JP4816201B2 - Speech processing apparatus and method, text speech synthesis apparatus, and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To output a synthesized speech of high quality even when an LSP coefficient group does not meet stableness conditions any more. <P>SOLUTION: Time-series data of the LSP coefficient group are generated and if an abrupt LSP coefficient is obtained and the stableness conditions are not met any more, a correcting method is switched according to the position of the frame where the abrupt LSP coefficient is output while other LSP coefficients corresponding to the same phoneme HMM and the same state of a hidden Markov model is referred to, thereby minimizing deterioration in speech quality. <P>COPYRIGHT: (C)2008,JPO&amp;INPIT

Description

The present invention relates to a speech processing apparatus and method, a text-to-speech synthesizer, and a program that are applied in the process of creating speech data from a given text character string.
The present invention also relates to a speech synthesizer that creates speech data from a given text string.

  Speech recognition technology and speech synthesis technology based on Hidden Markov Models (hereinafter HMM) have been very successful.

A speech recognition technique and a speech synthesis technique based on the HMM are disclosed in, for example, Patent Document 1.
JP 2002-62890 A

  As a spectral parameter used in the speech synthesis method based on the HMM, there is an LSP (Line Spectrum Pair) coefficient group.

A group of LSP coefficients is arranged for each time segment called a frame. In the case of using an m-dimensional feature vector in adopting the HMM, a group of LSP coefficient groups is composed of m LSP coefficients corresponding to each dimension. In addition, all LSP coefficients always exist between 0 and π. Further, if the m-dimensional LSP coefficient in the frame fm is expressed by ω ₁ [fm], ω ₂ [fm],..., Ω _m [fm] in order from the first dimension, ω ₁ [fm ] <Ω ₂ [fm] <... <Ω _m [fm].

  The above-described limitation on the existence range of the LSP coefficient and limitation on the magnitude relationship are hereinafter referred to as a stable condition.

  Usually, the stability condition is satisfied. 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.

  However, in reality, an LSP coefficient that does not satisfy the stability condition may be obtained in the process of maximizing the likelihood of the phoneme HMM sequence in units of sentences, for example.

  Hereinafter, this case is referred to as “a sudden LSP coefficient has been obtained”.

  If sudden LSP data is left unattended, the quality of the synthesized speech deteriorates.

  Therefore, Patent Document 1 suggests a method for suppressing the deterioration of voice when a sudden LSP coefficient is obtained. This method uses the LSP coefficients of the frames adjacent to the frame to which the sudden LSP coefficient is assigned, calculates an average value for each dimension, and replaces the average value with the sudden LSP coefficient. Correct to meet the stability condition.

  Hereinafter, such replacement processing is referred to as simple adjacent average interpolation processing.

In the simple adjacent average interpolation processing, ω _d [fm] = (ω _d [fm−) for an arbitrary dimension d (1 ≦ d ≦ m) regardless of the frame fm corresponding to the sudden LSP coefficient. 1] + ω _d [fm + 1]) / 2, an average value is calculated for each dimension using adjacent LSP coefficient groups, and the average value is interpolated.

  However, in the simple adjacent average interpolation processing, even if the stability condition of the LSP coefficient is satisfied, if the difference between adjacent LSP coefficients of the same dimension is large, interpolation may be performed with an inappropriate LSP coefficient. In particular, when the state is switched in the HMM, the difference between adjacent LSP coefficients may increase. In such a case, deterioration of the synthesized speech may not be suppressed.

  In addition, in the simple adjacent average interpolation process, LSP coefficients other than the order that caused the failure to satisfy the stability, that is, LSP coefficients that were originally not sudden and need not be interpolated are also interpolated. Therefore, the quality of the synthesized speech may be degraded.

  The present invention has been made in view of the above-described drawbacks of the prior art, and provides a speech processing device and a text speech synthesizer capable of generating high-quality synthesized speech by performing appropriate correction processing. For the purpose.

  Another object of the present invention is to enable generation of high-quality synthesized speech.

