JP5228283B2 - Speech synthesis dictionary construction device, speech synthesis dictionary construction method, and program - Google Patents

Description

  The present invention relates to a speech synthesis dictionary construction apparatus and a speech synthesis dictionary construction method for constructing a database used for speech synthesis by referring to a given speech database.

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

  Examples of documents related to the speech recognition technology and speech synthesis technology based on the HMM are those described in Patent Documents 1 to 3.

JP 2002-62890 A Japanese Patent Laid-Open No. 2002-244689 JP 2002-268660 A

  In speech recognition and speech synthesis based on the HMM, a speech synthesis dictionary that records the correspondence between phoneme labels and spectral parameter data strings 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 a speech synthesis dictionary is constructed, the speech synthesis parameter dictionary is constructed by using the speech spectrum parameter data string calculated from the speech data as it is for learning based on the HMM without any particular processing.

  Also, the learning process based on HMM often includes a single or multiple re-learning processes for improving the likelihood.

  When the learning process based on the HMM includes a learning process of a plurality of stages as described above, the phoneme HMM for the phoneme label is determined for each stage, and the corresponding relationship is transmitted to the next learning stage, and the learning proceeds. become.

  Conventionally, the phoneme HMM generated at each stage has been sent to the next stage without any special processing.

  The speech synthesis dictionary constructed in this way is used in a speech synthesizer.

  The LSP coefficient group time-series data generated by performing the LSP analysis on the speech data and the parameters defining the phoneme HMM related to the LSP coefficients as a result of the phoneme HMM learning, for example, the average value may be disturbed.

  The conventional speech synthesis dictionary construction method ignores the disturbance of the LSP coefficient group time series data and proceeds to the phoneme HMM learning process, or ignores the disturbance of the parameter defining the phoneme HMM related to the LSP coefficient and advances the learning. ing. For this reason, in the conventional speech synthesis dictionary construction method, a speech synthesis dictionary with insufficient quality may be constructed.

  The present invention has been made in view of the above circumstances, and an object of the present invention is to provide a speech synthesis dictionary construction apparatus and method capable of synthesizing high-quality text speech.

The speech synthesis dictionary construction apparatus according to the present invention is
An LSP coefficient group time series data generating unit that performs LSP analysis on voice data and generates LSP coefficient group time series data including multidimensional LSP (Line Spectrum Pair) coefficients;
A pre-learning spectral parameter correction unit that performs a correction process on the LSP coefficient group time-series data generated by the LSP coefficient group time-series data generation unit so as to satisfy a predetermined stability condition;
The phoneme label sequence and the corrected LSP coefficient group time-series data output from the pre-learning spectrum parameter correction unit are received, and phoneme HMM learning is performed in a plurality of stages by associating the phoneme HMM for each phoneme label by learning based on the hidden Markov model. A phoneme HMM learning unit;
The LSP coefficient defining the phoneme HMM associated with each phoneme label obtained in the phoneme HMM learning at each stage of the phoneme HMM learning unit is subjected to correction processing so as to satisfy a predetermined stability condition. A learning spectral parameter correction unit that is handed over to the phoneme HMM learning of
A data writing unit for recording a phoneme label and a phoneme HMM associated by the phoneme HMM learning unit in a speech synthesis dictionary;
It is characterized by providing.

  According to the present invention, since the parameters defining the phoneme HMM related to the LSP coefficient group time series data and the LSP coefficients are corrected, a speech synthesis dictionary having appropriate data is constructed. As a result, a speech synthesizer that synthesizes speech by referring to the speech synthesis dictionary can emit high-quality synthesized speech.

  Hereinafter, a speech synthesis dictionary construction apparatus and method according to an embodiment of the present invention will be described.

(Embodiment 1)
FIG. 1 is a schematic configuration diagram of a speech synthesis dictionary construction device 11 having a spectrum parameter correction function according to the first embodiment.

