for actually reading out text.
However, the quality of synthesized speech is still not as high as that of human live speech, and it is considered that quality improvement is indispensable for further market expansion.
A waveform-connected speech synthesis technique (Patent Document 1), which is one of speech synthesis techniques for converting arbitrary text into speech, searches for and connects fragments (speech segments) of speech waveforms of arbitrary length from a large number of speech corpora. To generate synthesized speech.

By selecting speech segments from a large number of speech corpora, this speech synthesis method, which covers prosodic variations, basically does not perform signal processing, so the quality of the synthesized speech is maintained.
A speech segment search approach is greatly involved in the quality of synthesized speech by the waveform-connected speech synthesis method.
A general search approach first creates a target phoneme sequence, prosodic pattern, etc. by text analysis of an input kanji sentence. Next, a search for a combination of speech waveforms that is as close as possible to these targets is performed by comprehensively evaluating a plurality of reference parameters.

The most common method for comprehensive evaluation of reference parameters is to calculate the sub-costs related to multiple parameters such as fundamental frequency, prosodic environment, phoneme duration, number of segment connections, etc. In this method, a speech unit sequence is used as output speech.
The weight set, the cost of each parameter, and the total cost in the above-described speech element sequence selection method can be expressed as follows.

この式による評価方法は、各パラメータを総合的に判定しているので、あるパラメータ、例えば、イントネーションに深く影響する、基本周波数に関するパラメータのコストが高く計算されても、他の基準パラメータにおいてコストが低く算出されれば、結果として総合コストが低くなる場合がある。そのため、音声データベース中に、より適切な音声素片の組み合わせが存在しているにもかかわらず、誤ったイントネーションになってしまう音声素片系列が選択されることがしばしばある。

Since the evaluation method based on this formula comprehensively determines each parameter, even if a certain parameter, for example, a parameter related to the fundamental frequency, which has a deep influence on intonation, is calculated at a high cost, the cost for other reference parameters is also low. If it is calculated to be low, the total cost may be low as a result. For this reason, a speech unit sequence that is erroneously entered even though there is a more appropriate speech unit combination in the speech database is often selected.

The value of the weight set for generating high-quality synthesized speech should be set to such a value that such a case does not occur as much as possible. However, depending on the degree of consistency between the input text and the speech database, such as the vocabulary included in the input text, the phoneme sequence included in the input text, and how many speech segments that can reproduce the target prosody pattern are included in the speech database. The optimal weight set values are different. In extreme terms, the optimal weight set value may be different for each input text.
Therefore, a method is proposed in which the quality of the generated synthesized speech is evaluated, and if the quality is low, the synthesis processing method is changed, and a series of processes of synthesizing again is repeated until a high-quality synthesized speech is generated. (Non-Patent Document 1, Patent Document 2).

In Non-Patent Document 1, when the synthesized speech has evaluated whether there is a fundamental frequency drop that can be perceived by humans in the accent nucleus when the fundamental frequency is falling, and it is determined that there is no drop In this case, the weight set is changed and the synthesis process is performed again. In Patent Document 2, the quality evaluation of synthesized speech is determined based on whether or not a certain reference value is satisfied.
Japanese Patent No. 2761552 JP 2004-354644 A Mirai Hasegawa, Hideyuki Mizuno, “Study on segment selection in corpus-based speech synthesis”, Acoustical Society of Japan, 1-7-2, PP.215-216, Mar.2004.

In the said nonpatent literature 1, only the local part called an accent nucleus is visible, and it is not judged whether the intonation is suitable in the place other than that. In addition, only measures related to the fundamental frequency are evaluated. For these reasons, there is a problem that correct quality evaluation cannot be performed.
In Patent Document 2, the quality evaluation of the synthesized speech is determined based on whether or not a certain reference value is satisfied. However, in the method of simply evaluating the synthesized speech quality based on whether or not a certain value is satisfied, It is difficult to evaluate quality with high accuracy, and in some cases, problems such as re-synthesizing synthesized speech that does not have any quality problem or outputting synthesized speech with low quality as it is, cause high quality. It is difficult to generate synthesized speech stably.

Further, in Patent Document 2, when it is determined that the quality is low, the synthesis processing method is changed based on the control information included in the input information. However, the change method based on the quality evaluation result of the actually generated synthesized speech Therefore, the predetermined synthesis processing means cannot always generate high-quality synthesized speech.
An object of the present invention is to propose a speech synthesis method, apparatus, program, and recording medium that can eliminate these disadvantages and generate high-quality synthesized speech.

