JP2008026777A - Speech synthesis dictionary structuring device, speech synthesis dictionary structuring method, and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To structure a speech synthesis dictionary with which an articulate speech can be synthesized. <P>SOLUTION: A formant emphasizing unit 145 or 445 converts melcepstrum coefficient sequence data generated by an analyzing unit 141 or 441 into emphasized melcepstrum coefficient sequence data for emphasizing formants of a speech spectrum. The formant emphasizing unit is disposed in front or in a phoneme HMM learning unit 151 or 431 which writes a phoneme HMM to a speech synthesis dictionary 173 or 453 while making it correspond to each phoneme label. <P>COPYRIGHT: (C)2008,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 speech synthesis dictionary used for speech synthesis and the like.

  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.

  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.

  The speech synthesis dictionary is constructed by a speech synthesis dictionary construction device. A speech synthesis dictionary construction device generally uses a mel cepstrum analysis and a pitch for data recorded in a database (hereinafter referred to as a speech database) configured from a set of phoneme label sequences and speech data corresponding thereto. A speech synthesis dictionary is constructed by performing extraction and a learning process based on the HMM.

  A conventional speech synthesis dictionary construction device constructs a speech synthesis dictionary by using speech data recorded in the speech database as it is for learning based on the HMM without any special processing when constructing the speech synthesis dictionary. Was.

  Further, the conventional speech synthesis dictionary construction device uses the mel cepstrum coefficient series data generated as a result of the mel cepstrum analysis in the learning process based on the HMM as it is for learning based on the HMM without performing any particular processing. I was building a speech synthesis dictionary.

  However, when speech is synthesized using the speech synthesis dictionary constructed as described above, the shape of the valley of the spectrum envelope of the speech data is smoothed compared to the shape of the valley of the spectrum envelope of the original speech data.

  The synthesized speech using the speech synthesis dictionary constructed by the conventional speech synthesis dictionary construction device is smoother than the natural speech of human beings as a result of smoothing the shape of the spectrum envelope of the speech data. It was damaged.

  The present invention has been made in view of the above circumstances, and a speech synthesis dictionary construction device, a speech synthesis dictionary construction method, and a program capable of constructing a speech synthesis dictionary capable of synthesizing clear speech. The purpose is to provide.

In order to achieve the above object, a speech synthesis dictionary construction device according to the first aspect of the present invention provides:
An input unit for inputting a phoneme label string and corresponding voice data;
An audio data processing unit for emphasizing spectral parameters for the audio data and converting the audio data into emphasized audio data;
A phoneme HMM learning unit that associates a phoneme HMM (Hidden Markov Model) for each phoneme label from the phoneme label string and the emphasized speech data;
A data writer for recording the learning results in the speech synthesis dictionary;
Is provided.

  Rather than using the original speech data as it is, the phoneme HMM learning is performed using the speech data that has been subjected to enhancement processing in advance, so that the speech synthesis dictionary that the speech synthesizer refers to generate clear synthesized speech A suitable speech synthesis dictionary is constructed.

  The voice data processing unit is defined by the in-processing mel cepstrum coefficient series data that performs mel cepstrum analysis on the voice data to generate first mel cepstrum coefficient series data, and is defined by the first mel cepstrum coefficient series data. An inverse filter that generates excitation sound source data from the data, and an in-process formant emphasis unit that generates a predetermined emphasis mel cepstrum coefficient series data in which the first mel cepstrum coefficient series data is subjected to a predetermined emphasis process to emphasize the spectral parameters; And a synthesis filter that is defined by the emphasized mel cepstrum coefficient series data and that generates the formant-enhanced processed speech data by inputting the excitation sound source data.

  Editing the mel cepstrum coefficient series data after the mel cepstrum analysis in this way is a reliable and simple method for performing formant emphasis processing on speech data.

  The phoneme HMM learning unit performs a mel cepstrum analysis on the formant-emphasized processed speech data to generate a second mel cepstrum coefficient sequence data; and a pitch sequence from the formant-enhanced processed speech data A pitch extraction unit that extracts data; a mel cepstrum learning unit that associates a phoneme HMM related to a mel cepstrum for each phoneme label from the phoneme label sequence and the second mel cepstrum coefficient sequence data; A pitch learning unit that associates a phoneme HMM related to the pitch for each phoneme label from the pitch series data may be provided.

  Accordingly, it is possible to generate both a mel cepstrum speech synthesis dictionary storing the phoneme HMM learning results related to the mel cepstrum and a pitch speech synthesis dictionary storing the phoneme HMM learning results related to the pitch.

In order to achieve the above object, a speech synthesis dictionary construction device according to the second aspect of the present invention provides:
An input unit for inputting a phoneme label string and corresponding voice data;
A mel cepstrum analysis unit in a learning unit that performs mel cepstrum analysis on the speech data to generate mel cepstrum coefficient series data;
A pitch extraction unit that extracts pitch series data from the audio data;
A formant emphasis unit in a learning unit that generates a emphasized mel cepstrum coefficient series data in which a predetermined emphasis process is performed on the mel cepstrum coefficient series data to emphasize spectrum parameters;
A mel cepstrum learning unit that associates a phoneme HMM (Hidden Markov Model) related to a mel cepstrum for each phoneme label from the phoneme label string and the emphasized mel cepstrum coefficient series data;
From the phoneme label string and the pitch series data, a pitch learning unit that associates a phoneme HMM related to pitch for each phoneme label;
A data writer for recording the learning results in the speech synthesis dictionary;
Is provided.

