JP2010145794A - Voice synthesis dictionary construction device, voice synthesis dictionary construction method, and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To generate a voice synthesis dictionary deleted with a pitch portion series data having significant pitch fluctuation, thereby to prevent a synthesized voice from getting unnatural. <P>SOLUTION: A pitch series data extracting part 220 extracts a pitch series data from a voice data taken out from a voice database 201. A voice cut-out part 225 cuts out the pitch portion series data in each phoneme from the extracted pitch series data. A pitch data evaluating part 230 determines whether the pitch fluctuation of each pitch portion series data is significant or not. A pitch data editing part 240 deletes the pitch portion series data having the significant pitch fluctuation. A phoneme HMM learning part 250 finds thereafter a correspondence relation between a phoneme label and a phoneme pitch data. A data writing-out part 260 writes the correspondence relation between the phoneme label and the phoneme pitch data after learning, into the voice synthesis dictionary 202. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a speech synthesis dictionary construction device, a speech synthesis dictionary construction method, and a program for constructing a database used for speech synthesis.

  Speech recognition technology and speech synthesis technology based on a hidden Markov model (hereinafter referred to as HMM) are widely used as speech recognition and speech synthesis technology.

  Speech recognition technology and speech synthesis technology based on HMM are disclosed in Patent Documents 1 and 2, for example.

  In speech synthesis based on the HMM, a speech synthesis dictionary in which a correspondence relationship between phoneme labels and spectrum parameter data strings is recorded is required.

  A speech synthesis dictionary usually performs spectrum analysis and pitch extraction on data recorded in a database (hereinafter referred to as a speech database) composed of a set of phoneme label sequences and corresponding speech data. It is constructed through a learning process based on HMM.

  Conventionally, when constructing a speech synthesis dictionary, even if pitches extracted from speech data include those that are significantly different from other pitches, they are used as they are for learning based on the HMM.

Japanese Patent Laid-Open No. 2002-244689 JP 2002-268660 A

  However, the synthesized speech generated using the speech synthesis dictionary constructed as described above includes a portion where the pitch variation is steep.

  For this reason, when a speech synthesis dictionary is constructed with a conventional speech synthesis dictionary construction device, the synthesized speech may be unnatural compared to natural human speech.

  The present invention has been made in view of the above circumstances, and a speech synthesis dictionary construction apparatus and speech synthesis dictionary construction method capable of constructing a speech synthesis dictionary that does not synthesize unnatural speech in which rapid pitch fluctuation occurs. And to provide a program.

In order to achieve the above object, a speech synthesis dictionary construction device according to the first aspect of the present invention provides:
An apparatus for constructing a speech synthesis dictionary,
A receiver for receiving a phoneme label string and corresponding voice data;
A pitch sequence data extraction unit for extracting pitch sequence data from the received audio data;
A phoneme cutout unit that cuts out pitch partial sequence data for each phoneme from the extracted pitch sequence data;
A pitch data evaluation unit that evaluates whether the extracted pitch partial series data is suitable for constructing a speech synthesis dictionary;
A pitch data editing unit for editing pitch partial series data evaluated to be suitable for building a speech synthesis dictionary;
A phoneme HMM learning unit that associates a phoneme HMM related to a pitch with each phoneme label by learning based on a hidden Markov model from the phoneme label sequence and the edited pitch subsequence data;
A data writer for recording the learning results in the speech synthesis dictionary;
It is characterized by providing.

In the speech synthesis dictionary construction device according to the first aspect,
The pitch data evaluation unit may evaluate the combination of the mel cepstrum partial sequence data corresponding to the pitch partial sequence data evaluated to be unsuitable for constructing a speech synthesis dictionary and not suitable for constructing a speech synthesis dictionary. Good.

  It is desirable that the pitch data evaluation unit evaluates that it is suitable for building a speech synthesis dictionary when the difference between the pitch partial series data and the average value does not exceed a predetermined threshold value in all frames.