The speech synthesizer of the present invention includes a database storing speech corpora holding speech data, and searches and connects speech segments of any length included in the speech corpus in order to convert any input text into speech. A speech synthesizer that generates synthesized speech whose content matches the input text,
The characteristic structure is that the input text is read by morphological analysis, the number of accent phrases required for prosodic pattern generation and tone combination information are obtained, and the information obtained by the text analysis processing section A prosody pattern generation unit that generates a target prosody pattern, a processing target accent phrase setting unit that sets an accent phrase to be processed according to the target prosody pattern, and a process that sets a weight set to an initial value for each accent phrase And a weight set setting and setting changing unit for executing a process for changing the set value, and a speech unit for selecting a speech unit based on the prosody pattern generated in the prosody pattern generation step and setting a speech unit sequence A product that evaluates the quality of synthesized speech for each of a plurality of evaluation scales with respect to a speech unit sequence searched in the search unit and the speech unit search step. The evaluation unit and the evaluation result of the quality evaluation unit detect the presence / absence of an evaluation scale determined to be of poor quality, and if the detection result is present, the above voice is set through the processing of the weight set and the setting change unit Return to the speech unit sequence search process in the unit sequence search unit, and if none, output speech unit sequence to determine the speech unit sequence searched at that time as the speech unit sequence of the accent phrase to be processed Each time the speech unit sequence for each accent phrase is determined by the determination unit and the output speech unit sequence determination unit, it is determined whether or not the processing has been completed for all of the accent phrases included in the input text. If it is determined, the speech unit sequence is output to the synthesized speech processing unit for each accent phrase constituting the input text, and if it is determined to be incomplete, the next accent phrase is set as the processing target accent phrase, A weight set setting / setting change unit, a speech unit sequence search unit, a quality evaluation unit, and an input text end determination unit that repeatedly operates the output speech unit sequence determination unit for a preset reselection allowable number of times. .

  Furthermore, in the configuration characterized by the present invention, when the evaluation result of the quality evaluation unit includes an evaluation measure determined to be of poor quality, the number of executions of the speech unit sequence search unit and the quality evaluation unit becomes the reselection allowable number of times. Re-selection allowable number determination unit that returns to the search operation of the speech unit sequence search unit through the weight set setting change processing of the weight set setting and the setting change unit if not achieved, If the reselection allowable number determination unit determines that the achievement is achieved, the current speech processing unit is the speech unit sequence searched at that time together with information that requires signal processing for the accent phrase that is currently being processed. If there is no signal processing necessity determination unit to determine the speech unit sequence of the accent phrase and the quality evaluation unit determines that the quality evaluation unit determines that the quality is poor, for the accent phrase currently being processed, Signal system Together with the information for processing unnecessary, a point obtained by adding a signal processing required determination unit for determining a speech unit sequence being searched at that time the speech unit sequence of the prosodic phrase.

Further, the configuration characterized by the present invention is that the evaluation means provided in the quality evaluation unit is characterized in that the speech unit sequence determined based on the prosodic pattern generated from the input text is at least evaluation means for intonation, evaluation means for speech rate, or continuity. It is a point including any of the evaluation means regarding.
Further, the configuration characterized by the present invention is the weight set set by the weight set setting and setting change unit, the weight of the parameter relating to the fundamental frequency of the speech unit sequence determined based on at least the prosody pattern generated from the input text, This is a point including any combination of the parameter weight related to the duration and the parameter weight related to the gap of the connection point.

According to the present invention, it is possible to evaluate global quality evaluation, that is, quality in a certain long range, which is organized as a prosody, instead of local quality evaluation by setting an accent phrase as a processing target section. In addition, the quality of the entire synthesized speech can be correctly evaluated by evaluating the quality of evaluation scales other than the fundamental frequency.
According to the present invention, by using a statistical method such as SVM, for example, it is possible to associate a non-linear relationship between a plurality of feature quantities and human right / wrong judgment, thereby enabling more accurate quality evaluation and stability. High-quality synthesized speech.

  According to the present invention, quality evaluation by a statistical method is performed using a plurality of evaluation scales, the result is reflected in a weight set, and a speech unit sequence is reselected to stabilize high-quality synthesized speech. Can be obtained.