In order to achieve the above object, a speech processing apparatus according to a first aspect of the present invention includes a phoneme label string conversion unit that converts a given text character string into a phoneme label string, and an output from the phoneme label string conversion unit. Time-series data generating means for generating time-series data of LSP coefficient groups including multi-dimensional LSP (Line Spectrum Pair) coefficients from the phoneme label sequence thus generated, and the LSP coefficient group time generated by the time-series data generating means For each LSP coefficient group constituting the series data, it is determined whether or not a predetermined stability condition is satisfied. When it is determined that the predetermined stability condition is not satisfied, the LSP coefficient group satisfies the predetermined stability condition. Depending on whether the phoneme HMM (Hidden Markov Model) that output the LSP coefficient group is in the 1 state, the correction method is switched depending on whether it is located at the head, tail, or other position. Instead, the speech processing apparatus includes a correcting unit that corrects the LSP coefficient group, wherein the predetermined stability condition is that a plurality of LSP coefficients constituting the individual LSP coefficient group are all greater than zero. and smaller than [pi, and, when arranged in ascending order of the dimensions of the LSP coefficients, I der be arranged in ascending order, the correction means 1 of the LSP coefficient group said predetermined stable condition is not met When the first position is located in the first state, at least the second LSP coefficient group is corrected in the state using the LSP coefficient group located second, and when the last position is located in the first state, at least When the LSP coefficient group located at the second position from the tail in that state is used to correct the one LSP coefficient group and the other one is located in the first state, at least The one LSP coefficient group is corrected using adjacent LSP coefficient groups in the state of (1 ).

In order to achieve the above object, a text-to-speech synthesizer according to a second aspect of the present invention supplies text data to be synthesized to the speech processing apparatus and the phoneme label string conversion means of the speech processing apparatus. means, and synthetic speech generating means to LSP synthesis filter by entering an excitation sound source data for generating a synthesized speech time series data LSP coefficient group was LSP synthesis filter coefficients generated by the audio processing apparatus, Ru provided with.

In order to achieve the above object, a speech processing method according to a third aspect of the present invention includes a phoneme label string conversion step for converting a text character string into a phoneme label string, and an LSP (Line Spectrum Pair) from the phoneme label string. ) A time series data generation step of generating time series data of LSP coefficient groups including coefficients, and when a predetermined stability condition is not satisfied for each LSP coefficient group constituting the LSP coefficient group time series data, In order for this LSP coefficient group to satisfy the predetermined stability condition, the correction method is switched by switching the correction method depending on whether the LSP coefficient group is located at the head, tail, or other position in the state of the phoneme HMM that has output the LSP coefficient group. Correcting the coefficient group, wherein the predetermined stability condition is a plurality of L constituting the individual LSP coefficient group. P coefficients are all smaller than larger and π than 0, and, when arranged in ascending order of the dimensions of the LSP coefficients, it der be arranged in ascending order, in the step of correcting said predetermined stable condition met If one LSP coefficient group that has not been placed is at the head in the state 1, the one LSP coefficient group is corrected using at least the second LSP coefficient group in the state, and the state 1 If the LSP coefficient group is corrected at least using the LSP coefficient group located second from the tail in that state, and if it is located in the other state, at least The one LSP coefficient group is corrected using LSP coefficient groups adjacent in the state .

In order to achieve the above object, a computer program according to a fourth aspect of the present invention provides a computer with a phoneme label string conversion step for converting a text character string into a phoneme label string, and from the phoneme label string, an LSP (Line (Spectrum Pair) When a time series data generation step for generating time series data of LSP coefficient groups including coefficients and individual LSP coefficient groups constituting the LSP coefficient group time series data does not satisfy a predetermined stability condition In addition, the correction method is switched depending on whether the LSP coefficient group is located at the head, tail, or other position in the 1 state of the phoneme HMM that has output the LSP coefficient group so that the predetermined stability condition is satisfied. A step of correcting the LSP coefficient group, wherein the predetermined stability condition is: A plurality of LSP coefficients constituting the serial individual LSP coefficient group are all smaller than the larger and π than 0, and, when arranged in ascending order of the dimensions of the LSP coefficients, I der be arranged in ascending order, the correction In this step, when one LSP coefficient group that does not satisfy the predetermined stability condition is located at the head in the first state, the one LSP coefficient group located at least in the second state is used as the first LSP coefficient group. When the LSP coefficient group is corrected and is located at the end in the state 1, the LSP coefficient group is corrected at least using the LSP coefficient group located at the second position from the end in the state. In the case of being located in the other state, at least one LSP coefficient group adjacent in that state is used to correct the one LSP coefficient group .

According to the present invention, even if the sudden LSP coefficients in the course of the speech synthesis based on HMM is obtained, A better correction is achieved, as a result, synthesis of a higher quality of the text speech can be realized.

  Hereinafter, a text-to-speech synthesizer according to an embodiment of the present invention will be described in detail.

(Embodiment 1)
First, the configuration of the text-to-speech synthesizer according to this embodiment will be described. FIG. 1 is a functional configuration diagram of a text-to-speech synthesizer 81 according to the present embodiment.

  The text-to-speech synthesizer 81 includes a text capturing unit 13, a phoneme label string converter 15, a time-series data generator 83, a spectrum parameter corrector 19, an excitation sound source generator 85, a synthesis filter, as shown in the figure. 87 and a speaker unit 23.