  The pitch data evaluation unit is suitable for constructing a speech synthesis dictionary when a difference between an average value of pitch partial series data and an average value of pitch partial series data belonging to the same monophone label does not exceed a predetermined threshold. You may make it evaluate.

In the speech synthesis dictionary construction device according to the first aspect,
The pitch partial series data may be deleted, and the pitch series data cut out from the pitch partial series data may be deleted.

In addition to this, when deleting the mel cepstrum partial series data corresponding to the deleted pitch partial series data,
The mel cepstrum partial sequence data may be deleted and the mel cepstrum sequence data obtained by cutting out the mel cepstrum partial sequence data may be deleted.

In the speech synthesis dictionary construction device according to the first aspect,
The number of audio data from which the pitch partial series data has been deleted may be counted, and if the number exceeds a predetermined threshold, all audio data may be collected again.

In order to achieve the above object, a speech synthesis dictionary construction method according to a second aspect of the present invention includes:
A method for building a speech synthesis dictionary,
A receiving step of receiving a phoneme label string and corresponding voice data from the database;
A pitch sequence data extraction step for extracting pitch sequence data from the received audio data;
A phoneme extraction step of extracting pitch partial sequence data for each phoneme from the extracted pitch sequence data;
A pitch data evaluation step for evaluating whether the extracted pitch subsequence data is suitable for constructing a speech synthesis dictionary;
A pitch data editing step for editing pitch subsequence data evaluated to be suitable for building a speech synthesis dictionary;
A phoneme HMM learning step in which a phoneme HMM associated with a pitch is associated with each phoneme label by learning based on a hidden Markov model from the phoneme label sequence and the edited pitch subsequence data;
An output step for outputting the learning result;
It is comprised from these.

In order to achieve the above object, a program according to the third aspect of the present invention provides:
On the computer,
A receiving step of receiving a phoneme label string and corresponding voice data from the database;
A pitch sequence data extraction step for extracting pitch sequence data from the received audio data;
A phoneme extraction step of extracting pitch partial sequence data for each phoneme from the extracted pitch sequence data;
A pitch data evaluation step for evaluating whether the extracted pitch subsequence data is suitable for constructing a speech synthesis dictionary;
A pitch data editing step for editing pitch subsequence data evaluated to be suitable for building a speech synthesis dictionary;
A phoneme HMM learning step in which a phoneme HMM associated with a pitch is associated with each phoneme label by learning based on a hidden Markov model from the phoneme label sequence and the edited pitch subsequence data;
An output step for outputting the learning result;
Is executed.

  According to the present invention, from the pitch partial series data extracted from the voice data, those having a remarkable pitch variation compared to other pitch partial series data are deleted and used for learning the voice synthesis dictionary. For this reason, the synthesized speech obtained by using the speech synthesis dictionary can be of high quality without causing an unnatural speech part without including a sudden pitch fluctuation.

Hereinafter, the speech synthesis dictionary construction device according to the embodiment of the present invention will be described in detail.
(Embodiment 1)
First, the configuration of the speech synthesis dictionary construction device according to the present embodiment will be described.
FIG. 1 is a diagram illustrating a configuration of the speech synthesis dictionary construction device according to the first embodiment.

A speech synthesis dictionary construction apparatus 100 shown in FIG. 1 includes a ROM 101, a storage unit 102, a data input / output interface (I / F) 103, a user interface (I / F) 104, and a CPU 105. Are connected to each other.
The ROM 101 stores an operation program for learning based on the HMM, in particular, in this embodiment, an operation program including operations for evaluating and editing the pitch subsequence data.

  The storage unit 102 includes a RAM 121 and a hard disk 122, and includes constants for learning, phoneme label strings, speech data, pitch series data, pitch partial series data, data phoneme labels and phoneme pitch information associated with each other. Remember.

  The data input / output I / F 103 is an interface for connecting to the original data-containing hard disk 111, the processed data recording hard disk 112, and the like. The original data-containing hard disk 111 corresponds to the speech database 201 shown in FIG. 2, and the processed data recording hard disk 112 corresponds to the speech synthesis dictionary 202 shown in FIG.