The synthetic speech quality evaluation method using the statistical method according to the present invention is preferably used by installing a program for realizing it in a computer, and by connecting and combining variable-length speech waveforms, any text can be obtained. An embodiment in which waveform-connected speech synthesis means for converting to speech is implemented in a program and used is the best mode for operating this quality evaluation method.
In order to make the speech synthesizer of the present invention function using a computer, a speech corpus having a large amount of speech data is provided, and any text included in the speech corpus is used to convert any input text into speech. In a speech synthesizer that generates synthesized speech whose contents match the input text by searching for and connecting long speech segments, the input text is read by morphological analysis, and the information required for prosody generation A text analysis processing unit that gives an accent type / tone combination type, a prosody generation unit that generates a target prosody pattern using the information obtained by the text analysis processing unit, and a synthesized speech prosody pattern, etc. Comprehensive evaluation of multiple reference parameters, and a speech segment sequence search unit that selects speech segments based on the evaluation results, and the quality of the generated synthesized speech A quality evaluation unit capable of globally evaluating each of multiple evaluation scales using a statistical method capable of associating a human quality evaluation judgment with a nonlinear relationship between multiple features, and a quality evaluation result Set the output speech unit sequence determination unit that determines whether or not to adopt the speech unit sequence and the weight set that reflects the evaluation result obtained by the quality evaluation unit when the quality is bad, and set the next accent A process for setting a weight set to an initial value for performing a segment search for a phrase, a weight set setting / setting changing unit for performing a setting value changing process, and an input text end determination for determining whether the input text has been searched to the end This is realized by constructing a synthesis processing unit that smoothly connects selected speech element sequences by using a signal processing method and the like.
With this configuration, it is possible to obtain the unique operational effect of the present invention that a high-quality synthesized speech can be stably obtained.

FIG. 1 shows the overall configuration of a speech synthesizer 100 according to the present invention. The speech synthesis method according to the present invention is basically a waveform-connected speech synthesis method, and a speech synthesizer 100 that operates according to this waveform-connected speech synthesis method includes a text analysis processing unit 10, a prosody generation unit 20, a speech segment. The search unit 30, the synthesis processing unit 40, and the voice database 50 are configured. The voice database 50 includes a voice corpus that holds a large amount of voice data.
The flow of processing in the speech synthesizer 100 is that an accent type, which is information necessary for prosody generation processing, is obtained by performing a morphological analysis on the input text TX that is the target of speech synthesis processing by the text analysis processing unit 10. -Tone combination type is given.

The prosody generation unit 20 uses these pieces of information to generate a target prosody pattern. The speech segment search unit searches the speech database 50 for a speech segment sequence in which the prosodic sequence matches the input text and the obtained prosodic pattern and the speech environment of the speech segment match as much as possible. At 30. The speech unit series selected last is smoothly connected in the synthesis processing unit 40 to generate synthesized speech data OD.
The processing flow up to this point is equivalent to the conventional speech synthesis method.
The feature of the present invention is the processing in the speech unit search unit 30, and particularly the processing in the quality evaluation unit 33 provided in the speech unit search unit 30.

FIG. 2 shows the internal configuration of the speech segment search unit 30, and FIG. 7 shows the processing flow. Since the text analysis processing unit 10, the prosody generation unit 20, the synthesis processing unit 40, and the speech database 50 other than these are not different from those in the past, detailed description thereof is omitted.
The speech unit search unit 30 includes a weight set setting / setting change unit 31, a speech unit sequence search unit 32, a quality evaluation unit 33, an output speech unit sequence determination unit 34, and an input text end determination unit 35. And is configured. The target prosody pattern and the input text information ST output from the prosody generation unit 20 are input to the weight set setting and setting change unit 31, and the speech unit is output from the input text end determination unit 35 as the output of the speech unit search unit 30. Series data OS is output.

The flow of processing in the speech segment search unit 30 is as close as possible to the target prosody pattern generated by the prosody generation unit 20 as in the conventional method, and includes phoneme sequences obtained from the text analysis processing unit 10. The speech unit sequence search unit 32 performs a process of searching the speech database 50 for a speech unit sequence that matches the input text information TX.
The selection of a speech segment sequence as close as possible to the target prosody pattern is as follows. As shown in equation (1), sub-costs related to a plurality of parameters such as fundamental frequency, phoneme environment, phoneme duration, number of segment connections, etc. For example, with respect to the fundamental frequency, the degree of coincidence of the fundamental frequency pattern, the gap value of the fundamental frequency at the connection point of the speech unit is calculated as a cost, and the total cost obtained by weighting and adding them is evaluated. The initial value of the weight set used here is a value tuned in advance so that high-quality synthesized speech can be obtained on average for various input texts. The initial value of the weight set is given by the weight set setting / setting changing unit 31 before searching for speech segments.