  The text speech synthesizer 81 is connected to the speech synthesis dictionary 25. The speech synthesis dictionary 25 stores a phoneme HMM related to LSP coefficients and a phoneme HMM related to pitch for each phoneme label. The parameters that define the phoneme HMM related to the LSP coefficient include the average value, the variance value, the transition probability for each state, and the output probability of the LSP coefficient in each state. The parameters that define the phoneme HMM related to the pitch include the average value and the variance Value, transition probability for each state, and output probability of pitch in each state. These data are stored in a hard disk or the like.

  The text capturing unit 13 of the text speech synthesizer 81 captures a given text character string.

  The phoneme label string conversion unit 15 converts each text character constituting the captured text character string into a phoneme label string by sequentially converting the text characters to the phoneme label in the order of capture, and passes it to the spectrum and pitch time series data generation unit 83.

  The specific form of the phoneme label is arbitrary as long as it can be adapted to the specification of the speech synthesis dictionary 25. For example, in the case of a triphone that regards a triplet of phonemes as one speech unit, information on phonemes before and after the search target phoneme is also required when searching in the speech synthesis dictionary 25. Therefore, information of phonemes before and after the phoneme label is taken into consideration.

  Using the received phoneme label string as a clue, the time-series data generation unit 83 searches the phoneme HMM for the corresponding LSP coefficient and pitch from the speech synthesis dictionary 25 and connects them in the order in which the phoneme labels appear. These connections are phoneme HMM sequences related to LSP coefficients and pitch, respectively, and LSP coefficient group time series data and pitch time series data are respectively set so as to maximize the likelihood from these phoneme HMM sequences related to LSP coefficients and pitch. Generate.

  That is, the time series data generation unit 83 generates LSP coefficient group time series data and pitch time series data from the received phoneme label string.

  The time series data generation unit 83 delivers the generated LSP coefficient group time series data to the spectrum parameter correction unit 19 and delivers the generated pitch time series data to the excitation sound source generation unit 85.

  The spectrum parameter correction unit 19 performs a predetermined correction process on the received LSP coefficient group time series data.

  Predetermined correction processing refers to checking whether or not a predetermined stability condition is satisfied for each frame of LSP coefficient group time-series data and, if not, the LSP coefficient group Is corrected with reference to another LSP coefficient group generated from the state of the phoneme HMM that has output. Details of this correction processing will be described later with reference to FIGS.

  The spectrum parameter correction unit 19 passes the corrected LSP coefficient group time series data to the synthesis filter 87.

  On the other hand, the excitation sound source generation unit 85 receives the pitch time series data from the time series data generation unit 83, generates excitation sound source data, and passes it to the synthesis filter 87.

  The synthesized filter 87 receives the corrected LSP coefficient group time-series data and excitation sound source data. The synthesis filter 87 generates the synthesized voice data by inputting the excitation sound source data to the synthesis filter 87 defined by the corrected LSP coefficient, and transmits the synthesized voice data to the speaker unit 23.

  The speaker unit 23 outputs the delivered synthesized voice data. The speaker unit 23 has functions necessary for outputting synthesized speech, such as a sound source and an amplifier.

  The text-to-speech synthesizer 81 shown in FIG. 1 physically has a user I / F 39, a CPU 33, a ROM 35, a storage unit 37, a data input I / F 41, and a bus for interconnecting them as shown in FIG. 47. The computer 31 is provided with 47.

  Further, the computer 31 includes a speaker 45 and a speaker drive sound board 43 connected to the speaker, and the sound board 43 is connected to the bus 47 in order to emit synthesized speech.

  The ROM 35 stores an operation program for text-to-speech synthesis, and in particular, in this embodiment, an operation program including an operation for correcting the LSP coefficient of speech data.

  The storage unit 37 includes a RAM 51 and a hard disk 53, and includes a text character string, phoneme label string, phoneme HMM string, phoneme spectrum string, pitch information, LSP coefficient group time series data, pitch time series data, and corrected LSP coefficients. Group time series data, excitation sound source data, and voice data are stored.

  The data input I / F 41 is connected to the speech synthesis dictionary 25 and inputs data.

  The user I / F 39 includes a keyboard 55 and a monitor 57, and the user gives a text character string to the text-to-speech synthesizer.

(Operation)
Next, the operation of the speech synthesizer shown in FIGS. 1 and 2 will be described with reference to FIGS.

  When a text character string is given from the user I / F 39 and a voice synthesis start instruction is given, the CPU 33 executes a voice synthesis operation by executing an operation program stored in the ROM 35.

  First, the CPU 33 stores the given text character string in the storage unit 37.

  Next, the CPU 33 sequentially reads the text character strings stored in the storage unit 37, converts them into phoneme label strings, and temporarily stores them in the storage unit 37. Note that the method of converting a text character string into a phoneme label sequence is arbitrary, and any known method can be used.