  The speech synthesis dictionary construction device 11 includes a data extraction unit 13, a spectrum analysis unit 15, a pre-learning spectrum parameter correction unit 17, a phoneme HMM learning unit 19, and a data writing unit 29.

  The phoneme HMM learning unit 19 includes a first phoneme HMM learning unit 21, a learning spectral parameter correction unit 23, and a second phoneme HMM learning unit 25 to an Nth phoneme HMM learning unit 27.

  The speech synthesis dictionary construction device 11 is connected to the speech database 31 and the speech synthesis dictionary 33.

  The voice database 31 is a database composed of a set of phoneme label strings and corresponding voice data, and is stored in, for example, a hard disk device.

  The speech synthesis dictionary 33 is a database that stores phoneme HMMs for each phoneme label generated by the speech synthesis dictionary construction device 11, and is stored in, for example, a hard disk. This speech synthesis dictionary 33 is used for speech synthesis.

  The data extraction unit 13 reads a set of a voice label string and voice data from the voice database 31 and separates it into a phoneme label string and voice data. The data extraction unit 13 delivers the phoneme label string to the first phoneme HMM learning unit 21 and delivers the speech data to the spectrum analysis unit 15.

  The spectrum analysis unit 15 analyzes the voice data delivered from the data extraction unit 13, generates a voice spectrum parameter data string representing the spectrum envelope of the voice data, and delivers it to the pre-learning spectrum parameter correction unit 17. In the present embodiment, the speech spectrum parameter data sequence is an LSP coefficient sequence.

  The pre-learning spectrum parameter correction unit 17 performs a correction operation on the speech spectrum parameter data sequence delivered from the spectrum analysis unit 15 so as to satisfy a predetermined stability condition, and converts the corrected speech spectrum parameter data sequence into the first phoneme HMM. This is supplied to the learning unit 21.

  The predetermined stability condition is that LSP coefficients in the same frame are all larger than 0 and smaller than π, and are arranged in ascending order when the coefficients are arranged in ascending order. That is, in principle, when LSP coefficients of the same frame are connected by lines along the time axis between LSP coefficient groups of adjacent frames, the lines do not cross each other, and these lines do not intersect with LSP. The region between 0 and π for the coefficient does not protrude.

  Here, the term “frame” refers to a cycle in which audio data is LSP analyzed.

  The first phoneme HMM learning unit 21 uses the group of pairs of the phoneme label sequence supplied from the data extraction unit 13 and the corrected speech spectrum parameter data sequence supplied from the pre-learning spectrum parameter correction unit 17 as a phoneme HMM. Based on the learning, a phoneme HMM is constructed for each phoneme label, and these are supplied to the in-learning spectral parameter correction unit 23.

  The second phoneme HMM learning unit 25 to the Nth phoneme HMM learning unit 27 re-learns the corrected phoneme HMM supplied from the learning spectral parameter correction unit 23 based on the phoneme HMM, and learns these spectral parameter correction units. 23.

  The in-learning spectral parameter correction unit 23 satisfies a predetermined stability condition with respect to the average value among the parameters defining the phoneme HMM supplied from the first phoneme HMM learning unit 21 to the Nth phoneme HMM learning unit 27. Make corrections. Thereafter, the in-learning spectral parameter correction unit 23 supplies the phoneme HMM to the next phoneme HMM learning unit. That is, the in-learning spectral parameter correction unit 23 calculates the average value of the parameters defining the phoneme HMM supplied from the first to (N-1) th phoneme HMM learning units, and the second to Nth phoneme HMM learning. Supply to the department. The in-learning spectral parameter correction unit 23 supplies the corrected phoneme HMM supplied from the Nth phoneme HMM learning unit 27 to the data writing unit 29.

  The data writing unit 29 records the phoneme HMM in the speech synthesis dictionary 33.

  The speech synthesis dictionary construction device 11 shown in FIG. 1 is physically configured by a general computer device 41 as shown in FIG. The CPU 43, ROM 45, storage unit 47, data input / output I / F 51, and user I / F 49 are connected to each other via a bus 71.