  In this case, the original speech data is input as it is to the phoneme learning unit, but since a formant emphasis process is inserted into the phoneme HMM learning, the speech that is referred to by the speech synthesizer to generate clear synthesized speech A speech synthesis dictionary suitable as a synthesis dictionary is constructed.

  The predetermined enhancement process is, for example, a process of multiplying a degree greater than a predetermined order among the mel cepstrum coefficient series data by a predetermined enhancement coefficient greater than 1.

  By amplifying the higher-order mel cepstrum coefficient series data, the difference between the peak and valley of the spectral envelope is amplified compared to the spectral envelope of the original speech data.

  The predetermined enhancement coefficient may be different for each audio data that is a generation source of mel cepstrum coefficient series data multiplied by the enhancement coefficient.

  By appropriately changing the emphasis coefficient for each voice data, it can be adjusted even when there is a change in the speaker's voice over time when collecting a large amount of voice data, for example, and an appropriate voice synthesis dictionary is constructed can do.

  The predetermined enhancement coefficient may vary depending on the order of the mel cepstrum coefficient series data multiplied by the enhancement coefficient.

  If it varies depending on the order of the mel cepstrum coefficient series data, more appropriate formant emphasis can be performed.

In order to achieve the above object, a speech synthesis dictionary construction method according to the third aspect of the present invention provides:
An input step in which a phoneme label string and corresponding speech data are input from the database;
An audio data processing step of enhancing spectral parameters with respect to the audio data and converting them into emphasized audio data;
A phoneme HMM learning step in which a phoneme HMM (Hidden Markov Model) is associated with each phoneme label from the phoneme label string and the emphasized speech data;
A data writing step for recording the learning results in the speech synthesis dictionary;
Consists of

In order to achieve the above object, a speech synthesis dictionary construction method according to the fourth aspect of the present invention provides:
An input step in which a phoneme label string and corresponding speech data are input from the database;
A mel cepstrum analysis step in a learning unit that performs mel cepstrum analysis on the voice data to generate mel cepstrum coefficient series data;
A pitch extraction step of extracting pitch series data from the audio data;
In-learning formant emphasis step for generating emphasized mel cepstrum coefficient series data in which the mel cepstrum coefficient series data is subjected to predetermined emphasis processing to emphasize spectral parameters;
A mel cepstrum learning step for associating a phoneme HMM (Hidden Markov Model) related to a mel cepstrum for each phoneme label from the phoneme label string and the emphasized mel cepstrum coefficient series data;
From the phoneme label string and the pitch series data, a pitch learning step for associating a phoneme HMM related to pitch for each phoneme label;
A data writing step for recording the learning results in the speech synthesis dictionary;
Consists of

In order to achieve the above object, a computer program according to the fifth aspect of the present invention provides:
On the computer,
An input step in which a phoneme label string and corresponding speech data are input from the database;
An audio data processing step of enhancing spectral parameters with respect to the audio data and converting them into emphasized audio data;
A phoneme HMM learning step in which a phoneme HMM (Hidden Markov Model) is associated with each phoneme label from the phoneme label string and the emphasized speech data;
A data writing step for recording the learning results in the speech synthesis dictionary;
It is a computer program that executes

In order to achieve the above object, a computer program according to the sixth aspect of the present invention provides:
On the computer,
An input step in which a phoneme label string and corresponding speech data are input from the database;
A mel cepstrum analysis step in a learning unit that performs mel cepstrum analysis on the voice data to generate mel cepstrum coefficient series data;
A pitch extraction step of extracting pitch series data from the audio data;
In-learning formant emphasis step for generating emphasized mel cepstrum coefficient series data in which the mel cepstrum coefficient series data is subjected to predetermined emphasis processing to emphasize spectral parameters;
A mel cepstrum learning step for associating a phoneme HMM (Hidden Markov Model) related to a mel cepstrum for each phoneme label from the phoneme label string and the emphasized mel cepstrum coefficient series data;
From the phoneme label string and the pitch series data, a pitch learning step for associating a phoneme HMM related to pitch for each phoneme label;
A data writing step for recording the learning results in the speech synthesis dictionary;
It is a computer program that executes

  According to the present invention, after speech data is subjected to formant emphasis processing, a speech synthesis dictionary is constructed using phoneme HMM learning. Alternatively, a speech synthesis dictionary is constructed by using the mel cepstrum coefficient series data generated in the phoneme HMM learning process for phoneme HMM learning while performing formant emphasis processing. For this reason, the synthesized speech obtained by using the speech synthesis dictionary can be made clear with formants emphasized.

  Hereinafter, the speech synthesis dictionary construction device according to the embodiment of the present invention will be described in detail.

(Embodiment 1)
FIG. 1 is a functional configuration diagram of the speech synthesis dictionary construction device 111 according to Embodiment 1 of the present invention.

  As shown in the figure, the speech synthesis dictionary construction device 111 includes an input unit 121, a data writing unit 123, a speech data processing unit 131, and a phoneme HMM learning unit 151.

  The voice data processing unit 131 includes a processing unit internal mel cepstrum analysis unit 141, an inverse filter 143, a processing unit internal formant emphasis unit 145, and a synthesis filter 147.

  The phoneme HMM learning unit 151 includes an in-learning unit mel cepstrum analysis unit 161, a pitch extraction unit 163, a mel cepstrum learning unit 165, and a pitch learning unit 167.

  The speech synthesis dictionary construction device 111 is connected to a speech database 171 and a speech synthesis dictionary 173 as shown in FIG.

  The voice database 171 is composed of a hard disk or the like, and stores a plurality of sets of phoneme label strings and corresponding voice data. The voice database 171 and the input unit 121 are connected.