  The data input / output I / F 103 is connected to the speech database 201 shown in FIG. 2, reads out a pair of phoneme label strings to be learned and speech data under the control of the CPU 105, and stores them in the storage unit 102.

  The data input / output I / F 103 is connected to the speech synthesis dictionary 202 shown in FIG. 2, and outputs a correspondence relationship between the phoneme label and the phoneme pitch information, which is a result of the processing by the CPU 105, to the speech synthesis dictionary 202.

  The user I / F 104 includes a keyboard (KB) 141 and a display (DSP) 142, and is provided for inputting arbitrary instructions, data, and programs. In particular, in the pitch evaluation / editing process, the user needs to give a predetermined constant through the I / F.

  The CPU 105 executes the operation program stored in the ROM 101, thereby realizing each unit shown in the functional configuration of FIG.

FIG. 2 is a functional configuration diagram of the speech synthesis dictionary construction device 100 shown in FIG.
A speech synthesis dictionary construction apparatus 100 shown in FIG. 1 is connected to a speech database 201 and a speech synthesis dictionary 202 as shown in FIG.

  The voice database 201 is a database composed of a set of phoneme label strings and corresponding voice data, and is stored in a hard disk or the like.

  The speech synthesis dictionary 202 is a database constructed by the speech synthesis dictionary construction apparatus 100 shown in FIG. 1, stores phoneme labels and phoneme learning results in association with each other, and is stored in a hard disk or the like.

  The phoneme learning result includes phoneme pitch information. Other spectral information necessary for speech synthesis varies depending on the specifications of the speech synthesizer, and the phoneme learning result includes such various information.

  Functionally, the speech synthesis dictionary construction apparatus 100 shown in FIG. 1 functionally has a data extraction unit 210, a pitch sequence data extraction unit 220, a phoneme segmentation unit 225, a pitch data evaluation unit 230, as shown in FIG. A pitch data editing unit 240, a phoneme HMM learning unit 250, and a data writing unit 260.

  The data extraction unit 210 reads data from the voice database 201 and separates it into a phoneme label string and voice data. The phoneme label string is delivered to the phoneme HMM learning unit 250, and the speech data is delivered to the pitch sequence data extraction unit 220.

  The pitch sequence data extraction unit 220 extracts predetermined pitch sequence data from the audio data delivered from the data extraction unit 210 and delivers it to the phoneme segmentation unit 225.

  The phoneme cutout unit 225 cuts out the pitch partial sequence data for each phoneme from the pitch sequence data delivered from the pitch sequence data extraction unit 220 and delivers it to the pitch data evaluation unit 230.

  The pitch data evaluation unit 230 evaluates whether each pitch partial series data delivered from the phoneme extraction unit 225 is suitable for constructing a speech synthesis dictionary. Details of this evaluation process will be described later with reference to FIGS. 3 and 4.

  The pitch data editing unit 240 performs a predetermined editing process on the pitch partial series data delivered from the pitch data evaluation unit 230 to generate edited pitch partial series data. Details of the predetermined editing process will be described later with reference to FIGS.

  The edited pitch partial series data is delivered to the phoneme HMM learning unit 250.

  The phoneme HMM learning unit 250 converts the correspondence relationship between the phoneme label string and the edited pitch partial series data into a correspondence relationship between the phoneme label and the phoneme pitch information (phoneme HMM) by learning based on the HMM. The data is written to the data writing unit 260.

  The data writing unit 260 records the correspondence relationship between the phoneme label and the phoneme pitch information in the speech synthesis dictionary 202.

  The predetermined evaluation process executed by the pitch data evaluation unit 230 may be any process as long as the process evaluates whether the pitch data is suitable for constructing a speech synthesis dictionary. An example will be described.

  In the following description, a frame means a time segment used for pitch extraction and is represented by the symbol fm.

(Specific example 1 of evaluation / editing process)
A specific example 1 of the evaluation / editing process will be described with reference to the flowchart shown in FIG.