Next, the quality evaluation unit 33 evaluates the overall quality of the synthesized speech generated from the speech element sequence selected once. The scale of quality to be evaluated (quality evaluation scale) is not limited to one, but includes a scale related to the fundamental frequency deeply related to intonation and a scale related to phoneme duration related to speech rate.
The unit for evaluating quality is preferably as long as possible in the present invention for the purpose of evaluating global quality. However, if the unit is too long, there are too many parameters to be considered when evaluating the quality, and it is considered that the evaluation becomes difficult. Therefore, it is necessary to evaluate the quality of a phoneme segment sequence having an appropriate length.

When evaluating quality using measures of intonation and utterance speed, these measures are elements that represent prosody, so in the following example, the accent phrase that is a prosodic unit is used as the quality evaluation unit. The case will be described.
In addition, the quality of the synthesized speech, taking into account the spectrum that is deeply related to the real voice, the continuity of the fundamental frequency, and the measure that matches the prosodic environment of the speech unit that is deeply related to the clarity at the connection point of the speech unit. Make an evaluation.
From the above, the quality evaluation unit 33 performs quality evaluation using a statistical method for a plurality of evaluation scales for each accent phrase unit. An example of a statistical method is machine learning. In this specification, a case where SVM (Support Vector Machine) is used among a number of machine learners will be described as an example.

  SVM uses a large number of samples as learning data, but learns the relationship between a class assigned to a certain learning sample and the multidimensional feature value extracted from that sample, and where is the class in the multidimensional feature space. This is a type of pattern recognition technique for estimating which class belongs when a certain feature value is input by associating whether the feature is assigned. The SVM can be expressed as shown in equation (2) and as shown in FIG. 6 in the figure. Below, the case where it aims at classifying into two classes is described.

２つのクラスA、Bがあるとし、f(x⁻)の値はクラスAの場合に正と、クラスＢの場合は負と、クラスの境界では０となるように、学習サンプルからw⁻，ｂの値を学習する。これらの値を学習することで、特徴空間上のクラスの境界を推定する。この推定は、境界と学習サンプルの最小距離が最大となるように、つまり２つのクラスの真ん中に境界が位置するように行われる（図６）。
以下では，基本周波数に関する品質評価のためのSVMの作成について具体例を示す。学習データには、あらかじめ作成しておいた合成音声や人間の生の音声などを使うことができる。基本周波数に関するSVMの学習データのクラスは、イントネーションが正しいか、間違っているかのどちらかをクラスとして与える。このクラスの付与は人手を用いて行うが、発声速度や明瞭度など他の要因は一切気にせず、イントネーションだけに注目した評価をアクセント句単位で行い、その結果得られる正誤をクラスとして付与する。

Two classes A, consider a B, f (x ^-) and positive in the case of values Class A, and negative in the case of class B, as becomes zero at the boundary of the class, w from learning samples ^-, Learn the value of b. By learning these values, class boundaries on the feature space are estimated. This estimation is performed so that the minimum distance between the boundary and the learning sample is maximized, that is, the boundary is located in the middle of the two classes (FIG. 6).
In the following, a specific example is shown for creating an SVM for quality evaluation of the fundamental frequency. For the learning data, synthesized speech or human live speech prepared in advance can be used. The SVM learning data class related to the fundamental frequency is given as a class indicating whether the intonation is correct or incorrect. This class is assigned manually, but we do not care about other factors such as utterance speed and intelligibility at all, and we give an evaluation focusing on only intonation on an accent phrase basis. .

  On the other hand, with respect to the feature quantity that can be used to determine whether the intonation is correct or not, the quality can be evaluated with higher accuracy by using not only the local quality evaluation parameter but also the global parameter. An example of quality evaluation parameters is given below. As shown in FIG. 3, a regression line is drawn for each mora for each of the target prosody pattern and the synthesized speech prosody pattern, and the slope, that is, the regression coefficient is obtained to obtain a series of regression coefficients corresponding to the number of mora. The correlation coefficient of the obtained two sequences is obtained and the value is used as a parameter. Similarly, the correlation coefficient is obtained for the Δ component of the regression coefficient sequence and used as a parameter. Similarly, the correlation coefficient is obtained for the Δ component of the regression coefficient sequence and used as a parameter. In addition, the difference between the regression coefficients of the mora with the accent kernel and the next mora, the gap of the fundamental frequency between the mora, and the like can be mentioned.