  The ROM 45 stores an operation program for learning based on the HMM, and in particular, in this embodiment, an operation program including an operation for correcting a spectrum parameter.

  The storage unit 47 includes a RAM 65 and a hard disk 67 and stores learning constants, phoneme label strings, speech data, spectrum parameter data, and correspondences between phoneme labels and spectrum parameter data strings.

  The data input / output I / F 51 is an interface for connecting to the original data-containing hard disk 55 and the like and the processed data recording hard disk 57 and the like.

  The data input / output I / F 51 is connected to the speech database 31 shown in FIG. 1, reads out a pair of phoneme label strings to be learned and speech data, and stores them in the storage unit 37 under the control of the CPU 43 shown in FIG. To do.

  The data input / output I / F 51 is connected to the speech synthesis dictionary 33 shown in FIG. 1, and outputs the phoneme HMM, which is the result of the processing by the CPU 43 shown in FIG. 2, to the speech synthesis dictionary 33 shown in FIG.

  The user I / F 49 shown in FIG. 2 includes a keyboard 61 and a monitor 63, and is provided for inputting arbitrary instructions, data, and programs.

  The CPU 43 executes a speech synthesis dictionary generation operation by executing an operation program stored in the ROM 45.

  As shown in FIG. 1, the feature of the speech synthesis dictionary construction device 11 according to the present embodiment is that a predetermined correction process is performed so that one or two or more spectral parameter correction units are provided so that the spectral parameter data satisfies a predetermined stability condition. Is to do.

  The correction process executed by the pre-learning spectral parameter correction unit 17 and the in-learning spectral parameter correction unit 23 may be any operation as long as the operation satisfies a predetermined stability condition. Hereinafter, in order to facilitate understanding, a specific operation of the speech synthesizer will be described with reference to specific correction processing.

  First, the data extraction unit 13 sequentially extracts pairs of phoneme labels and audio data stored in the audio database 31, separates them into phoneme labels and audio data, and stores the phoneme label string in the first learning unit 21. Data is supplied to the spectrum analyzer 15.

  The spectrum analysis unit 15 analyzes the supplied voice data by a known method, sequentially generates LSP coefficients, and supplies the LSP coefficients to the pre-learning spectrum parameter correction unit 17.

  The correction operation performed by the pre-learning spectrum parameter correction unit 17 will be described with reference to the flowchart shown in FIG.

First, N _Sp pieces of audio data are extracted and stored (step S301), and the m-th audio data Sp _m is specified from among them (step S303). Next, the LSP coefficient group ω _{m, k} [fm] of the fm-th frame (step S305) is extracted from the LFM coefficient group of N _fm frames of the m-th audio data Sp _m (where 1 ≦ k ≦ N _d and N _d are the orders of the LSP coefficients.) (Step S307). Next, the extracted LSP coefficient ω _{m, k} [fm] (where 1 ≦ k ≦ N _d , 0 ≦ fm ≦ N _fm [m], N _d is the order of the LSP coefficient, and N _fm [m] is the voice data. in the number of frames) with respect sp _m, it determines whether a predetermined stability condition of the LSP coefficients is satisfied (step S309).

Here, the predetermined stability condition of the LSP coefficient is
0 <ω _{m, 1} [fm] <ω _{m, 2} [fm] <...... <ω _{m, Nd} [fm] <π
It is.

Incidentally, the LSP coefficients omega _{m, k,} for each frame fm, i.e., determines whether or meets the stability condition for each LSP coefficients ω _m, k [fm]. That is, a loop is set such that fm is scanned in the range of 0 ≦ fm ≦ N _fm [m] (step S315; fm ← fm + 1 if not completed).

  When the predetermined stability condition is satisfied (step S309; satisfied), the LSP coefficient confirmation flag cf [fm] is set to 1 (cf [fm] = 1) (step S311). Step S309; not satisfied), the LSP coefficient confirmation flag cf [fm] is set to 0 (cf [fm] = 0) (Step S313).