  The voice data is delivered from the voice database 171 to the in-process mel cepstrum analysis unit 141 inside the voice data processing unit 131 via the input unit 121. The audio data processing unit 131 emphasizes the spectrum parameter for the audio data.

  The phoneme label string is transferred from the speech database 171 to the mel cepstrum learning unit 165 and the pitch learning unit 167 inside the phoneme HMM learning unit 151 via the input unit 121.

  The speech synthesis dictionary 173 is composed of a hard disk or the like, and stores a plurality of sets of phoneme labels and corresponding phoneme HMMs. The speech synthesis dictionary 173 and the data writing unit 123 are connected. The speech synthesis dictionary 173 is constructed by the speech synthesis dictionary construction device 111.

  The in-processing unit mel cepstrum analysis unit 141 performs mel cepstrum analysis on the voice data to generate mel cepstrum coefficient series data, and passes it to the in-processing unit formant emphasis unit 145.

  The inverse filter 143 is defined by the generated mel cepstrum coefficient series data.

  Audio data is input to the inverse filter 143. As a result, excitation sound source data is generated and delivered to the synthesis filter 147.

  The in-process formant emphasizing unit 145 performs formant emphasis processing on the mel cepstrum coefficient series data delivered from the in-process mel cepstrum analysis unit 141 by enlarging the higher-order mel cepstrum coefficient series data, thereby Emphasized mel cepstrum coefficient series data with enhanced spectral parameters is generated and delivered to the synthesis filter 147.

  A detailed procedure of the formant emphasis process will be described later with reference to FIGS.

  The synthesis filter 147 is an MLSA (Mel Log Spectrum Approximation) synthesis filter that uses enhanced mel cepstrum coefficient series data as coefficients, and generates formant-enhanced voice data by inputting excitation sound source data.

  The synthesis filter 147 passes the generated formant-emphasized speech data to the in-learning mel cepstrum analysis unit 161 and the pitch extraction unit 163 inside the phoneme HMM learning unit 151.

  In the phoneme HMM learning unit 151, formant-emphasized speech data that has been formant-enhanced in advance by the speech data processing unit 131 is referred to instead of directly referring to the speech data recorded in the speech database 171. In other respects, a method similar to the known method is employed. Then, so-called phoneme HMM learning is performed in which a phoneme HMM corresponding to each phoneme label is determined.

  The in-learning unit mel cepstrum analysis unit 161 extracts mel cepstrum coefficient series data from the delivered formant-emphasized speech data by mel cepstrum analysis.

  The pitch extraction unit 163 extracts pitch series data from the delivered formant-enhanced voice data.

  In the mel cepstrum learning unit 165, phoneme HMM learning related to the mel cepstrum is performed from the phoneme label string delivered from the speech database 171 via the input unit 121 and the mel cepstrum coefficient sequence data generated by the in-learning mel cepstrum analysis unit 161. Is done. The learning result is recorded in the speech synthesis dictionary 173 via the data writing unit 123.

  Further, the phoneme HMM learning regarding the pitch is performed in the pitch learning unit 167 from the phoneme label string delivered from the speech database 171 via the input unit 121 and the pitch sequence data generated by the pitch extraction unit 163. The learning result is recorded in the speech synthesis dictionary 173 via the data writing unit 123.

  The speech synthesis dictionary construction device 111 shown in FIG. 1 is physically configured by a general computer device 211 as shown in FIG.

  The CPU 221, ROM 223, storage unit 225, user interface (hereinafter referred to as I / F) 227, and data input / output I / F 229 are connected to each other via a bus 231.

  In addition to the known operation program for learning based on the HMM, the ROM 223 stores an operation program for performing formant emphasis processing on audio data, particularly in this embodiment.

  The storage unit 225 includes a RAM 241 and a hard disk 243, and supports a phoneme HMM for each formant emphasis constant, phoneme label string, voice data, mel cepstrum coefficient series data, excitation source data, pitch series data, and phoneme label. Temporarily memorize the attached ones.

  The data input / output I / F 229 is an interface for connecting to the original data-containing hard disk 261 or the like and the processed data recording hard disk 263 or the like. The original data hard disk 261 corresponds to the speech database 171 in FIG. 1, and the processed data recording hard disk 263 corresponds to the speech synthesis dictionary 173 in FIG.

  The data input / output I / F 229 is connected to the voice database 171 shown in FIG. 1, reads out a pair of phoneme label strings and voice data and stores it in the storage unit 225 under the control of the CPU 221 shown in FIG.

  The CPU 221 sequentially reads out the voice data from the storage unit 225 according to the operation program for performing the formant emphasis process stored in the ROM 223, performs the formant emphasis process, and stores the processed voice data in association with the phoneme label string. Stored in the unit 225.

  As a result of the processing of the CPU 221, the storage unit 225 stores a pair of phoneme label strings and formant-enhanced voice data. This is used for phoneme HMM learning.

  The CPU 221 executes a synthesis dictionary generation operation by executing an operation program for phoneme HMM learning stored in the ROM 223.

  The data input / output I / F 229 is connected to the speech synthesis dictionary 173 shown in FIG. 1, and is a result of the processing by the CPU 221 shown in FIG. 2, and the phoneme HMM related to the mel cepstrum for each phoneme label and the phoneme related to the pitch for each phoneme label. The HMM is output to the speech synthesis dictionary 173 shown in FIG.

  The user I / F 227 shown in FIG. 2 includes a keyboard 251 and a monitor 253, and is provided for inputting arbitrary instructions, data, and programs. In particular, in formant emphasis processing, the user needs to give various constants via the I / F.

