KR101567566B1 - System and Method for Statistical Speech Synthesis with Personalized Synthetic Voice - Google Patents

Abstract

The present invention relates to a statistical voice synthesis system applying personalized voice and a method thereof wherein synthesized sound to which voice characteristics of a patient with speech disorder are applied is generated in the statistical voice synthesis system to improve communication of the patient. The statistical voice synthesis system includes: a short vowel voice collection unit which collects short vowel voice from the patient with speech disorder; a speaker adaptation parameter extraction unit which compares short vowel voice, collected in the short vowel voice collection unit, with a relevant short vowel model from sound models of the statistical voice synthesis system and extracts a bilinear conversion speaker adaptation parameter based on a formant; a synthesized sound generation unit which generates synthesized sound to which the characteristics of a speaker are applied by applying the speaker adaptation parameter extracted in the speaker adaptation parameter extraction unit; and a synthesized sound tuning unit which selectively tunes timbre of the synthesized sound generated in the synthesized sound generation unit.

Description

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a statistical speech synthesis system and method,

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a statistical speech synthesis technology, and more particularly, to a statistical speech synthesis system and a method for generating a synthetic speech reflecting a tone characteristic of a horse disordered person in a statistical speech synthesis system, will be.

Currently, commercialized speech synthesis systems mainly use a corpus-based synthesis method. In this method, a lot of work is required in addition to a large amount of speech data in order to express the tone of a specific speaker.

Nonetheless, if the speech impairment progresses slowly, you can record a large amount of voice from the patient before it becomes severe and produce a synthetic voice with your own voice from it.

However, this method can not be applied to patients who have already developed a lot of speech disorders.

The statistical speech synthesis method developed after the corpus-based synthesis method is a method of synthesizing speech using a statistical acoustic model that has been used for speech recognition in the past. Even if only a relatively small amount of specific speaker speech data is used, A speaker adaptation function for generating a speaker adaptation function.

1 shows a configuration diagram of a statistical speech synthesis system having a speaker adaptation function according to the prior art.

A statistical speech synthesis system having a speaker adaptation function of the prior art includes collecting large-scale speech data of a specific speaker or a plurality of speakers (S100), performing acoustic modeling for statistical speech synthesis (S101) (S102), a speaker adaptation step (S104) using the established specific speaker or the average speech acoustic model and the small-scale speech data S103 of the target speaker, and the acoustic model S105) to generate a synthesized sound (S106).

In this statistical speech synthesis method, the scale of voice data necessary for speaker adaptation is known to be about 6 to 7 minutes (about 100 sentences), and it is very difficult for the patients with speech disorders to utter this amount, The ambiguity and the noise characteristic of the obstacle voice are reflected together with the tone color characteristic desired by the user, so that the clarity of the speaker-adapted synthetic sound also deteriorates seriously.

In order to solve this problem, in the prior art, there are a method of replacing some parameters of an acoustic model adapted to a speaker with a voice of a horse disorder patient by parameters of an existing average voice model, And a speaker adaptation model is constructed from the result.

In both of these methods, the clarity of the synthesized voice is improved to some extent, but the clarity of the synthesized voice is still much lower than that of the normal person.

Recently, a method of applying speaker adaptation to a specific speaker voice having a short length of about one sentence by applying the vocal tract length normalization (VTLN) technique to statistical speech synthesis has been proposed. When applied to adaptation, the patient's recording burden is greatly reduced.

However, even in this method, it is difficult to automate the process of matching the voice of one sentence with the acoustic model of the existing phoneme due to the characteristics of pronunciation imprecision of the patient with the horse disorder, Since the VTLN is performed using the MGC (Mel-Generalized Cepstrum) parameter directly, there is a problem that is not effective in identifying the personality of a patient with a horse disorder.

Korean Patent Publication No. 10-2011-0021944 Korean Patent Publication No. 10-2013-0109902

In order to solve the problem of the conventional statistical speech synthesis system, the present invention provides a statistical speech synthesis system, in which a synthetic tone reflecting the tone characteristics of a horse disordered person is generated, And to provide a synthesis system and method.

The present invention is directed to patients with speech disorders who do not have a severely impaired voice so that the voice of a person's voice can not be clearly identified by the general public, And a method for generating a synthetic sound reflecting a personal timbre characteristic in a statistical speech synthesis system.

In the present invention, patients with speech disorders have difficulty in communicating normally with the general public. Therefore, Augmented and Alternative Communication (AAC) devices having voice synthesis functions are sometimes used as auxiliary means. A personal voice which reflects personal characteristics is learned by learning the tone color characteristics of the user from the voice data of the patient himself or herself, And to provide a synthesis system and method.

The objects of the present invention are not limited to the above-mentioned objects, and other objects not mentioned can be clearly understood by those skilled in the art from the following description.

According to another aspect of the present invention, there is provided a statistical speech synthesis system that reflects individual tone colors, comprising: a short vowel sound collection unit for collecting short vowel sound from a patient with a horse disorder; a short vowel sound collection unit for collecting short vowel sounds and a statistical speech synthesis system A speaker adaptation parameter extracting unit for extracting a formant-based bilinear transformed speaker adaptation parameter by comparing the corresponding short sound model among the acoustic models, a synthesized sound reflecting the speaker characteristics by applying the speaker adaptation parameter extracted from the speaker adaptation parameter extracting unit And a synthesized sound tuning unit for selectively tuning a tone color of the synthesized sound generated by the synthetic sound generating unit.