When fm has been scanned to N _fm [m] (step S315; completion), it is then determined whether or not there is an LSP coefficient confirmation flag that satisfies cf [fm _NG ] = 0 (step S317).

If there is no LSP coefficient confirmation flag for cf [fm _NG ] = 0 (step S317; does not exist), the correction operation for the m-th audio data Sp _m ends.

If there is an LSP coefficient confirmation flag for cf [fm _NG ] = 0 (step S317; present), the LSP coefficient ω _{m, k} [fm for the frame fm _NG for cf [fm = fm _NG ] = 0. the _NG], corrected (step S321, S323, S325), and terminates the correction operation for the m-th audio data Sp _m.

The contents of correction will be described. First, it is determined whether fm _NG is 0, 0 <fm _NG <N _fm [m], or N _fm [m] (step S319).

If it is determined in step S319 that fm _NG = 0, correction is performed with reference to the LSP coefficients of the second and subsequent frames (step S321). For example, ω _{m, k} [fm _NG ] = ω _{m, k} [fm _{OK, H} ] (where 1 ≦ k ≦ N _d , fm _{OK, H} is larger than fm _NG , cf [fm _{OK, H} ]) = 1 is the minimum value satisfying 1).

If it is determined in step S319 that 0 <fm _NG <N _fm [m], refer to the LSP coefficient of the frame adjacent to fm _NG such that 0 ≦ fm ≦ N _fm [m]. Then, correction is performed (step S323). For example, ω _{m, k} [fm _NG ] = ω _{m, k} [fm _{OK, L} ] × α + ω _{m, k} [fm _{OK, H} ] × β (where 1 ≦ k ≦ N _d , fm _{OK, L} Is a maximum value smaller than fm _NG and satisfies cf [fm _{OK, L} ] = 1, fm _{OK, H} is a minimum value larger than fm _NG and satisfies cf [fm _{OK, H} ] = 1, and α and β are weighting factors ).

When it is determined in step S21 that fm _NG = N _fm [m], correction is performed with reference to the second and previous frames from the tail (step S325). For example, ω _{m, k} [fm _NG ] = ω _{m, k} [fm _{OK, L} ] (where 1 ≦ k ≦ N _d , fm _{OK, L} is smaller than fm _NG , cf [fm _{OK, L} ]) = 1 is the maximum value satisfying 1).

  The pre-learning spectrum parameter correction unit 17 repeats the above correction process while sequentially updating the audio data and the frame (step S315; fm ← fm + 1 if not completed, step S327; m ← m + 1 if not completed). On the other hand, when the processing for all the frames of all the audio data is completed as a result of the continuation of the update, the process exits the loop (step S315; completion, step S327; completion).

  The pre-learning spectrum parameter correction unit 17 supplies the corrected speech spectrum parameter data string (corrected LSP coefficient group string) to the first phoneme HMM learning unit 21.

  The first phoneme HMM learning unit 21 associates the phoneme label with the corrected speech spectrum parameter data string (corrected LSP coefficient group string), and performs learning based on the phoneme HMM. As the learning method itself, a known arbitrary method can be adopted.

  The in-learning spectral parameter correction unit 23 performs correction processing shown in a specific example described later on the phoneme HMM supplied from the first phoneme HMM learning unit 21 and supplies the phoneme HMM to the next phoneme HMM learning unit.

  Thereafter, the same processing is repeated, and the final phoneme HMM is supplied to the data writing unit 29 and written into the speech synthesis dictionary 33.

  In the speech synthesis dictionary construction device of the present embodiment, the pre-learning spectral parameter correction unit 17 and the in-learning spectral parameter correction unit 23 make corrections so as to satisfy a predetermined stability condition each time a spectral parameter is generated. Therefore, the construction of a speech synthesis dictionary that contributes to outputting synthesized speech with higher sound quality can be achieved.