  As shown in FIG. 1, the feature of the speech synthesis dictionary construction apparatus 111 according to the present embodiment is that the speech data processing unit 131 performs a predetermined enhancement process for enhancing the formant of each speech data.

  The voice data is processed into formant-emphasized voice data by passing through the voice data processing unit 131. More specifically, the higher-order mel cepstrum coefficient series data is amplified by the in-process formant emphasis unit 145.

  The predetermined emphasis process executed by the audio data processing unit 131 may be any process as long as it finally enhances the formant of the audio data. Below, the typical procedure of an emphasis process is demonstrated.

  In the following description, a frame means a time segment used to convert audio data into a spectrum, and is represented by the symbol fm.

  With reference to the flowchart shown in FIG. 3, the specific example of the emphasis process which the audio | voice data processing part 131 performs is demonstrated.

In this example, in advance, the user is performed in the processing portion mel cepstrum analysis section 141, the order D _a mel cepstrum analysis of the voice data, via a user I / F227 of FIG. 2, set in the storage unit 225 It is assumed that this is done (step S315).

  As shown in FIG. 1, when the speech synthesis dictionary construction device 111 constructs the speech synthesis dictionary 173, the speech database 171 and, for example, an empty speech synthesis dictionary 173 are connected to the speech synthesis dictionary construction device 111. The

  When an instruction to start construction of the speech synthesis dictionary 173 is issued from the user I / F 227 of FIG. 2, 1 is stored as an initial value of the counter n in the counter register in the CPU 221 of FIG. 2 (step S311). This n is a variable for identifying the audio data recorded in the audio database 171 of FIG.

The input unit 121 in FIG. 1 extracts the voice data voice Sp _n (where 1 ≦ n ≦ N _SP , where N _SP is the number of data in the voice database) from the voice database 171, and the storage unit in FIG. 2. It memorize | stores in 225 (step S313).

The in-process mel cepstrum analysis unit 141 in FIG. 1 performs a Da _c order (order is given in step S315 as described above) on the audio data Sp _n (step S317).

As a result of the analysis, mel cepstrum coefficient series data MC _{n, d} [fm] (1 ≦ n ≦ N _SP , 0 ≦ d ≦ D _a , 0 ≦ fm ≦ N _fm [n], where N _fm [n] is is the number of frames to speech data Sp _n.) because is generated, and stores them in the storage unit 225 of FIG. 2 (step S319).

  An inverse MLSA filter is defined by these mel cepstrum coefficient series data (step S321 in FIG. 3 and a dotted arrow in the audio data processing unit 131 in FIG. 1).

As a result of inputting the audio data Sp _n to the inverse MLSA filter defined in this way (step S323 in FIG. 3), excitation sound source data ExData _n is generated and stored in the storage unit 225 in FIG. 2 (step in FIG. 3). S325).

The CPU 221 in FIG. 2 reads the mel cepstrum coefficient series data MC _{n, d} [fm] from the storage unit 225 in accordance with an instruction of the audio data processing program stored in the ROM 223, and performs a predetermined emphasis process described in detail later. Thus, the emphasized mel cepstrum coefficient series data EMMC _{n, d} [fm] is generated and stored in the storage unit 225 of FIG. 2 (step S327 of FIG. 3). In FIG. 1, this processing is performed by the in-process formant emphasis unit 145.

The enhanced mel cepstrum coefficient series data EmMC _{n, d} [fm] generated in this way is used as the coefficient of the MLSA synthesis filter, and when the excitation sound source data ExData _n is input, formant enhanced speech data EmSp _n is generated. . This is stored in the storage unit 225 of FIG.

That is, in order to convert the excitation sound source data ExData _n into formant-enhanced processed speech data EmSp _n by MLSA synthesis, the operation content of the MLSA synthesis filter is defined by the enhanced mel cepstrum coefficient series data EmMC _{n, d} [fm] ( Step S329). When the excitation sound source data ExData _n called from the storage unit 225 in FIG. 2 is input to the MLSA synthesis filter defined in this way (step S331), formant-emphasized voice data EmSp _n is calculated. The calculation result is stored in the storage unit 225 of FIG. 2 (step S333).

  It is determined whether or not all the audio data recorded in the audio database 171 in FIG. 1 has been converted to formant-enhanced audio data (step S335).

Specifically, when the counter n is smaller than the number N _{Sp of} audio data in the counter register inside the CPU 221 in FIG. 2 (step S335; Yes), after incrementing n by 1 (step S337), Back to the audio data Sp _n extraction step (step S313).

If n is greater than or equal to N _Sp (step S335; No), it means that the conversion of all audio data recorded in the audio database 171 of FIG. 1 into formant enhanced audio data has been completed. Then, the formant emphasis processing in the data processing unit 131 ends.

  Thereafter, the formant-enhanced speech data is delivered to the phoneme HMM learning unit 151 in FIG. As described above, the phoneme HMM learning unit 151 is known except that instead of directly referring to the voice data recorded in the voice database 171, the voice data that has been subjected to formant enhancement in advance by the voice data processing unit 131 is referred to. Phoneme HMM learning is performed by a method similar to the above method.

  According to the speech synthesis dictionary construction device 111 according to the present embodiment, phoneme HMM learning is performed with reference to speech data that has been subjected to formant emphasis in advance. Therefore, the speech synthesis dictionary constructed by the device is a speech synthesis device. Contributes to generating clear synthesized speech.

(Embodiment 2)
FIG. 4 is a functional configuration diagram of the speech synthesis dictionary construction device 411 according to the second embodiment.

  In the case of the above-described embodiment, speech data processing is performed before phoneme HMM learning, whereas in this embodiment, a formant emphasis unit is provided inside the phoneme HMM learning unit.