  The CPU 33 sequentially extracts phoneme HMMs one by one from the speech synthesis dictionary 25 based on the phoneme label sequence stored in the storage unit 37, and generates a phoneme HMM sequence (see FIG. 3B).

  The CPU 33 also generates an LSP coefficient group from the phoneme HMM sequence (see FIG. 3C).

  There are usually a plurality of LSP coefficient groups in one state period. A time span corresponding to each LSP coefficient group is called one frame.

  Here, as shown in FIG. 3, it is assumed that each phoneme HMM has three states as states for outputting the LSP coefficient group, and a plurality of frames fm are outputted from each state. Each frame has 10 LSP coefficients. The horizontal axis corresponds to the time axis.

Of the frames corresponding to the state S _j of the phoneme HMMλ _i , the first one is called fm λ _i , S _{j , S,} and the last one is called fm λ _i , S _{j , E.} The storage unit 37 stores the frame identification information together with the LSP coefficient group data output from each state.

That is, as shown in FIG. 3, assuming that the number of frames is N _fm and the dimension of the LSP coefficient is N _d dimension, the LSP coefficient group time series data uses the LSP coefficient ω _d [fm],
{Ω _d [1], ω _d [2], ..., ω _d [fm], ..., ω _d [N _fm -1], ω _d [N _fm ]}
(However, 1 ≦ d ≦ N _{d = 10.} )
It is expressed.

  Next, the CPU 33 performs a predetermined correction process on the LSP coefficient group time series data stored in the storage unit 37 and stores it in the storage unit 37 as corrected LSP coefficient group time series data. The predetermined correction process will be described later with reference to FIG.

  In the above description, it is assumed that the user inputs a text string to be read out to the text-to-speech synthesizer, then issues a speech synthesis start instruction, and waits for the output of synthesized speech. Yes. This is called a batch method.

  On the other hand, at the time of text-to-speech synthesis, there may be a request to read a text character string sequentially by synthesized speech while reading the text character string for each predetermined division unit. This is called a real-time method.

  In the real-time method, the text character string to be read out by the speech synthesizer is interpreted as a series of shorter text character strings, and the short text character strings are processed sequentially by the batch method.

  That is, the operation of the text-to-speech synthesizer is essentially the same whether the batch method or the real-time method is realized. Therefore, this embodiment can realize both the batch method and the real-time method.

  However, in realizing the real-time method, it is necessary to pay attention to what division unit the text character string is read into the text-to-speech synthesizer. Depending on the structure of the speech synthesis dictionary 25 of FIG. 1, it may be necessary to input a certain amount of text character string to the speech synthesizer as one division unit. For example, when the speech synthesis dictionary 25 requires information on the context or the like for the search, it is necessary to use a text character string having a length sufficient to find the context or the like as a classification unit.

  As shown in FIG. 1, the feature of the text-to-speech synthesizer 81 according to the present embodiment is that the spectrum parameter correction unit 19 performs a predetermined correction process on the LSP coefficient group time series data.

Next, the spectrum correction process executed by the CPU 33 will be described with reference to FIGS.
First, FIG. 3 shows the correspondence relationship between the phoneme HMM, each state of the phoneme HMM, the LSP coefficient, and the frame. In order to facilitate understanding, it is assumed that a phoneme HMM having a state number 5 (S ₀ is an initial state, S ₄ is an end state, and no LSP coefficient is output) is applied, and the LSP coefficient is 10-dimensional.

  Each phoneme HMM has three states S1, S2, and S3 as states for outputting LSP coefficient groups, and LSP coefficient groups for a plurality of frames (fm) are output from each state. Each frame has 10 LSP coefficients. The horizontal axis corresponds to the time axis.

Of the frames corresponding to the state S _j of the phoneme HMMλ _i , the first one is called fm λ _i , S _{j , S,} and the last one is called fm λ _i , S _{j , E.}

Assuming that the number of frames is N _fm and the dimension of the LSP coefficient is the N _d dimension, the spectral time series data uses the LSP coefficient ω _d [fm],
{Ω _d [1], ω _d [2], ..., ω _d [fm], ..., ω _d [N _fm -1], ω _d [N _fm ]}
(However, 1 ≦ d ≦ N _d .)
It is represented by In the case of FIG. 3, N _d = 10.

  For some reason, there is a case where the stability condition regarding the LSP coefficient is not satisfied in any frame. If this is left unattended, the sound quality of the synthesized speech will be adversely affected. Therefore, the LSP coefficient is corrected.

The correction process will be described with reference to the flowchart shown in FIG.
As a premise, it is assumed that LSP coefficient group time-series data is generated by the above-described processing, and all LSP coefficients ω _d [fm] are obtained and stored in the storage unit 37 of FIG.