When a speech unit selection result is input, an SVM that can determine whether the intonation of the synthesized speech created from the speech unit is correct or incorrect is defined as the above feature quantity and class ( It is created in advance using a large amount of learning data associated with correctness of intonation.
In the quality evaluation unit 33 (FIG. 2) that evaluates the quality of the synthesized speech using the SVM, first, the speech unit sequence search unit 32 (FIG. 3) obtains the feature quantity necessary for the input of the SVM related to intonation. Extracted from the selected speech segment selection result. Next, by inputting the feature amount into the SVM, quality of the synthesized speech intonation obtained from the speech unit selection result can be obtained.

As in the case of speech segment search, a method may be considered in which feature values such as those described above are individually evaluated, and the total cost obtained by linearly summing the evaluation values is used to determine whether the intonation is correct or incorrect. However, since the SVM can associate a non-linear relationship between a plurality of feature amounts and correctness / incorrectness determination by a person, it is possible to perform correctness / incorrectness determination with higher accuracy than the total cost of a simple linear sum.
In addition to intonation obtained from the fundamental frequency, quality evaluation using the SVM is obtained from the utterance speed obtained from the phoneme duration, the continuity obtained from the fundamental frequency and spectrum gap values at the unit connection points, and the coincidence of the phoneme environment The same shall be done for clarity.

  An example of the quality evaluation unit 33 using a plurality of quality evaluation scales is shown in FIG. In this example, a case is described in which quality evaluation is performed for three types of evaluation scales: quality evaluation related to intonation, quality evaluation related to speech rate, and quality evaluation related to continuity, but the number of evaluation scales is limited to three. Not a thing. Prior to performing each quality evaluation, feature quantity extraction means 33A-1, 33B-1, and 33C-1 are provided, and each of these feature quantity extraction means 33A-1, 33B-1, and 33C-1 has a quality related to intonation. In the evaluation, a correlation function of a regression coefficient, a correlation function of a Δ regression coefficient, a difference of regression coefficients in the nucleus, and the like are extracted as feature quantities. In addition, in the quality evaluation related to speech rate, the duration of each mora, the vowel duration of each mora, the average duration of each vowel, etc. are extracted as features, and in the quality evaluation related to continuity, the gap value and spectrum of the fundamental frequency are extracted. Are extracted as feature values, and evaluation is performed by each of the SVMs 33A-2, 33B-2, and 33C-2 using these feature values. In addition, the case where the quality evaluation result which SVM33A-2, 33B-2, 33C-2 regarding each quality evaluation scale outputs is good is shown as ◯, and the case where it is bad is shown as x.

The output phoneme sequence determination unit 34 determines from the quality evaluation result obtained by the quality evaluation unit 33 whether to use the selected speech unit or discard it and reselect the speech unit.
If there is a measure that is determined to be of poor quality among all quality evaluation measures in an accent phrase, the output speech segment sequence determination unit 34 re-emphasizes the weight coefficient of the segment search parameter corresponding to that measure. Shift processing to selection. For example, when the SVM related to intonation determines that the quality is poor, a certain amount is added to the weight count of the segment search parameter related to the fundamental frequency, such as 50% of the initial value. Alternatively, the weighting factor is changed by dynamically determining how much the weighting factor is to be added from the value output by the SVM, that is, the value of f (x ⁻ ) expressed by the equation (2). The value of f (x ⁻ ) is a value indicating a class with poor quality, and if the absolute value is large, it means that it is far from a class with good quality. If it is small, it is close to a class with good quality, so the addition amount of the weight coefficient is also reduced.

These processes are performed by the weight setting units 31-1, 31-2, and 31-3 of the weight set setting / setting changing unit 31 (FIG. 5). Then, the speech unit coefficient search unit 32 performs re-selection of the unit, and the quality evaluation unit 33 performs the process of determining the quality again until the quality is improved. Even if there are a plurality of measures that are judged to have poor quality, all the weighting factors corresponding to the measures that are judged to be bad are changed. If it is determined that the quality is poor even if the reselection is repeated a predetermined number of reselections, it is determined to output the speech segment sequence searched with the initial weight set.
In the quality evaluation unit 33, when all the quality evaluation SVMs 33A-2, 33B-2, and 33C-2 determine that the quality of the accent phrase is good, the output speech unit sequence determination unit 34 determines the voice that expresses the accent phrase. It is determined that the segment sequence is output.

Since the quality evaluation is performed in units of accent phrases, the speech unit search in the speech unit sequence search unit 32, the quality evaluation in the quality evaluation unit 33, and the output speech unit sequence in the output speech unit sequence determination unit 34 so far are performed. The determination, weight set setting, and change of the weight set in the setting change unit 31 are performed for each accent phrase from the beginning of the input text.
This series of processing is performed for each accent phrase up to the end of the input text. Whether or not the input text has been completed is determined by the input text end determination unit 35. If the input text has not been completed, the weight set setting and setting change unit 31 performs weighting. The set is set to the initial value (FIG. 5), and the next accent phrase segment search is performed. If it is determined that the processing has been completed, the synthesis processing unit 40 (FIG. 1) smoothly connects adjacent speech units of the speech unit series used as synthesized speech and outputs the synthesized speech data.