  The speech synthesis dictionary construction apparatus 411 basically uses a speech database 451 to perform a phoneme HMM learning related to a mel cepstrum and a phoneme HMM learning related to a pitch, and write a learning result to the speech synthesis dictionary 453. It has the same configuration as the construction device.

  That is, as shown in the figure, an input unit 421, a data writing unit 423, and a phoneme HMM learning unit 431 are provided. The phoneme HMM learning unit 431 includes a learning-in-part mel cepstrum analysis unit 441, a pitch extraction unit 443, A mel cepstrum learning unit 447 and a pitch learning unit 449 are provided.

  However, the speech synthesis dictionary construction device 411 according to the present embodiment further includes an in-learning formant emphasis unit 445 inside the phoneme HMM learning unit 431.

  In the known speech synthesis dictionary construction device, the analysis result by the in-learning unit mel cepstrum analysis unit 441 is directly transferred to the mel cepstrum learning unit 447.

  On the other hand, in the speech synthesis dictionary construction device 411 according to the present embodiment, the in-learning unit mel cepstrum analysis unit 441 first passes the mel cepstrum coefficient series data generated from the speech data to the in-learning unit formant emphasizing unit 445.

  The in-learning formant emphasizing unit 445 performs predetermined formant emphasis processing on the delivered mel cepstrum coefficient series data, converts the data into emphasized mel cepstrum coefficient series data, and then delivers the mel cepstrum coefficient series data to the mel cepstrum learning unit 447.

  The predetermined formant emphasis process is a process of increasing mel cepstrum coefficient series data of higher order than a predetermined order. Details of such enhancement processing will be described later with reference to FIG.

  Similar to the device 111 according to the previous embodiment, the speech synthesis dictionary construction device 411 shown in FIG. 4 is also physically configured by a general computer device 211 as shown in FIG.

  In addition to the known operation program for learning based on the HMM, the ROM 223 stores an operation program for performing formant emphasis processing on the mel cepstrum coefficient series data in this embodiment.

  In the formant emphasis process, the user needs to give various constants via the user I / F 227.

  Hereinafter, a predetermined enhancement process executed by the in-learning formant emphasis unit 445 inside the phoneme HMM learning unit 431 will be described with reference to FIG.

First, the user stores the order _C of the mel cepstrum analysis executed in the in-learning unit mel cepstrum analysis unit 441 in FIG. 4 in the storage unit 225 via the user I / F 227 (step S515).

  The counter n for voice data identification is stored in the counter register inside the CPU 221 in FIG. The initial value of n is 1 (step S511).

The CPU 221 in FIG. 2 follows the instruction of the program stored in the ROM 223, and the nth audio data Sp among the N _Sp audio data recorded in the audio database 451 in FIG. 4 via the data input / output I / F 229. _n is extracted (step S513), stored in the storage unit 225, and subjected to D _{c -th} order mel cepstrum analysis (step S517).

As a result of the analysis, the mel cepstrum coefficient series data MC _{n, d} [fm] (1 ≦ n ≦ N _Sp , 0 ≦ d ≦ D _c , 0 ≦ fm ≦ N _fm [n], where N _fm [n] is a voice a number of frames for data Sp _n.) are generated, CPU 221 of FIG. 2 stores them in the storage unit 225 (step S519).

The CPU 221 in FIG. 2 sequentially calls the mel cepstrum coefficient series data MC _{n, d} [fm] from the storage unit 225 and performs predetermined enhancement processing described in detail later, thereby performing the emphasized mel cepstrum coefficient series data EMMC _{n, d} [ fm] is generated and stored in the storage unit 225 (step S521).

It is determined whether or not the generation of the emphasized mel cepstrum coefficient series data has been completed for all of the N _Sp speech data recorded in the speech database of FIG. 4 (step S523).

  If not completed yet (step S523; Yes), the next audio data (step S525) is repeated from the audio data capturing operation (step S513).

If completed (step S523; No), the phoneme HMM related to the mel cepstrum is obtained from the audio data Sp _n and the emphasized mel cepstrum coefficient series data EMMC _{n, d} [fm] stored in the storage unit 225 of FIG. Learning is performed (step S527).

  The learning data obtained as a result of the phoneme HMM learning is sent to the data writing unit 423 in FIG. 4 (step S529).

As described above, when performing phoneme HMM learning related to the mel cepstrum, unlike the known method, not the processed mel cepstrum coefficient series data MC _{n, d} [fm] but the emphasized mel cepstrum coefficient series data EMMC _{n. , D} [fm] is used, the speech synthesis dictionary constructed by the speech synthesis dictionary construction device 411 according to the present embodiment helps the speech synthesis device to generate clear synthesized speech.

(Embodiment 3)
In the speech synthesis dictionary construction device 111 according to the first exemplary embodiment illustrated in FIG. 1, the speech data processing unit 131 and the phoneme HMM learning unit 151 are housed in the same casing. Then, the formant-emphasized speech data generated by the synthesis filter 147 is stored in the storage unit 225 of FIG. 2, that is, the speech synthesis dictionary construction device 111, and then extracted and used by the phoneme HMM learning unit 151. .

  Here, the formant emphasis processing data generated by the synthesis filter 147 does not necessarily have to be stored in a storage device inside the single speech synthesis dictionary construction device 111, for example, the hard disk 243 shown in FIG.

  Therefore, the following can be considered as the device set according to the third embodiment. That is, the speech data processing unit 131 and the phoneme HMM learning unit 151 of the speech synthesizer according to Embodiment 1 shown in FIG. A set of an independent speech data processing device and a phoneme HMM learning device is used.