Up to this point, it has been assumed that when cf [fm _NG ] = 0, correction of ω _{m, k} [fm _NG ] is performed for all orders k satisfying 1 ≦ k ≦ N _d. The correction of ω _{m, k} [fm _NG ] may be performed only for the order k that caused the condition not to be satisfied.

In such a case, the LSP coefficient is corrected for the required order k and the stability condition is satisfied. However, unlike the case where all LSP coefficients belonging to the frame fm _NG are to be corrected, it is inherently stable. Since the LSP coefficient for the order k that satisfies the condition does not need to be subjected to extra correction, a more appropriate correction is realized.

(Specific example of correction during learning 1)
A specific example of the inside of the phoneme HMM learning unit 19 in the speech synthesis dictionary construction device 11 according to the first embodiment will be described with reference to the flowchart shown in FIG.

  The speech synthesis dictionary construction device according to this example is characterized in that the phoneme HMM learning unit includes first to fifth phoneme HMM learning units and a learning spectral parameter correction unit.

  The first to fifth phoneme HMM learning units respectively perform initialization learning of the monophone HMM, re-learning of the monophone HMM, initialization learning of the triphone HMM, re-learning of the triphone HMM, and clustering processing using a decision tree Take on.

An HMM with 5 states is used as the monophone HMM. State S ₀ is an initial state and state S ₄ is an end state, and neither state outputs an LSP coefficient. LSP coefficients are output from states S ₁ , S ₂ , S ₃ .

The first phoneme HMM learning unit uses the phoneme label string and the speech spectrum parameter data string ω _{m, k} [fm] as learning data to initialize and learn the monophone HMM related to the LSP coefficient for each phoneme label (step S401).

  The learning result is delivered to the in-learning spectrum parameter correction unit 23 shown in FIG.

In the learning spectral parameter correction unit, the average value ω _{k, Ave} [S _i ] of the LSP coefficients in each state (where i is 1 to 3 and 1 ≦ 1) with respect to the monophone HMM related to the LSP coefficient of each phoneme label k ≦ N _d , where N _d is the order of the LSP coefficient.) is the stability condition 0 <ω _{1, Ave} [S _i ] <ω _{2, Ave} [S _i ] <…… <ω _{Nd, Ave} [S _i ] <π is discriminated (step S425). Since it is necessary to perform discrimination for all of S ₁ , S ₂ , and S ₃ , processing is sequentially performed using the counter i (steps S 423 and S 429).

If the stability condition is satisfied (step S425; satisfied), the correction operation is ended if the processing for all of the states S ₁ , S ₂ , and S ₃ is completed (step S429; YES).

  If the stability condition is not satisfied (step S425; not satisfied), an operation of correcting to satisfy is performed (step S427).

  The result of the correction operation is delivered to the second phoneme HMM learning unit.

  The second phoneme HMM learning unit relearns the monophone HMM related to the LSP coefficient for each phoneme label.

  The learning result is transferred to the in-learning spectrum parameter correction unit 23.

  The in-learning spectral parameter correction unit 23 performs the same correction operation as described above, and passes the result to the third phoneme HMM learning unit.

  The third phoneme HMM learning unit performs initialization learning of the triphone HMM. That is, the monophone HMM related to the LSP coefficient is copied to the triphone HMM related to the LSP coefficient considering the preceding and following phoneme labels, and is subjected to initialization learning.

Similarly to the monophone HMM, the triphone HMM is an HMM having five states, and the average value of the LSP coefficients in each state in the HMM is represented by ω _{k, Ave} [S _i ].

  The learning result is transferred to the in-learning spectrum parameter correction unit 23.

  The in-learning spectral parameter correction unit 23 performs the same correction operation as described above, and passes the result to the fourth phoneme HMM learning unit.

  The fourth phoneme HMM learning unit performs relearning of the triphone HMM.

  The learning result is transferred to the in-learning spectrum parameter correction unit 23.

  The in-learning spectral parameter correction unit 23 performs the same correction operation as described above, and passes the result to the fifth phoneme HMM learning unit.