FIG. 7 shows a flowchart for explaining the operation of the speech element search unit 30 described above. FIG. 7 shows a basic processing flow of the speech segment search unit 30 used in the speech synthesis method of the present invention.
By the text analysis process of the input text, the number N of accent phrases of the input text is calculated, and the reselection allowable number S is set (step S7-1).
The input text end determination unit 35 determines whether the processing has been completed for all accent phrases (step S7-2).
The nth accent phrase is set as a processing target (step S7-3).

The weight set of the weight set setting / setting changing unit 31 is set to an initial value (step S7-4).
1 is substituted into the reselection allowable number of times S (step S7-5).
The speech unit sequence search unit 32 searches for speech unit sequences (step S7-6).
The quality evaluation unit 33 performs quality evaluation for each of the plurality of scales, and determines whether “quality is good” or “quality is bad” (step S7-7).
It is determined whether or not there is a scale determined by the output speech segment series determination unit 34 as having poor quality (step S7-8).

If the determination result in step S7-8 is NO, the speech unit sequence searched at that time is determined as the “speech unit sequence” of the nth accent phrase (step S7-11).
The number of accent phrases n is incremented by 1, and the process returns to step S7-2 (step S7-12).
It is determined whether or not the processing has been completed for all accent phrases (step S7-2), and if completed, the process branches to step S7-13, and “speech segment series” for each accent phrase constituting the input text. Is output to the composition processing unit 40 (FIG. 1).

If the processing has not been completed for all accent phrases, steps S7-3 to S7-10 are repeated to search for the speech unit sequence of the next accent phrase.
If it is determined in step S7-8 that there is an evaluation scale with poor quality, the weight set setting value is changed in weight set setting changing step S7-9, and the count value of the reselection allowable number of times S is incremented by +1 (step S7- 10), the process returns to step S7-6, and the search for the speech segment sequence of the accent phrase is continued.

  8 and 9 show the processing flow of the speech synthesizer that implements the speech synthesis method proposed in claim 2 of the present invention. In this embodiment, as an exception process, the synthesis processing unit 40 manipulates the fundamental frequency of the speech waveform and the phoneme duration. For example, signal processing as disclosed in Patent Document 3 (Japanese Patent No. 3557124) is performed. When only synthesized speech with low quality can be obtained simply by considering the combination of speech units, it is possible to generate synthesized speech with higher quality than when no signal processing is performed by performing such signal processing. Whether or not to perform signal processing is determined by the output speech segment sequence determination unit 34 based on whether reselection has been repeated a predetermined number of times, and the determination result is transmitted to the synthesis processing unit 40.

That is, in the processing flow shown in FIGS. 8 and 9, steps S7-1 to S7-13 correspond to the steps shown in FIG. 7, but in the second embodiment, steps S8-1 and S8- shown in FIG. 2 and S8-3 are provided.
Step S8-1 shows a reselection allowable number determination step. In this reselection allowable number determination step S8-1, the number of executions of the speech element sequence search step S7-6 and the quality evaluation step S7-7 is the reselection allowable number of times. It is determined whether or not S has been reached. If it is N0, the routine returns to step S7-6 via weight set setting change step S7-9, and the speech unit sequence search is continued.

On the other hand, if there is no measure that is determined to be of poor quality in the output speech segment sequence determination step S7-8, it is determined in step S8-2 that the nth accent phrase currently being processed does not require signal processing, In step S7-11, the speech unit sequence searched at that time is determined as the speech unit sequence of the nth accent phrase, and the speech unit sequence is added to the synthesis processing unit 40 together with information that does not require signal processing. Output.
When it is determined that the number of executions of steps S7-7 and S7-8 has reached the reselection allowable number S in the reselection allowable number determination step S8-1 even though all the quality evaluation results want to reach good. The process branches to step S8-3. In step S8-3, the nth accent phrase currently being processed is determined to require signal processing, and the speech unit sequence searched at that time is output to the synthesis processing unit 40 together with information that requires signal processing. To do.