  First, the CPU 33 initializes a pointer fm that designates a number for identifying a frame to 1 (step S11).

The CPU reads LSP coefficients ω ₁ [fm], ω ₂ [fm],..., Ω _Nd [fm] from the storage unit 37 for the frame fm (step S13).

Check whether the stability condition of the LSP coefficient is satisfied. That is, it is determined whether or not 0 <ω ₁ [fm] <ω ₂ [fm] <... <Ω _Nd−1 [fm] <ω _Nd [fm] <π is satisfied (step S15).

  If it is determined that the stability condition is satisfied (step S15; satisfied), the LSP coefficient confirmation flag cf [fm] is set to a value 1 indicating that the stability condition is satisfied (cf [fm] = 1). 2 is stored in association with the frame number fm (step S17).

  If it is determined in step S15 that the stability condition is not satisfied (step S15; not satisfied), cf [fm] = 0 is stored in the storage unit 37 (step S19).

  It is determined whether or not processing has been completed for all fm (step S21).

If not completed, fm is incremented by 1, and the process returns to step S13. On the other hand, if completed, the generated LSP coefficient confirmation flag sequence data {cf [1], cf [2],..., Cf [fm],..., Cf [N _fm −1], cf [ N _fm ]}, it is determined whether or not fm _NG with cf [fm _NG ] = 0 exists (step S23).

If there is no such fm _NG , the correction process is terminated.

In step S23, when it is determined that fm _NG with cf [fm _NG ] = 0 exists, it is determined which of the states S _{1 to} S ₃ corresponds to the position of the frame fm _NG. (Step S25).

That is, in FIG. 3, if fm _{λi, Sj, S} ≦ fm _NG ≦ fm _{λi, Sj, E} , the frame fm _NG corresponds to the state S _j .

Result of the position determination in step S25, the frame _{fm NG} is to have been determined to correspond to a state _{S 1.} In such a case, subsequently, it is determined based on the frame identification information whether the frame fm _NG is in a finer position among the frames corresponding to the state S ₁ (step S27).

The discrimination result in step S27 is divided into three types. That is, when the frame fm _NG is the _first frame among the frames corresponding to S ₁ (fm _NG = fm _{λi, S1, S} ), it is neither the head nor the tail (fm _{λi, S1, S} <fm _NG < fm _{λi, S1, E} ) and the last case (fm _NG = fm _{λi, S1, E} ).

If the beginning of, the correction is performed with reference to the LSP coefficients of the second and subsequent frame corresponding to S ₁ (step S29). If neither end at the top, corrected by referring to the LSP coefficients of the in position, such as adjacent or neighboring adjacent to the corresponding vital fm _NG in S ₁ (near) frame (step S31). For the last, it is corrected by referring to the LSP coefficients of the second previous frame from the end corresponding to S ₁ (step S33).

In step S25, the case where it is determined to correspond to the case and _{S 3} that is determined to correspond to _{S 2} also performs similar processing to that of Step S29～S33. That is, correction is performed using LSP coefficient groups that are output from the same state and are adjacent.

As described above, when the LSP coefficient corresponding to the phoneme HMM corresponding to the phoneme HMMλ _i includes an abrupt LSP coefficient as a result, the stability condition is not satisfied, the CPU 33 corrects the adjacent phoneme HMMλ when correcting the LSP coefficient. _The correction is performed using only the LSP coefficients that satisfy the stability condition obtained from the phoneme HMMλ _i without referring to the LSP coefficients corresponding to the phoneme HMM before _{i−1 and} the phoneme HMM after the phoneme HMMλ _{i + 1} .

Further, if the LSP coefficient corresponding to the state S _j (1 ≦ j ≦ 3) needs to be corrected, the correction is performed with reference to only the other LSP coefficient corresponding to S _j and corresponding to the same phoneme HMMλ _i . However, the LSP coefficient output from the other state S _k (k ≠ j) is not referred to.

  Therefore, even when there is a large gap in the LSP when the state is switched, since excessive interpolation is not performed, it is possible to provide high-quality synthesized speech without any unreasonable speech data synthesis. In other words, by providing more appropriate correction to the LSP coefficient time series data, it is possible to provide higher quality synthesized speech.

  Next, preferred specific examples of the correction processing according to this embodiment, in particular, the case classification of steps S25, S27, S29, S31, S33, etc. in FIG. 4 and the processing corresponding to each case will be described below.

(Specific example 1 of processing according to case classification and each case)
Case classification is case classification as shown in the following (a) to (h). Depending on the position of the frame fm _NG , the LSP coefficient is corrected as follows.