  The fifth phoneme learning unit performs clustering using a decision tree on the triphone HMM related to the LSP coefficient, generates a model that can also deal with combinations of phoneme labels that do not exist in the learning data, and again generates the LSP coefficient. Re-learn triphone HMM for

  The learning result is transferred to the in-learning spectrum parameter correction unit 23.

  The in-learning spectral parameter correction unit performs a correction operation similar to the above-described operation, and delivers the result to the data writing unit 29 shown in FIG.

  In this manner, in the phoneme HMM learning unit, every time learning is completed, the LSP coefficient correction operation is performed by the in-learning spectral parameter correction unit. Below, the specific content of the said correction operation is demonstrated with reference to the flowchart shown in FIG.

The average values ω _{k and Ave} [S _i ] of the LSP coefficients in each state of the monophone HMM or the triphone HMM are read (step S501), and for these, the stability condition 0 <ω _{1 and Ave} [S _i ] <ω _{2 , Ave} [S _i ] <... <Ω _{Nd, Ave} [S _i ] <π is determined.

If the stability condition for all states _{S 1} to _{S 3} is satisfied (step S503; YES, Step S505; YES, Step S509; YES), without correction, and terminates the correction operation.

Although the state _{S 1} is not satisfied the stability condition (step S503; NO), if it is satisfied for the _{S 2} and _{S 3} (step S507; YES, Step S513; YES), the state _{S 1} For the mean value ω _{k, Ave} [S ₁ ] of the LSP coefficients,
ω _{k, Ave} [S ₁ ] = ω _{k, Ave} [S ₂ ] (where 1 ≦ k ≦ N _d )
Is corrected (step S521), LSP coefficients for the state _{S 2} and _{S 3} as without correction, and terminates the correction operation.

Although the state _{S 2} has not been met the stability criteria, if it is satisfied for _{S 1} and _{S 3} (step S503; YES, Step S505; NO, step S511; YES), the state _{S 2} LSP For the mean value ω _{k, Ave} [S ₂ ] of the coefficients,
ω _{k, Ave} [S ₂ ] = (ω _{k, Ave} [S ₁ ] + ω _{k, Ave} [S ₃ ]) / 2 (where 1 ≦ k ≦ N _d ) (step S519) The correction operation is terminated without correcting the LSP coefficients related to S ₁ and S ₃ .

Although the state _{S 3} is not the stability condition is satisfied, if it is satisfied for _{S 1} and _{S 2} (step S503; YES, Step S505; YES, Step S509; NO) is, LSP state _{S 3} For the mean value ω _{k, Ave} [S ₃ ] of the coefficients,
ω _{k, Ave} [S ₃ ] = ω _{k, Ave} [S ₂ ] (where 1 ≦ k ≦ N _d )
Is corrected (step S517), LSP coefficients for the state _{S 1} and _{S 2} as uncorrected, and terminates the correction operation.

Although the state _{S 1} and the state _{S 2} is not the stability condition is satisfied, if it is satisfied for _{S 3} (step S503; NO, step S507; NO, step S515; YES), the state _{S 1} and For the mean value ω _{k, Ave} [S ₁ ] and ω _{k, Ave} [S ₂ ] of the LSP coefficients in state S ₂ ,
ω _{k, Ave} [S ₁ ] = ω _{k, Ave} [S ₃ ], ω _{k, Ave} [S ₂ ] = ω _{k, Ave} [S ₃ ] (where 1 ≦ k ≦ N _d )
Is corrected (step S527), LSP coefficients for the state _{S 3} as without correction, and terminates the correction operation.

Although for the state _{S 1} and the state _{S 3} not the stability condition is satisfied, if it is satisfied for _{S 2} (step S503; NO, step S507; YES, Step S513; NO), the state _{S 1} and For the mean value ω _{k, Ave} [S ₁ ] and ω _{k, Ave} [S ₃ ] of the LSP coefficients in state S ₃ ,
ω _{k, Ave} [S ₁ ] = ω _{k, Ave} [S ₂ ], ω _{k, Ave} [S ₃ ] = ω _{k, Ave} [S ₂ ] (where 1 ≦ k ≦ N _d )
Is corrected (step S525), LSP coefficients for the state _{S 2} as uncorrected, and terminates the correction operation.