  In this specific example, it is assumed that the user previously sets the threshold value thpit of the pitch difference in the storage unit 102 via the user I / F 104 in FIG.

  When the speech synthesis dictionary construction apparatus 100 shown in FIG. 1 constructs the speech synthesis dictionary 202, the speech synthesis dictionary construction apparatus 100 is connected to the speech database 201 and, for example, an empty speech synthesis dictionary 202. .

When an instruction to start speech synthesis dictionary 202 generated is from the user I / F 104, the data extraction unit 210, from the speech database 201, a phoneme label sequence MnLab _m and voice data Sp _m (however, 1 ≦ m ≦ _{M SP} Yes, _MSP is the number of data in the speech database.) The pairs are sequentially read out and stored in the storage unit 102.

Pitch series data extraction unit 220, the voice data Sp _m from the pitch-series data Pit _m [fm] (where a _{0 ≦ fm ≦ FLab m, FLab} m is the number of frames of the audio data Sp _m.) Extract And it memorize | stores in the memory | storage part 102 (step S1).

Then, the phoneme cutout unit 225 performs phoneme cutout on the pitch sequence data Pit _m using the phoneme label string MnLab _m to generate pitch partial sequence data pit _{L, n} (step S2). However, a _{0 ≦ L ≦ M L, M} L is a phoneme label total. Further, 0 ≦ n ≦ N _L , and N _L is the total number of pitch partial series data for the phoneme label.

The pitch data evaluation unit 230 calculates the pitch partial sequence data group pit _{L, n} [fm] (where 0 ≦ fm ≦ F _{L, n} , where F _{L, n} is the total number of frames for the pitch partial sequence data pit _{L, n.} For each of the pitch partial series data pit _{L, n} , an average value ave _{L, n} is calculated according to the following equation (step S3).
ave _{L, n} = {1 / (F _{L, n} + 1)} · Σ _{fm = 0} ^{FL, n} pit _{L, n} [fm]

Then, according to the following steps, the difference between the pitch partial series data pit _{L, n} and the average value ave _{L, n} is calculated for each phoneme label and for each pitch partial series data pit _{L, n.} The series data pit _{L, n} is evaluated.

  First, a pointer L for designating a number for identifying a phoneme label is initialized to “1” (step S4).

  Subsequently, a pointer n for designating a number for identifying each pitch partial series data is initialized to “1” (step S5).

With these pointers L, n, attention is paid to the pitch partial series data group pit _{L, n} [fm] (where 0 ≦ fm ≦ F _{L, n} ) on the storage unit 102.

  Then, the pointer fm indicating the frame number for the L, n-th pitch partial series data is initialized to “0” (step S6).

With this pointer fm, paying attention to the pitch partial series data pit _{L, n} [fm], it is determined according to the following equation (1) whether the difference from the average value of the processing target frame fm is equal to or less than a predetermined threshold thpit ( Step S7).
| Pit _{L, n} [fm] −ave _{L, n} | ≦ thpit (1)

If it is determined in step S7 that it is not less than the predetermined threshold (step S7; No), the pitch partial series data pit _{L, n} is deleted (step S8).

  If it is determined in step S7 that the value is equal to or smaller than the predetermined threshold (step S7; Yes), it is determined whether or not the processing has been completed for all fm (step S9).

  If it is determined that the process has not been completed (step S9; No), fm is incremented by “1” in step S12, and the process returns to step S7.

  If it is determined in step S9 that the processes for all fms have been completed (step S9; Yes), it is determined whether or not the processes have been completed for all n (step S10).

  If it is determined that the process has not been completed (step S10; No), n is incremented by “1” in step S13, and the process returns to step S6.

  If it is determined in step S10 that processing for all n has been completed (step S10; Yes), it is determined whether or not processing has been completed for all L (step S11).

  If it is determined that the process has not been completed (step S11; No), L is incremented by “1” in step S14, and the process returns to step S5.

  If it is determined in step S11 whether or not the processing has been completed for all Ls, it is determined that the processing has been completed (step S11; Yes).