_{(A) fm NG = fm λi} , S1, when _{S (S25; S 1, S27} ; top)
ω _d [fm _NG ] = ω _d [fm _{λi, S1, S} + 1]
(However, 1 ≦ d ≦ N _d )
(Step S29).
That is, the LSP group one frame behind is copied.
_{(B) fm λi, S1,} S <fm NG <fm λi, S1, when _{E (S25; S 1, S27} ; even at the beginning, nor the last),
ω _d [fm _NG ] = (ω _d [fm−1] + ω _d [fm + 1]) / 2
(However, 1 ≦ d ≦ N _d )
(Step S31).
That is, it is corrected to the average value of the LSP group of one frame before and after.
(C) When fm _NG = fm _{λi, S1, E} (S25; S1, S27; tail)
ω _d [fm _NG ] = ω _d [fm _{λi, S1, E} −1]
(However, 1 ≦ d ≦ N _d )
age,
(D) When fm _{λi, S2, S} ≦ fm _NG <fm _{λi, S2, E} ,
ω _d [fm _NG ] = ω _d [fm _{λi, S2, S} + 1]
(However, 1 ≦ d ≦ N _d )
age,
(E) When fm _NG = fm _{λi, S2, E} ,
ω _d [fm _NG ] = ω _d [fm _{λi, S2, E} −1]
(However, 1 ≦ d ≦ N _d )
age,
(F) When fm _NG = fm _{λi, S3, S} ,
ω _d [fm _NG ] = ω _d [fm _{λi, S3, S} + 1]
(However, 1 ≦ d ≦ N _d )
age,
(G) When fm _{λi, S3, S} <fm _NG <fm _{λi, S3, E} ,
ω _d [fm _NG ] = (ω _d [fm−1] + ω _d [fm + 1]) / 2
(However, 1 ≦ d ≦ N _d )
age,
(H) When fm _NG = fm _{λi, S3, E} ,
ω _d [fm _NG ] = ω _d [fm _{λi, S3, E} −1]
(However, 1 ≦ d ≦ N _d )
And

  This specific example is characterized in that an unexpected LSP coefficient is referenced with reference to the LSP coefficient of another frame generated from the state of the phoneme HMM that output the coefficient.

(Specific example 2 of processing according to case classification and each case)
In this specific example, unlike the specific example described above, in some cases, the LSP coefficient of the frame adjacent to the frame to which the correction target LSP coefficient belongs and the frame adjacent to the adjacent frame are also referred to as follows. There are features.

In the case of (a), ie, fm _NG = fm _{λi, S1, S} ,
ω _d [fm _NG ] = ω _d [fm _{λi, S1, S} + 1] × 2 + ω _d [fm _{λi, S1, S} + 2] × (−1)
(However, 1 ≦ d ≦ N _d .)
age,
In the case of (c), ie, fm _NG = fm _{λi, S1, E} ,
ω _d [fm _NG ] = ω _d [fm _{λi, S1, E} −2] × (−1) + ω _d [fm _{λi, S1, E} −1] × 2
(However, 1 ≦ d ≦ N _d .)
age,
In the case of (f), ie, fm _NG = fm _{λi, S3, S} ,
ω _d [fm _NG ] = ω _d [fm _{λi, S3, S} +1] × 2 + ω _d [fm _{λi, S3, S} +2] × (−1)
(However, 1 ≦ d ≦ N _d .)
age,
In the case of (h), ie, fm _NG = fm _{λi, S3, E} ,
ω _d [fm _NG ] = ω _d [fm _{λi, S3, E} −2] × (−1) + ω _d [fm _{λi, S3, E} −1] × 2
(However, 1 ≦ d ≦ N _d .)
And

(Example 3 of case classification and processing according to each case)
This specific example is the same as the two specific examples in that the LSP coefficient correction method is switched depending on the position of the frame fm _NG . In this specific example, only the LSP coefficient of the order that needs correction is corrected so as to satisfy the stability, and the corrected LSP coefficient data is output. Here, if the order d of the LSP coefficient that needs to be corrected is α ≦ d ≦ β (where α ≦ β, 1 ≦ α ≦ N _d , 1 ≦ β ≦ N _d ), The LSP coefficient is corrected as follows.