That is, the processing flow shown in FIGS. 8 and 9 is a processing flow in which the signal processing proposed in Patent Document 3 described above is added to the processing flow shown in FIG. An effect is obtained in which adjacent speech elements of the speech element series to be connected are smoothly connected.
In the above-described speech synthesis method according to the present invention and the speech synthesizer operating using this speech synthesis method, the embodiment in which the speech synthesis program proposed in the present invention is installed in a computer and realized by the computer is the most desirable embodiment. . The speech synthesis program proposed in the present invention is written in a computer-readable program language, recorded on a recording medium such as a magnetic disk or CD-ROM that can be read by the computer, and from these recording media to a computer through a communication line. It is installed, decoded by a CPU provided in the computer, and executed.

  The outline of the program is composed of routines for constructing the constituent elements 10 to 40 shown in FIG. 1, and particularly the feature of the present invention is the construction of the speech segment search unit 30 shown in FIG. Further details are present in the processing procedure shown in FIG. 7 or FIG. 8 and FIG.

  The voice synthesizing method and apparatus according to the present invention is used as a voice guide device for various devices.

1 is a block diagram for explaining an outline of the present invention. The block diagram for demonstrating the detail of the speech unit search part shown in FIG. The figure which showed the regression line of the target prosody and the regression line of the synthetic speech prosody. The block diagram for demonstrating the detail of the quality evaluation part which can be used for the speech unit search part shown in FIG. The block diagram for demonstrating an example of a structure of the weight set setting part shown in FIG. 5 is a flowchart for explaining an example of a processing flow of a speech element search unit shown in FIG. 4. The flowchart for demonstrating the example which added the signal type | system | group processing function to the processing flow shown in FIG. 8 is a flowchart for explaining an example in which a signal processing function is added to the processing flow shown in FIG. 9 is a flowchart for explaining the continuation of the processing flow shown in FIG.

Explanation of symbols

TX input text 32 speech unit sequence search unit OD synthesized speech data 33 quality evaluation unit 10 text analysis processing unit 34 output speech unit sequence determination unit 20 prosody generation unit 35 input text end determination unit 30 speech unit search unit 40 synthesis processing Unit 31 Weight set setting and setting change unit 50 Voice database

Claims (10)