Through the above processing, all pit _{L, n} [fm] that can be extracted from the speech database 201 of FIG. 2 and used for learning are stored in the storage unit 102 of FIG. The pit _{L, n} [fm] is used in the phoneme HMM learning unit 250 in FIG.

  In this specific example, the pitch subsequence data used in the phoneme HMM learning unit 250 shown in FIG. 2 is excluded from the pitch subsequence data group by the pitch data editing unit 240 in advance. By using the pitch partial series data, it is possible to achieve the construction of a speech synthesis dictionary that can synthesize more natural synthesized speech.

(Specific example 2 of evaluation / editing process)
Next, specific example 2 of the evaluation / editing process will be described with reference to the flowchart shown in FIG.

  In specific example 1, evaluation and editing are performed by repeating the process for the number of fm in step S7 of FIG. 3. However, if the fm value is large, a processing time is required.

  Therefore, in this specific example, instead of making a determination for each frame, the processing time is shortened by making a determination according to equation (2) described later.

  The flow of the operation of this example is as shown in FIG. 4 and is basically the same as that of Example 1 described with reference to FIG.

  Therefore, in FIG. 4, the same reference numerals are given to steps for performing the same processing as in FIG. 3.

  This example is different from Example 1 mainly in that Step S24 is provided, and Step S27 is provided instead of Step S7, and the loop of Steps S9 and S12 is removed.

In step S24, an average value AVE _L of the pitch partial series data belonging to the same monophone label is calculated according to the following equation.
AVE _L = (1 / N _L ) · Σ _{n = 0} ^NL ave _{L, n}

In step S27, the pitch partial series data for each monophone label is evaluated according to the following equation (2) _{, where} the difference between the average values ave _{L, n} and AVE _L for each frame of the pitch partial series data is a predetermined threshold value. This is done by determining whether or not
| Ave _{L, n} −AVE _L | ≦ THpit (2)

  In this specific example, it is assumed that the user previously sets a threshold value THpit for pitch difference in the storage unit 102 via the user I / F 104 in FIG.

  Thus, according to this example, it is possible to provide a speech synthesis dictionary construction device that contributes to the output of natural synthesized speech that does not cause rapid pitch fluctuations, and a device with high construction efficiency.

  Two specific examples of the evaluation / editing process have been described above, but the evaluation / editing process is not limited to these. Any device may be used as long as it removes a sudden change in pitch data.

For example, for editing, the formula (1), (2) condition instead of deleting the pitch partial sequence data that does not meet the data pit _L does not satisfy the _condition, previous data pit _L of _{n, n-1} or It may be replaced with subsequent data pit _{L, n + 1} .

  In the above, in order to facilitate understanding, an example in which all data is once read from the voice database 201 in FIG. 2 into the storage unit 102 shown in FIG. 1 by the data extraction unit 210 has been described. It is not an essential requirement. For example, depending on the specifications of the phoneme HMM learning unit 250 shown in FIG. 2, it is possible to construct a speech synthesis dictionary more dynamically.

(Embodiment 2)
In the first embodiment, the phoneme label is associated with the phoneme pitch information by the speech synthesis dictionary construction device 100 shown in FIG. The present invention is not limited to this, and can also be applied to the case where phoneme labels are associated with phoneme pitch information and phoneme spectrum parameter information.

  The following describes the speech synthesis dictionary construction apparatus 500 according to the second embodiment in which phoneme labels, phoneme pitch information, and phoneme spectrum parameter information are associated and written to the speech synthesis dictionary.

  As shown in FIG. 5, the speech synthesis dictionary construction apparatus 500 according to the present embodiment includes a data extraction unit 210 and a pitch data evaluation unit 230. These units have the same configurations and functions as the corresponding units of the speech synthesis dictionary construction apparatus 100 shown in FIG. 1 according to the first embodiment.

  The speech synthesis dictionary construction apparatus 500 further includes a sequence data extraction unit 520, a phoneme segmentation unit 525, a data editing unit 540, a phoneme HMM learning unit 550, and a data writing unit 560.