  In this case, the two devices are connected via an external recording medium such as a hard disk. That is, the speech data processing device records a pair of speech label string and formant-emphasized speech data on an external recording medium, while the phoneme HMM learning device reads the pair from the external recording medium.

  In other words, the speech data processing device is a device that recreates the original speech database into a formant-emphasized speech database, while the phoneme learning device has the same configuration as a known device, but also refers to speech The apparatus is characterized in that the database is different from the known one.

(Embodiment 4)
The speech data processing device in the device set according to Embodiment 3 and the speech synthesis dictionary construction device 411 according to Embodiment 2 shown in FIG. 4 may be connected via an external recording medium such as a hard disk.

  Alternatively, the learning shown in FIG. 4 is performed between the in-learning mel cepstrum analysis unit 161 and the mel cepstrum learning unit 165 in the phoneme HMM learning unit 151 of the speech synthesis dictionary construction apparatus 111 according to the first embodiment shown in FIG. Even if the internal formant emphasizing unit 445 is inserted, it has substantially the same function.

  According to this embodiment, formant emphasis is performed twice, so that a more appropriate speech synthesis dictionary can be constructed as a dictionary that is referred to by the speech synthesis dictionary that generates clear synthesized speech.

(About emphasis processing)
The mel cepstrum coefficient series data qualitatively considers the horizontal axis of the speech spectrum that originally means the frequency as time, and what frequency component the waveform has in the real time domain, It can be said that it is the result of analyzing.

  In general, there are less obvious formants between clear formants appearing in the speech spectrum. That is, generally clear formants are distributed at relatively wide frequency intervals, while other formants tend to be distributed at relatively narrow frequency intervals between such prominent formants.

  Narrow frequency interval variation components correspond to "high frequencies" in the cepstrum. Such a high frequency corresponds to a higher order of the mel cepstrum coefficient series data.

  Therefore, amplifying higher-order mel cepstrum coefficient series data obtained by mel cepstrum analysis of voice data means emphasizing peaks and valleys of the envelope of the voice spectrum. By emphasizing the valley, the formant of the voice data is emphasized. Here, this is called enhancement processing. Below, the specific example of an emphasis process is shown.

  It should be noted that the formant emphasis unit that performs the emphasis process may have a different position in the speech synthesis dictionary construction device when the embodiment is different, but is the same in that a predetermined arithmetic process is performed on the mel cepstrum coefficient series data. Therefore, in the following, the distinction between the above-described embodiments is not limited.

(Specific example 1 of emphasis processing)
Specific example 1 of the enhancement process will be described with reference to FIG.

In this specific example, first, the user determines how many or more orders of mel cepstrum coefficients are to be amplified, and stores them in the storage unit 225 via the user I / F 227 of FIG. 2 (step S611). In the following _description , it is assumed that _dem (2 ≦ d _em ≦ D _b , where D _b is the order taken into consideration in the mel cepstrum analysis) and the mel cepstrum coefficient is amplified.

The mel cepstrum analysis for all N _SP speech data has been completed, and the mel cepstrum coefficient series data MC _{n, d} [fm] (1 ≦ n ≦ N _Sp , 0 ≦ d ≦ D _a , 0 ≦ fm ≦ N _fm [n] and N _fm [n] are the number of frames for the audio data Sp _n .) Are already stored in the storage unit 225 of FIG.

In the following, only the process related to the nth audio data will be described. In order to complete the formant emphasis processing, it is necessary to scan all n (1 ≦ n ≦ N _Sp ).

The CPU 221 in FIG. 2 loads the variable d as a counter into the register in accordance with the instruction of the operation program stored in the ROM 223. d represents the order of the mel-cepstrum, since it is 0 ≦ d ≦ _{D b,} the initial value is set to 0 (step S613).

CPU221 of Figure 2, according to an instruction operation program stored in the ROM 223, d is determined whether a given _{d em} or more in step S611 (step S615).

When d is smaller than d _em (step S615; No), the original mel cepstrum coefficient series data MC _{n, d} [fm] (0 ≦ fm ≦ N _fm [n]) is not amplified and is directly enhanced mel cepstrum. Coefficient series data EMMC _{n, d} [fm]. That is, EMMC _{n, d} [fm] = MC _{n, d} [fm] is set (step S619).

If d is greater than or equal _{d em} (step S615; Yes), the original Mel cepstrum coefficient series data _{MC n, d [fm] (} 0 ≦ fm ≦ N fm [n]) to amplify the emphasis mel cepstrum coefficient series data EMMC _{n, d} [fm].

In this specific example, amplification is performed by multiplying a predetermined enhancement coefficient larger than 1 regardless of the order. That is, EMMC _{n, d} [fm] = MC _{n, d} [fm] × (1 + β) (where β> 1) (step S617).

  Amplifying higher-order mel cepstrum coefficient series data in this way emphasizes the high-frequency component of the audio data, so that the formant is emphasized compared to the original audio data and becomes clear.

When finished processing in step S619 or step S617 of degree d, d, it is determined whether or not the _{D b} is smaller than (step S621). For d _{<D b} (step S621; Yes), increments d by 1 (step S623), the process proceeds to step S615 and subsequent steps for the next order. For d ≧ _{D b} (step S621; No), emphasis mel cepstrum coefficient series data _{eMMC n} for all orders, since the generation of d [fm] is completed, and ends the enhancement process.

  By performing the processing as described above, formant emphasis of audio data can be easily achieved. Then, if a speech synthesis dictionary construction device incorporating such a process is used, a speech suitable for making the synthesized speech clear as a speech synthesis dictionary to be referred to when the speech synthesizer generates synthesized speech. A consensus dictionary can be constructed.