It has a database that stores speech corpora holding speech data, and in order to convert any input text into speech, input text and content can be searched and connected by searching for speech segments of any length included in the speech corpus. In a speech synthesis method for generating synthesized speech in which
A text analysis processing step for obtaining a reading by morphological analysis of the input text and obtaining the number of accent phrases and tone combination information necessary for generating a prosodic pattern;
A prosodic pattern generation step for generating a target prosodic pattern using the information obtained in the text analysis processing step;
A processing target accent phrase setting step for setting a processing target accent phrase according to a target prosodic pattern;
A weight set setting and setting change step for performing a process of setting the weight set to an initial value for each accent phrase and a process of changing the set value of the weight set value;
A speech segment search step of selecting a speech segment based on the prosodic pattern generated in the prosodic pattern generation step and setting a speech segment sequence;
A quality evaluation step of performing a quality evaluation of the synthesized speech for each of a plurality of evaluation scales with respect to the speech unit sequence searched in the speech unit search step;
The presence / absence of an evaluation scale determined to be poor in the evaluation result of the quality evaluation step is detected, and if the detection result is present, the above speech unit series is subjected to the weight set setting and the setting change process of the setting change step. Processing to return to the search step, and if not, an output speech unit sequence determination step for determining the speech unit sequence searched at that time as the speech unit sequence of the processing target accent phrase,
When the speech unit sequence for each accent phrase is determined in this output speech unit sequence determination step, it is determined whether or not the processing has been completed for all the accent phrases included in the input text. Outputs the speech segment series for each accent phrase constituting the input text to the synthesized speech processing step, and if it is determined that it has not been completed, sets the next accent phrase as the processing target accent phrase, and sets and sets the above weight set A change step, a speech segment sequence search step, a quality evaluation step, an output speech segment sequence determination step, an input text end determination step that repeatedly executes a preset reselection allowable number of times,
A speech synthesis method comprising: 2. The speech synthesis method according to claim 1, wherein when the evaluation result of the quality evaluation step includes an evaluation scale determined to be of poor quality, the number of executions of the speech unit sequence search step and the quality evaluation step is the re-execution number. It is determined whether or not the allowable number of selections has been reached, and if not achieved, a reselection allowable number of times determination step that returns to the speech segment sequence search step through the weight set setting change step,
If it is determined that this reselection allowable number determination step is achieved, the speech unit sequence currently searched for along with the information that requires signal processing for the accent phrase currently being processed is currently processed. A signal system processing necessity determination step for determining the speech unit sequence of the accent phrase being
If the quality evaluation step determines that the evaluation scale is poor, the speech being searched at that time together with information that does not require signal processing for the accent phrase currently being processed A signal processing unnecessary determination step for determining a unit sequence as a speech unit sequence of the accent phrase;
A speech synthesis method characterized by adding   3. The speech synthesis method according to claim 1, wherein the evaluation scale used in the quality evaluation step is at least one of an evaluation scale related to intonation, an evaluation scale related to speech rate, or an evaluation scale related to continuity. A speech synthesis method comprising:   4. The speech synthesis method according to claim 1, wherein the weight set set in the weight set setting and setting change step includes at least a parameter weight related to a fundamental frequency, a parameter weight related to a duration, and a connection. A speech synthesis method comprising any combination of parameter weights for point gaps. It has a database that stores speech corpora holding speech data, and in order to convert any input text into speech, input text and content can be searched and connected by searching for speech segments of any length included in the speech corpus. In a speech synthesizer that generates synthesized speech in which
A text analysis processing unit that obtains readings by morphological analysis of input text and obtains the number of accent phrases and tone combination information necessary for prosody pattern generation;
Using the information obtained by the text analysis processing unit, a prosody pattern generation unit that generates a target prosody pattern,
A processing target accent phrase setting unit for setting an accent phrase to be processed according to a target prosodic pattern;
A weight set setting and setting change unit for performing a process of setting a weight set to an initial value for each accent phrase and a process of changing a weight set value;
A speech segment search unit that selects a speech unit based on the prosodic pattern generated by the prosody pattern generation unit;
A quality evaluation unit that evaluates the quality of the synthesized speech for each of a plurality of evaluation scales with respect to the speech unit sequence searched by the speech unit search unit;
The presence / absence of an evaluation scale determined to be poor in the evaluation result of the quality evaluation unit is detected, and if the detection result is present, the speech unit sequence is subjected to the weight set setting and the setting change processing of the setting change unit. Return to the speech unit sequence search process in the search unit, if not, an output speech unit sequence determination unit that determines the speech unit sequence searched at that time as the speech unit sequence of the accent phrase to be processed; ,
When the output speech segment sequence determination unit determines whether or not the processing has been completed for all accent phrases included in the input text each time the speech segment sequence for each accent phrase is determined, Outputs a speech segment sequence for each accent phrase constituting the input text to the synthesized speech processing unit, and when it is determined that the input is not completed, sets the next accent phrase as a processing target accent phrase, and the weight set setting unit A speech unit sequence search unit, a quality evaluation unit, and an input text end determination unit that repeatedly operates the output speech unit sequence determination unit for a preset number of reselection allowed,
A speech synthesizer comprising: 6. The speech synthesizer according to claim 5, wherein when the evaluation result of the quality evaluation unit includes an evaluation measure that is determined to be of poor quality, the number of executions of the speech unit sequence search unit and the quality evaluation unit is re-executed. It is determined whether or not the allowable number of selections has been reached, and if not achieved, the reselection allowable number determination is made to return to the search operation of the speech unit sequence search unit through the weight set setting change process of the weight set setting change unit. And
If this reselection allowable number determination unit determines that the achievement is achieved, for the accent phrase currently being processed, the current speech unit sequence searched for at that time together with information requiring signal processing is currently processed. A signal processing necessity determination unit for determining a speech unit sequence of an accent phrase being
When the quality evaluation unit determines that the quality measure is determined to be poor, the speech being searched at that time together with information that does not require signal processing for the accent phrase currently being processed A signal processing unnecessary determination unit for determining a unit sequence as a speech unit sequence of the accent phrase;
A speech synthesizer characterized by the addition of   7. The speech synthesizer according to claim 5, wherein the evaluation unit provided in the quality evaluation unit includes at least an evaluation unit for intonation of a speech unit sequence determined based on a prosodic pattern generated from input text. A speech synthesizer characterized by including either an evaluation means for speech rate or an evaluation means for continuity.   8. The speech synthesizer according to claim 5, wherein the weight set set by the weight set setting and setting change unit is a speech unit sequence determined based on at least a prosodic pattern generated from input text. A speech synthesizer comprising a combination of a parameter weight related to a fundamental frequency, a parameter weight related to a duration, and a parameter weight related to a gap at a connection point.   9. A speech synthesis program that is written in a computer-readable program language and causes the computer to function as the speech synthesizer according to claim 5.   A recording medium comprising a computer-readable recording medium, wherein at least the speech synthesis program according to claim 9 is recorded on the recording medium.

Images