In addition to the same function as pitch sequence data extraction unit 220, sequence data extraction unit 520 further performs D-order mel cepstrum analysis on the audio data extracted by data extraction unit 210, thereby obtaining mel cepstrum sequence data. Mcep ^d _m (0 ≦ d ≦ D) is generated.

In addition to the same function as the phoneme segmentation unit 225, the phoneme segmentation unit 525 further performs a phoneme segmentation process on the mel cepstrum sequence data Mcep ^d _m to obtain the mel cepstrum partial sequence data mcep ^{d for} each identical phoneme label. _{L and n} are generated.

In addition to the same function as that of the pitch data editing unit 240, the data editing unit 540 further includes a mel corresponding to the pitch partial series data pit _{L, n} that has been evaluated and deleted by the pitch data evaluation unit 230 as being unusable as learning data. The cepstrum partial series data mcep ^d _{L, n} is deleted.

  The phoneme HMM learning unit 550 learns the correspondence between the phoneme label string and the corresponding edited pitch sequence data based on the HMM, thereby indicating information indicating the correspondence between the phoneme label and the phoneme pitch information (phoneme HMM). To the data writing unit 560. Further, the phoneme HMM learning unit 550 learns the correspondence relationship between the phoneme label sequence and the mel cepstrum sequence data based on the HMM, converts it into information indicating the correspondence relationship between the phoneme label and the phoneme spectrum parameter information, and writes the data Delivered to part 560.

  The data writing unit 560 writes the correspondence relationship between the phoneme label and the phoneme pitch information and the correspondence relationship between the phoneme label and the phoneme spectrum parameter information to the speech synthesis dictionary 202.

  By using the speech synthesis dictionary 202 constructed in this way, high-quality speech can be synthesized using phoneme pitch information and phoneme spectrum parameter information for each phoneme label.

(Embodiment 3)
In the third embodiment, the pitch partial series data pit _{L, n} is deleted in the first embodiment, and the pitch series data Pit _m obtained by cutting out the pitch partial series data is also deleted.

  This is realized by making a change as shown in FIG. 6 in the specific example 1 of the evaluation / editing process shown in FIG. Hereinafter, the additional part (steps S61 to S68) will be described.

  After taking out the audio data (step S1), a counter m for counting the audio data is provided, an initial value “1” is set in the counter m (step S61), and whether or not the pitch partial series data has been deleted. A flag DEL is provided, and an initial value “0” is set in the flag DEL (step S63).

Then, after deleting the pitch partial series data pit _{L, n} (step S8), "1" is set to the deletion flag DEL (step S64).

After completing the processing for the audio data instructed by the counter m (step S11; Yes), the deletion flag DEL is determined (step S65). If “1” is set, the pitch sequence data Pit _m Is deleted (step S66).

  Thereafter, it is determined whether or not the processing has been completed for all audio data (step S67). If not, m is incremented (step S68), and the processing proceeds to the next audio data (step S63 and subsequent steps).

Further, in the above-described specific example 2 of the evaluation / editing process shown in FIG. 4, the pitch partial series data pit _{L, n} is deleted, and the pitch series data Pit _m obtained by cutting out the pitch partial series data is also deleted. Therefore, a change as shown in FIG. 7 may be added. The additional part of FIG. 7 is the same as the additional part of FIG.

  As a modification of the third embodiment, the pitch partial sequence data and the mel cepstrum partial sequence data are deleted, and the pitch sequence data and the mel cepstrum sequence data obtained by cutting out the pitch partial sequence data and the mel cepstrum partial sequence data are also deleted. You may make it do. This is realized by executing the processing of FIG. 6 or 7 by the apparatus of FIG.

(Embodiment 4)
In the fourth embodiment, the pitch partial series data and the pitch series data in the third embodiment are deleted, and further, the audio data generated from the pitch series data is also deleted. Then, the number of deleted audio data is counted, and when the counted value exceeds a predetermined threshold, all the audio data is recollected.