(A) When fm _NG = fm _{λi, S1, S} ,
ω _d [fm _NG ] = ω _d [fm _{λi, S1, S} + 1]
(However, α ≦ d ≦ β.)
age,
(B) When fm _{λi, S1, S} <fm _NG <fm _{λi, S1, E} ,
ω _d [fm _NG ] = (ω _d [fm−1] + ω _d [fm + 1]) / 2
(However, α ≦ d ≦ β.)
age,
(C) When fm _NG = fm _{λi, S1, E} ,
ω _d [fm _NG ] = ω _d [fm _{λi, S1, E} −1]
(However, α ≦ d ≦ β.)
age,
(D) When fm _{λi, S2, S} ≦ fm _NG <fm _{λi, S2, E} ,
ω _d [fm _NG ] = ω _d [fm _{λi, S2, S} + 1]
(However, α ≦ d ≦ β.)
age,
(E) When fm _NG = fm _{λi, S2, E} ,
ω _d [fm _NG ] = ω _d [fm _{λi, S2, E} −1]
(However, α ≦ d ≦ β.)
age,
(F) When fm _NG = fm _{λi, S3, S} ,
ω _d [fm _NG ] = ω _d [fm _{λi, S3, S} + 1]
(However, α ≦ d ≦ β.)
age,
(G) When fm _{λi, S3, S} <fm _NG <fm _{λi, S3, E} ,
ω _d [fm _NG ] = (ω _d [fm−1] + ω _d [fm + 1]) / 2
(However, α ≦ d ≦ β.)
age,
(H) When fm _NG = fm _{λi, S3, E} ,
ω _d [fm _NG ] = ω _d [fm _{λi, S3, E} −1]
(However, α ≦ d ≦ β.)
And

  In this specific example, only the LSP coefficient of the order that caused the loss of stability is corrected. Therefore, the LSP coefficient of the corresponding frame is satisfied as compared with the two specific examples while satisfying the stability condition of the LSP coefficient. It becomes possible to correct with an appropriate value.

  As in the case of specific example 2 as a variation of specific example 1, as a variation of this specific example, the frame adjacent to the frame to which the correction target LSP coefficient belongs and the LSP coefficient of the frame further adjacent to the adjacent frame are set. There is also a specific example of referring to.

(Embodiment 2)
In specific example 3 and the like as well, the sudden LSP coefficients are interpolated using only the LSP coefficients from the same phoneme HMM in the same state, but in these cases, since the influence is small, different phoneme HMMs, There is no problem even if interpolation is performed using LSP coefficients.

For example, if 2,3-D coefficient of the LSP coefficients of the first frame of the _{S 2} becomes sudden reversed, LSP of the second frame of 2,3-D and _{S 2} of the last LSP coefficients _{S 1} An average (or a weighted average) of two-dimensional and three-dimensional coefficients may be obtained, and interpolation may be performed using this average value. In this way, the influence of interpolation is limited to sudden dimensions, and using LSP coefficients in different states has no significant effect and may be good.

(Embodiment 3)
In the first embodiment, the sudden LSP coefficient is corrected using only the LSP coefficient generated from the same state of the same phoneme HMM. On the other hand, even if correction is performed using LSP coefficients output from different states of the same phoneme HMM, sufficiently high-quality text speech synthesis may be realized.

  According to the present embodiment, a more flexible correction method can be adopted depending on the situation.

It is also conceivable to omit the position determination (step S25, step S27) of fm _{NG in} the flowchart shown in FIG. Thereby, simpler LSP coefficient correction can be realized.

  In addition, this invention is not limited to the said embodiment, A various deformation | transformation and application are possible.

  For example, the hardware configuration, the block configuration, and the flowchart are examples and are not limited. The state and the position of the frame may be determined by analyzing the LSP without attaching the state information or the frame identification information.

  A combination of processing corresponding to case classification is also arbitrary within the scope of the gist of the invention.

  Furthermore, the present invention can be realized by using a general computer without using a dedicated device and system. That is, a computer program for executing the above-described processing may be manufactured and distributed on a computer, or the above-described processing may be executed by installing the computer program on the computer.

It is a function block diagram of the text speech synthesizer provided with the spectrum correction part based on Embodiment 1. FIG. 1 is a diagram illustrating a physical configuration of a text-to-speech synthesizer according to Embodiment 1. FIG. It is explanatory drawing which shows the correspondence of a phoneme HMM row | line | column, the number of states applied to each phoneme HMM, an LSP coefficient, and a frame. It is a figure which shows the flow of operation | movement in Embodiment 1 of a spectrum correction process.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 13 ... Text capture part, 15 ... Phoneme label sequence conversion part, 19 ... Spectral parameter correction part, 23 ... Speaker part, 25 ... Speech synthesis dictionary, 31 ... For text speech synthesis 33 ... CPU, 35 ... ROM, 37 ... storage unit, 39 ... user I / F, 41 ... data input I / F, 43 ... sound board, 45 ... Speaker, 47 ... Bus, 51 ... RAM, 53 ... Hard disk, 55 ... Keyboard, 57 ... Monitor, 81 ... Text-to-speech synthesizer according to Embodiment 1, 83 ... Time series data generation unit, 85 ... Excitation sound source generation unit, 87 ... Synthesis filter

Claims (6)