(Specific example 2 of emphasis processing)
In the above-described specific example, the emphasis coefficient is a constant of 1 + β, but the emphasis coefficient may be used depending on the audio data. That is, in step S617 of FIG. 6, the enhanced mel cepstrum coefficient series data may be obtained by setting β as a function of n and EMMC _{n, d} [fm] = MC _{n, d} [fm] × (1 + β _n ).

  As a more specific example, when the voice database 171 or 451 shown in FIG. 1 or 4 records voice data whose voice data has changed characteristics depending on the situation during recording, the voice data is recorded. For example, the emphasis coefficient may be changed in accordance with the situation at the time of being performed.

  Thus, by changing the enhancement coefficient for each voice data, even if the voice in the voice database is a voice under a plurality of situations, the enhancement of the formant of the voice data can be appropriately executed. Then, if a speech synthesis dictionary construction device incorporating such a process is used, a speech suitable for making the synthesized speech clear as a speech synthesis dictionary to be referred to when the speech synthesizer generates synthesized speech. A consensus dictionary can be constructed.

(Specific example 3 of emphasis processing)
The enhancement coefficient may be properly used depending on the order d of the mel cepstrum coefficient series data. That is, in step S617 of FIG. 6, the enhanced mel cepstrum coefficient series data may be obtained by setting β as a function of d and EMMC _{n, d} [fm] = MC _{n, d} [fm] × (1 + β _d ).

In emphasizing a formant in a speech spectrum, it is not always appropriate to amplify mel cepstrum coefficient series data having an order equal to or higher than _dem which is a threshold of the order to be emphasized. Furthermore, the uniform amplification will unnecessarily emphasize the peaks and valleys of the envelope of the speech spectrum, resulting in the appearance of formants that should not exist. On the contrary, there is a possibility that it is contrary.

  Thus, by using different emphasis coefficients depending on the order d of the mel cepstrum coefficient series data, it is possible to construct a speech synthesis dictionary in accordance with the object of the present invention.

(Specific example 4 of emphasis processing)
Specific example 2 and specific example 3 described above may be combined. That is, in step S617 of FIG. 6, β is a function of both n and d, and emmmel cepstrum coefficient series data is obtained as EMMC _{n, d} [fm] = MC _{n, d} [fm] × (1 + β _{n, d} ). May be.

  This can complicate the determination of what function of β and n is β, but formant emphasis processing that combines the advantages of Specific Example 2 and Specific Example 3 is realized.

(Specific example 5 of emphasis processing)
In step S611 of FIG. 6, the user, while giving d _em an order of the threshold emphasizing as a constant, which may be changed for each sound data.

  This is because there may be a case where it is appropriate to emphasize the mel cepstrum coefficient series data of the order of higher order for each audio data.

Incidentally, an attempt to perform this specific example, according to the step S611 of FIG. 6, but the user must give every time threshold d _em depending on every single audio data, for the processing of massive speech data, The user's burden may be reduced by automatically determining _dem as a function of n according to a predetermined rule and automating step S611.

  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, by distributing a computer program for causing a computer to execute the above-described processing through a recording medium or communication, and installing and executing the computer program on the computer, the computer can function as the speech synthesis dictionary construction device of the present invention. is there.

It is a functional block diagram of the speech synthesis dictionary construction apparatus which provided the formant emphasis part based on Embodiment 1 in the speech data processing part. It is a figure which shows the physical structure of the speech synthesis dictionary construction apparatus which concerns on Embodiment 1 and 2. FIG. It is a figure which shows the flow of operation | movement in the formant emphasis process in the audio | voice data processing part which concerns on Embodiment 1. FIG. It is a function block diagram of the speech synthesis dictionary construction | assembly apparatus which provided the formant emphasis part based on Embodiment 2 in the phoneme HMM learning part. It is a figure which shows the flow of operation | movement in the phoneme HMM learning regarding a mel cepstrum based on Embodiment 2. FIG. It is a figure which shows the flow of the operation | movement which produces | generates emphasis mel cepstrum coefficient series data.

Explanation of symbols

  111: Speech synthesis dictionary construction device according to embodiment 1, 121: input unit, 123: data writing unit, 131: speech data processing unit, 141: in-processing unit mel cepstrum analysis unit, 143... Inverse filter, 145... Formant emphasis unit in processing section, 147... Synthesis filter, 151... Phoneme HMM learning section, 161. 165 ... Mel cepstrum learning unit, 167 ... Pitch learning unit, 171 ... speech database, 173 ... speech synthesis dictionary, 211 ... computer device, 221 ... CPU, 223 ... ROM, 225 ... storage unit, 227 ... user I / F, 229 ... data input / output I / F, 231 ... bus, 241 ... RAM, 43: Hard disk, 251: Keyboard, 253: Monitor, 261: Hard disk with original data, 263: Hard disk for recording processed data, 411: Speech synthesis dictionary according to the second embodiment Construction device, 421 ... input unit, 423 ... data writing unit, 431 ... phoneme HMM learning unit, 441 ... in-learning mel cepstrum analysis unit, 443 ... pitch extraction unit, 445 ... In-learning formant emphasis unit, 447 ... mel cepstrum learning unit, 449 ... pitch learning unit, 451 ... speech database, 453 ... speech synthesis dictionary

Claims (11)