Although for the state _{S 2,} and the state _{S 3} not the stability condition is satisfied, if it is satisfied for _{S 1} (step S503; YES, Step S505; NO, step S511; NO), the state _{S 2} and For the mean value ω _{k, Ave} [S ₂ ] and ω _{k, Ave} [S ₃ ] of LSP coefficients in state S ₃ ,
ω _{k, Ave} [S ₂ ] = ω _{k, Ave} [S ₁ ], ω _{k, Ave} [S ₃ ] = ω _{k, Ave} [S ₁ ] (where 1 ≦ k ≦ N _d )
Is corrected (step S523), LSP coefficients for the state _{S 1} as without correction, and terminates the correction operation.

If the stability condition for the state _{S 1} to _{S 3} is not satisfied (step S503; NO, step S507; NO, step S515; NO) is returned to the pre-learning state (step S529), ends the correction operation To do.

(Specific example 2 of correction during learning)
In the specific example, the correction operation performed when the stability condition is not satisfied is performed for all orders k of the LSP coefficient, that is, all k satisfying 1 ≦ k ≦ N _d . On the other hand, in the present embodiment, only the order k that causes the stability condition not to be satisfied is performed.

  By limiting the correction target in this way, it is possible to prevent the correction from being performed up to the order k that originally did not need to be corrected. As a result, the speech synthesis dictionary construction device according to the fourth embodiment The construction of a speech synthesis dictionary that contributes to the output of synthesized speech with higher sound quality than the constructed speech synthesis dictionary is achieved.

(Specific example 3 of correction during learning)
In the present embodiment, in addition to the correction of the LSP coefficient in the specific example 1, a predetermined appropriateness determination criterion is imposed on the variance value ω _{k, Var} [S _i ] of the LSP coefficient in each state in the HMM, For the variance value of the order k having an inappropriate value, an operation for correcting to an appropriate value is added.

(Embodiment 2)
The speech synthesis dictionary construction device according to the present embodiment is a speech synthesis dictionary construction device in which the pre-learning spectrum parameter correction unit is omitted in the speech synthesis dictionary construction device according to the first embodiment.

  Since the correction is repeated by the in-learning spectral parameter correction unit 23 inside the phoneme HMM learning unit 19, when the speech synthesis dictionary constructed by the speech synthesis dictionary construction device according to the present embodiment is used for the speech synthesis device, Compared to the case where the speech synthesis dictionary constructed by the speech synthesis dictionary construction device is used in the speech synthesis device, the output of the synthesized speech with high sound quality is achieved.

  Since the pre-learning spectrum parameter correction unit is omitted, when the speech synthesis dictionary constructed by the speech synthesis dictionary construction device according to the present embodiment is used for the speech synthesis device, the speech synthesis dictionary construction device according to the first embodiment Compared with the case where the speech synthesis dictionary constructed by the above is used in a speech synthesizer, the quality of the synthesized speech output is inferior.

  However, by omitting the pre-learning spectrum parameter correction unit, the speech synthesis dictionary construction device according to the present embodiment has an advantage that the structure is simpler than the speech synthesis dictionary construction device according to the first embodiment. .

  In addition, this invention is not limited to rule 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 schematic block diagram of the speech synthesis dictionary construction apparatus and method by this invention. 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 flowchart which shows the flow of operation | movement in the correction before learning. It is a flowchart which shows the flow of learning in a 1st phoneme HMM learning part. It is a flowchart which shows the flow of operation | movement in correction | amendment during learning.