Phoneme label string conversion means for converting a given text string into a phoneme label string;
Time-series data generating means for generating time-series data of LSP coefficient groups including multi-dimensional LSP (Line Spectrum Pair) coefficients from the phoneme label string output from the phoneme label string converting means;
When it is determined whether or not a predetermined stability condition is satisfied for each LSP coefficient group constituting the LSP coefficient group time-series data generated by the time-series data generation means, and it is determined that it is not satisfied In order for this LSP coefficient group to satisfy the predetermined stability condition, correction is made depending on whether the phoneme HMM (Hidden Markov Model) that has output the LSP coefficient group is positioned at the head, tail, or other position. Correction means for switching the method to correct the LSP coefficient group;
A speech processing apparatus comprising:
The predetermined stability condition means that when a plurality of LSP coefficients constituting the individual LSP coefficient group are all larger than 0 and smaller than π and arranged in ascending order of the dimension of the LSP coefficient, they are arranged in ascending order. I Kotodea,
When the one LSP coefficient group that does not satisfy the predetermined stability condition is located at the head in the first state, the correcting means uses the LSP coefficient group located at least second in the state to use the 1 LSP coefficient group. When the first LSP coefficient group is corrected and is located at the tail in the state 1, the one LSP coefficient group is corrected at least using the LSP coefficient group located second from the tail in the state. The speech processing apparatus according to claim 1, wherein at least one LSP coefficient group is corrected using at least one adjacent LSP coefficient group in that state . When it is determined that the predetermined stability condition is not satisfied, the correcting unit determines that another LSP generated by the phoneme HMM that outputs the LSP coefficient group so that the LSP coefficient group satisfies the predetermined stability condition. Correcting the LSP coefficient group using the coefficient group;
The speech processing apparatus according to claim 1.   The speech processing apparatus according to claim 1, wherein the correction unit corrects only the LSP coefficient of a dimension that causes the stability condition not to be satisfied in the LSP coefficient group. The speech processing apparatus according to any one of claims 1 to 3,
Means for supplying text synthesis target text data to the phoneme label string conversion means of the speech processing apparatus;
Synthetic speech generation means for generating synthesized speech by inputting excitation sound source data to an LSP synthesis filter using LSP coefficient group time-series data generated by the speech processing apparatus as LSP synthesis filter coefficients;
A text-to-speech synthesizer. A phoneme label string conversion step for converting a text string into a phoneme label string;
A time series data generation step for generating time series data of LSP coefficient groups including LSP (Line Spectrum Pair) coefficients from the phoneme label string;
For each LSP coefficient group constituting the LSP coefficient group time series data, when the predetermined stability condition is not satisfied, the LSP coefficient group is output so that the LSP coefficient group satisfies the predetermined stability condition. A step of correcting the LSP coefficient group by switching a correction method depending on whether the phoneme HMM is located in the first, last, or other position in the 1 state of the phoneme HMM,
A voice processing method comprising:
The predetermined stability condition means that when a plurality of LSP coefficients constituting the individual LSP coefficient group are all larger than 0 and smaller than π and arranged in ascending order of the dimension of the LSP coefficient, they are arranged in ascending order. I Kotodea,
In the correcting step, when one LSP coefficient group that does not satisfy the predetermined stability condition is located at the head in the state 1, the LSP coefficient group located at the second position in the state is used at least. When one LSP coefficient group is corrected and is located at the end in the state 1, the LSP coefficient group is corrected at least using the LSP coefficient group located second from the end in the state, A speech processing method comprising: correcting one LSP coefficient group using at least one LSP coefficient group adjacent in that state when the other one is located in one state . On the computer,
A phoneme label string conversion step for converting a text string into a phoneme label string;
A time series data generation step for generating time series data of LSP coefficient groups including LSP (Line Spectrum Pair) coefficients from the phoneme label string;
For each LSP coefficient group constituting the LSP coefficient group time series data, when the predetermined stability condition is not satisfied, the LSP coefficient group is output so that the LSP coefficient group satisfies the predetermined stability condition. A step of correcting the LSP coefficient group by switching a correction method depending on whether the phoneme HMM is located in the first, last, or other position in the 1 state of the phoneme HMM,
A computer program for executing
The predetermined stability condition means that when a plurality of LSP coefficients constituting the individual LSP coefficient group are all larger than 0 and smaller than π and arranged in ascending order of the dimension of the LSP coefficient, they are arranged in ascending order. I Kotodea,
In the correcting step, when one LSP coefficient group that does not satisfy the predetermined stability condition is located at the head in the state 1, the LSP coefficient group located at the second position in the state is used at least. When one LSP coefficient group is corrected and is located at the end in the state 1, the LSP coefficient group is corrected at least using the LSP coefficient group located second from the end in the state, A computer program for correcting one LSP coefficient group using at least one LSP coefficient group adjacent in that state when positioned in the other state in one state .

Images