  This is realized by adding the changes shown in FIG. 8 to the above-described specific example 1 of the evaluation / editing process shown in FIG. Hereinafter, additional portions (steps S62, S80 to S83) will be described.

After taking out the audio data (step S1), a counter Cnt for counting the number of deleted audio data is provided, and an initial value “0” is set in the counter Cnt (step S62). Then, when deleting the pitch partial series data pit _{L, n} and the pitch series data Pit _m , the voice data Sp _m that generated them is deleted (step S80), and the counter Cnt is incremented (step S81).

  Then, after the processing is completed for all audio data (step S67; Yes), it is determined whether the count value of the counter Cnt has exceeded a predetermined threshold Thdel (step S82). When the count value of the counter Cnt exceeds a predetermined threshold Thdel (step S82; Yes), the CPU 105 in FIG. 1 notifies the user via the user I / F 104, for example, and re-records the user. And voice data is collected again (step S83).

Thereby, a higher quality speech synthesis dictionary can be constructed.
Further, in the specific example 2 of the evaluation / editing process shown in FIG. 7 described above, in order to delete the pitch partial series data and the pitch series data, and further delete the audio data that generated the pitch series data, FIG. You may make it add a change as shown in. This is because there is a high possibility that this audio data will generate pitch sequence data that causes a significant pitch fluctuation again. The additional part of FIG. 9 is the same as the additional part of FIG.

Further, as a modification of the fourth embodiment, the pitch partial series data and the mel cepstrum partial series data are deleted, and the pitch series data and the mel cepstrum series data cut out from the pitch partial series data and the mel cepstrum partial series data are deleted. Furthermore, the audio data that generated these data may also be deleted. This is realized by executing the processing of FIG. 8 or FIG. 9 by the apparatus of FIG.
In addition, this invention is not limited to the said embodiment, A various deformation | transformation and application are possible.

  For example, the above-described hardware configuration, block configuration, and flowchart are examples and are not limited.

  The present invention is not limited to the speech synthesis dictionary construction device, and can be constructed using any computer. For example, a program for causing a general-purpose computer to execute the above-described processing is distributed by a recording medium or communication, and this is installed in the computer and executed, thereby allowing the speech synthesis dictionary construction apparatus of the present invention to function. Is possible.

It is a figure which shows the physical structure of the speech synthesis dictionary construction apparatus which concerns on Embodiment 1. FIG. It is a functional block diagram of the speech synthesis dictionary construction apparatus provided with the pitch data evaluation and edit part based on Embodiment 1. FIG. It is a figure which shows the flow of operation | movement in the specific example 1 of pitch data evaluation and edit processing. It is a figure which shows the flow of operation | movement in the specific example 2 of pitch data evaluation and edit processing. It is a function block diagram of the speech synthesis dictionary construction apparatus with spectrum analysis based on Embodiment 2. FIG. It is a figure which shows the flow of operation | movement in the specific example 1 of the pitch data evaluation and edit process based on Embodiment 3. FIG. It is a figure which shows the flow of operation | movement in the specific example 2 of the pitch data evaluation and edit process based on Embodiment 3. FIG. It is a figure which shows the flow of operation | movement in the specific example 1 of the pitch data evaluation and edit process based on Embodiment 4. FIG. It is a figure which shows the flow of operation | movement in the specific example 2 of the pitch data evaluation and edit process based on Embodiment 4. FIG.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 100 ... Speech synthesis dictionary construction apparatus, 101 ... ROM, 102 ... Memory | storage part, 103 ... Data input / output I / F, 104 ... User I / F, 105 ... CPU, 106・ ・ ・ Bus, 111 ... Hard disk with original data, 112 ... Hard disk for recording processed data, 121 ... RAM, 122 ... Hard disk, 141 ... Keyboard, 142 ... Display, 201 ... Speech database, 202 ... Speech synthesis dictionary, 210 ... Data extraction unit, 220 ... Pitch sequence data extraction unit, 225 ... Phoneme extraction unit, 230 ... Pitch data evaluation unit, 240 ... Pitch data editing unit, 250 ... Phoneme HMM learning unit, 260 ... Data writing unit, 500 ... Speech synthesis dictionary construction device, 52 0 ... Series data extraction unit, 525 ... Phoneme extraction unit, 540 ... Data editing unit, 550 ... Phoneme HMM learning unit, 560 ... Data writing unit