An input unit for inputting a phoneme label string and corresponding voice data;
An audio data processing unit for emphasizing spectral parameters for the audio data and converting the audio data into emphasized audio data;
A phoneme HMM learning unit that associates a phoneme HMM (Hidden Markov Model) for each phoneme label from the phoneme label string and the emphasized speech data;
A data writer for recording the learning results in the speech synthesis dictionary;
A speech synthesis dictionary construction device comprising: The voice data processing unit
An in-process mel cepstrum analysis unit that performs mel cepstrum analysis on the voice data to generate first mel cepstrum coefficient series data;
An inverse filter that is defined by the first mel cepstrum coefficient series data and generates excitation sound source data from the audio data;
A processing unit formant emphasis unit that performs a predetermined emphasis process on the first mel cepstrum coefficient series data to generate the emphasized mel cepstrum coefficient series data that emphasizes the spectral parameters;
A synthesis filter that is defined by the emphasized mel cepstrum coefficient series data and generates the formant-enhanced processed speech data by inputting the excitation sound source data;
The speech synthesis dictionary construction apparatus according to claim 1. The phoneme HMM learning unit
An in-learning mel cepstrum analysis unit that performs mel cepstrum analysis on the formant-enhanced speech data to generate second mel cepstrum coefficient series data;
A pitch extraction unit that extracts pitch series data from the formant-enhanced speech data;
A mel cepstrum learning unit that associates phoneme HMMs related to mel cepstrum for each phoneme label from the phoneme label string and the second mel cepstrum coefficient series data;
From the phoneme label string and the pitch series data, a pitch learning unit that associates a phoneme HMM related to pitch for each phoneme label;
The speech synthesis dictionary construction apparatus according to claim 2. An input unit for inputting a phoneme label string and corresponding voice data;
A mel cepstrum analysis unit in a learning unit that performs mel cepstrum analysis on the speech data to generate mel cepstrum coefficient series data;
A pitch extraction unit that extracts pitch series data from the audio data;
A formant emphasis unit in a learning unit that generates a emphasized mel cepstrum coefficient series data in which a predetermined emphasis process is performed on the mel cepstrum coefficient series data to emphasize spectrum parameters;
A mel cepstrum learning unit that associates a phoneme HMM (Hidden Markov Model) related to a mel cepstrum for each phoneme label from the phoneme label string and the emphasized mel cepstrum coefficient series data;
From the phoneme label string and the pitch series data, a pitch learning unit that associates a phoneme HMM related to pitch for each phoneme label;
A data writer for recording the learning results in the speech synthesis dictionary;
A speech synthesis dictionary construction device comprising: The predetermined emphasis process is:
The mel cepstrum coefficient series data is a process of multiplying an order higher than a predetermined order by a predetermined enhancement coefficient greater than 1.
The speech synthesis dictionary construction device according to any one of claims 2 to 4, wherein the speech synthesis dictionary construction device. The predetermined enhancement factor is:
It can be different for each voice data that is the generation source of the mel cepstrum coefficient series data multiplied by the enhancement coefficient,
The speech synthesis dictionary construction device according to claim 5. The predetermined enhancement factor is:
Depending on the order of the mel cepstrum coefficient series data multiplied by the enhancement coefficient,
The speech synthesis dictionary construction device according to claim 5 or 6. An input step in which a phoneme label string and corresponding speech data are input from the database;
An audio data processing step of enhancing spectral parameters with respect to the audio data and converting them into emphasized audio data;
A phoneme HMM learning step in which a phoneme HMM (Hidden Markov Model) is associated with each phoneme label from the phoneme label string and the emphasized speech data;
A data writing step for recording the learning results in the speech synthesis dictionary;
A speech synthesis dictionary construction method comprising: An input step in which a phoneme label string and corresponding speech data are input from the database;
A mel cepstrum analysis step in a learning unit that performs mel cepstrum analysis on the voice data to generate mel cepstrum coefficient series data;
A pitch extraction step of extracting pitch series data from the audio data;
In-learning formant emphasis step for generating emphasized mel cepstrum coefficient series data in which the mel cepstrum coefficient series data is subjected to predetermined emphasis processing to emphasize spectral parameters;
A mel cepstrum learning step for associating a phoneme HMM (Hidden Markov Model) related to a mel cepstrum for each phoneme label from the phoneme label string and the emphasized mel cepstrum coefficient series data;
From the phoneme label string and the pitch series data, a pitch learning step for associating a phoneme HMM related to pitch for each phoneme label;
A data writing step for recording the learning results in the speech synthesis dictionary;
A speech synthesis dictionary construction method comprising: On the computer,
An input step in which a phoneme label string and corresponding speech data are input from the database;
An audio data processing step of enhancing spectral parameters with respect to the audio data and converting them into emphasized audio data;
A phoneme HMM learning step in which a phoneme HMM (Hidden Markov Model) is associated with each phoneme label from the phoneme label string and the emphasized speech data;
A data writing step for recording the learning results in the speech synthesis dictionary;
A computer program that executes On the computer,
An input step in which a phoneme label string and corresponding speech data are input from the database;
A mel cepstrum analysis step in a learning unit that performs mel cepstrum analysis on the voice data to generate mel cepstrum coefficient series data;
A pitch extraction step of extracting pitch series data from the audio data;
In-learning formant emphasis step for generating emphasized mel cepstrum coefficient series data in which the mel cepstrum coefficient series data is subjected to predetermined emphasis processing to emphasize spectral parameters;
A mel cepstrum learning step for associating a phoneme HMM (Hidden Markov Model) related to a mel cepstrum for each phoneme label from the phoneme label string and the emphasized mel cepstrum coefficient series data;
From the phoneme label string and the pitch series data, a pitch learning step for associating a phoneme HMM related to pitch for each phoneme label;
A data writing step for recording the learning results in the speech synthesis dictionary;
A computer program that executes

Images