Explanation of symbols

DESCRIPTION OF SYMBOLS 11 ... Speech synthesis dictionary construction apparatus, 13 ... Data extraction part, 15 ... Spectrum analysis part, 17 ... Pre-learning spectrum parameter correction part, 19 ... Phoneme HMM learning part, 21 ... 1st phoneme HMM learning unit, 23... Learning spectral parameter correction unit, 25... 2nd phoneme HMM learning unit, 27... Nth phoneme HMM learning unit, 29. .. Voice database, 33... Speech synthesis dictionary, 41... Computer device, 43... CPU, 45... ROM, 47. Data input / output I / F, 55: Hard disk with original data, 57: Hard disk for recording processed data, 61: Keyboard, 63: Monitor, 65 ... RAM, 67・Dodisuku, 71 ... bus

Claims (5)

An LSP coefficient group time series data generating unit that performs LSP analysis on voice data and generates LSP coefficient group time series data including multidimensional LSP (Line Spectrum Pair) coefficients;
A pre-learning spectral parameter correction unit that performs a correction process on the LSP coefficient group time-series data generated by the LSP coefficient group time-series data generation unit so as to satisfy a predetermined stability condition;
The phoneme label sequence and the corrected LSP coefficient group time-series data output from the pre-learning spectrum parameter correction unit are received, and phoneme HMM learning is performed in a plurality of stages by associating the phoneme HMM for each phoneme label by learning based on the hidden Markov model. A phoneme HMM learning unit;
The LSP coefficient defining the phoneme HMM associated with each phoneme label obtained in the phoneme HMM learning at each stage of the phoneme HMM learning unit is subjected to correction processing so as to satisfy a predetermined stability condition. A learning spectral parameter correction unit that is handed over to the phoneme HMM learning of
A data writing unit for recording a phoneme label and a phoneme HMM associated by the phoneme HMM learning unit in a speech synthesis dictionary;
A speech synthesis dictionary construction device comprising: The correction process, for LSP coefficient group, to determine whether the predetermined stability condition is satisfied, if not satisfied, and a process of replacing the LSP coefficient group that satisfies predetermined stability condition ,
The speech synthesis dictionary construction apparatus according to claim 1. The correction process includes a process of replacing the LSP coefficient group with a coefficient group that satisfies the predetermined stability condition,
The predetermined stability condition is that LSP coefficients are all larger than 0 and smaller than π, and are arranged in ascending order when the coefficients are arranged in ascending order.
The speech synthesis dictionary construction apparatus according to claim 1. LSP analysis is performed on voice data to generate LSP coefficient group time series data including multidimensional LSP (Line Spectrum Pair) coefficients,
A correction process is performed on the generated LSP coefficient group time series data so as to satisfy a predetermined stability condition,
Receiving a phoneme label sequence and the corrected LSP coefficient group time-series data, and performing phoneme HMM learning to associate a phoneme HMM for each phoneme label by learning based on a hidden Markov model;
In the phoneme HMM learning, the LSP coefficient defining the phoneme HMM associated with each phoneme label obtained by the phoneme HMM learning in each step is subjected to correction processing so as to satisfy a predetermined stability condition, and the next step Handed over to phoneme HMM learning
Record the associated phoneme label and the phoneme HMM in the speech synthesis dictionary;
A speech synthesis dictionary construction method characterized by that. On the computer,
LSP analysis is performed on voice data to generate LSP coefficient group time series data including multidimensional LSP (Line Spectrum Pair) coefficients,
A correction process is performed on the generated LSP coefficient group time series data so as to satisfy a predetermined stability condition,
Receiving a phoneme label sequence and the corrected LSP coefficient group time-series data, and performing phoneme HMM learning to associate a phoneme HMM for each phoneme label by learning based on a hidden Markov model;
In the phoneme HMM learning, the LSP coefficient defining the phoneme HMM associated with each phoneme label obtained by the phoneme HMM learning in each step is subjected to correction processing so as to satisfy a predetermined stability condition, and the next step Handed over to phoneme HMM learning
Record the associated phoneme label and the phoneme HMM in the speech synthesis dictionary;
A computer program for executing processing.

Images