Claims (9)

An apparatus for constructing a speech synthesis dictionary,
A receiver for receiving a phoneme label string and corresponding voice data;
A pitch sequence data extraction unit for extracting pitch sequence data from the received audio data;
A phoneme cutout unit that cuts out pitch partial sequence data for each phoneme from the extracted pitch sequence data;
A pitch data evaluation unit that evaluates whether the extracted pitch partial series data is suitable for constructing a speech synthesis dictionary;
A pitch data editing unit for editing pitch partial series data evaluated to be suitable for building a speech synthesis dictionary;
A phoneme HMM learning unit that associates a phoneme HMM related to a pitch with each phoneme label by learning based on the hidden Markov model HMM from the phoneme label sequence and the edited pitch subsequence data;
A data writer for recording the learning results in the speech synthesis dictionary;
A speech synthesis dictionary construction device comprising: Evaluating that it is not suitable for constructing a speech synthesis dictionary together with mel cepstrum partial sequence data corresponding to pitch partial sequence data evaluated as not suitable for constructing a speech synthesis dictionary by the pitch data evaluation unit,
The speech synthesis dictionary construction apparatus according to claim 1. The pitch data evaluation unit evaluates that it is suitable for building a speech synthesis dictionary when the difference between the pitch partial series data and the average value does not exceed a predetermined threshold in all frames.
The speech synthesis dictionary construction device according to claim 1 or 2. The pitch data evaluation unit is suitable for constructing a speech synthesis dictionary when a difference between an average value of pitch partial series data and an average value of pitch partial series data belonging to the same monophone label does not exceed a predetermined threshold. evaluate,
The speech synthesis dictionary construction device according to claim 1 or 2. Deleting the pitch partial series data and deleting the pitch series data cut out from the pitch partial series data;
The speech synthesis dictionary construction apparatus according to claim 1. Deleting the mel cepstrum partial series data and deleting the mel cepstrum series data cut out from the mel cepstrum partial series data;
The speech synthesis dictionary construction device according to claim 2. Count the number of audio data from which the pitch partial series data has been deleted, and re-collect all audio data when the count value exceeds a predetermined threshold.
The speech synthesis dictionary construction device according to claim 1 or 2. A method for building a speech synthesis dictionary,
A receiving step of receiving a phoneme label string and corresponding voice data from the database;
A pitch sequence data extraction step for extracting pitch sequence data from the received audio data;
A phoneme extraction step of extracting pitch partial sequence data for each phoneme from the extracted pitch sequence data;
A pitch data evaluation step for evaluating whether the extracted pitch subsequence data is suitable for constructing a speech synthesis dictionary;
A pitch data editing step for editing pitch subsequence data evaluated to be suitable for building a speech synthesis dictionary;
A phoneme HMM learning step in which a phoneme HMM associated with a pitch is associated with each phoneme label by learning based on a hidden Markov model from the phoneme label sequence and the edited pitch subsequence data;
An output step for outputting the learning result;
A speech synthesis dictionary construction method comprising: On the computer,
A receiving step of receiving a phoneme label string and corresponding voice data from the database;
A pitch sequence data extraction step for extracting pitch sequence data from the received audio data;
A phoneme extraction step of extracting pitch partial sequence data for each phoneme from the extracted pitch sequence data;
A pitch data evaluation step for evaluating whether the extracted pitch subsequence data is suitable for constructing a speech synthesis dictionary;
A pitch data editing step for editing pitch subsequence data evaluated to be suitable for building a speech synthesis dictionary;
A phoneme HMM learning step in which a phoneme HMM associated with a pitch is associated with each phoneme label by learning based on a hidden Markov model from the phoneme label sequence and the edited pitch subsequence data;
An output step for outputting the learning result;
A program that executes

Images