Here, the speaker adaptation parameter extracting unit, collected vowel 3 for audio data delimited units of frames separated by frame-by-frame unit, 2 K + 1 frames consisting of a frame energy as the largest frame, and before and after the K frame A formant frequency extracting unit for extracting a plurality of formant frequencies, that is, a first formant F1, a second formant F2, and a third formant F3, ( F 1, F 2, and F 3) extracted from a specific short vowel sound of a patient with a horse disorder is determined as a median of formant frequencies of each extracted frame, and an acoustic model corresponding to the short vowel A speaker adaptation parameter determining a bilinear transform coefficient? _SA whose weight sum of squared errors is minimum when expressed as a bilinear transform of the formant frequencies ( F 1 _M , F 2 _M and F 3 _M ) Parami Crystal characterized in that it comprises a.

In order to accomplish the other object, there is provided a method of statistically synthesizing a voice according to the present invention, comprising: collecting a short vowel voice from a patient with speech disorders; comparing the collected short vowel voice with a corresponding one of the acoustic models of the statistical speech synthesis system; Extracting a speaker-based bilinear transformed speaker adaptation parameter, and generating a synthesized speech that reflects the speaker characteristic by applying the extracted speaker adaptation parameter.

Herein, in the step of generating the synthetic sound reflecting the speaker characteristics by applying the extracted speaker adaptation parameter, the step of tuning the tone color based on the result of the generated synthetic sound is further performed.

And Four steps of treatment based on a bilinear transformation to extract the speaker adaptation parameters of the is provided with a step to separate the vowel sound collecting data on a frame-by-frame basis, the frame energy consisting of the largest frame and before and after the K frame 2 K + 1 Extracting a first formant frequency F1, a second formant F2, and a third formant frequency F3 for three frames, Four determining a median value (median) of the treatment frequency, formant frequencies extracted from a particular vowel of the end disorder (F 1, F 2 and F 3) of the formant frequency (F 1 of the acoustic model corresponding to the vowel _M, F 2 characterized by including the _M and obtaining F 3 _M bilinear transformation (bilinear transform) the weighted square error sum is bilinear transform coefficients is minimized α _SA when rendering of) a speaker adaptation parameters.

And, in the step of obtaining the bilinear transform coefficient? _SA with the weighted square error sum as a speaker adaptation parameter, the bilinear transform equation in frequency (Hz unit)

이고,

ego,

와

는 각각 쌍선형 변환 이전 및 이후의 주파수를 의미하며, α는 쌍선형 변환 파라미터인 것을 특징으로 한다.here

Wow

Respectively denote frequencies before and after the bilinear transformation, and? Is a bilinear transformation parameter.

에 대해 정리한 함수

는,

이고,And the bilinear transformation in frequency (Hz)

Functions that are summarized for

Quot;

ego,

를 이용하여 말장애 환자의 특정 단모음으로부터 추출한 포먼트 주파수(F1,F2 및 F3)들과 이 단모음에 해당하는 음향모델의 포먼트 주파수(F1_M,F2_M 및 F3_M)의 쌍선형 변환 값들과의 가중제곱오차합은,Binary Linear Conversion Functions

( F 1, F 2, and F 3) extracted from a specific short vowel of a horse disordered person and formant frequencies ( F 1 _M , F 2 _M, and F 3 _M ) of the acoustic model corresponding to the short vowel, The weighted squared error sum with the binarized transformed values of < RTI ID = 0.0 >

이고,

ego,

는, 말장애 환자의 음성의 경우 일반인의 음성에 비해 포먼트 주파수 추출의 신뢰도가 떨어짐을 감안하여, 추정된 i번째 포먼트 주파수의 신뢰도를 고려한 가중치인 것을 특징으로 한다.Here,

Is a weight considering the reliability of the estimated i- th formant frequency considering the reliability of the formant frequency extraction is lower than the voice of a general person in the case of speech of a horse disordered patient.

계산은,Weights that take into account the reliability of individual formant frequencies

The calculation,

이고, 여기서

와

는 사용된 2K + 1개 프레임의 i번째 포먼트 주파수

들의 평균과 표준 편차값, 함수 g(x)는 x ≥ 0인 범위에 대해 단조증가 특성을 가지는 함수인 것을 특징으로 한다.

, Where

Wow

Is the i- th formant frequency of the 2 K + 1 frames used

And the function g ( x ) is a function having a monotone increasing property with respect to a range of x ? 0.

라고 하면, 음색 변환은,And in the step of tuning the sound based on the listening results of the generated synthesized voice, of the prosody characteristic of the synthesized sound, intonation characteristics indicating high and low of sound is used to represent a log F0 parameters of the acoustic model, the old for the j-th frame of the synthesized sound The log F0 value is set to LF 0 ( j ), the tone-converted log F0 value is set to

In the tone color conversion,

으로 이루어지고,

Lt; / RTI >

Here, LF 0 _SA is a user-specified parameter for converting the intonation characteristics of the synthetic sound. If LF 0 _SA > 0, the sound becomes higher and LF 0 _SA <-Negative or 0 becomes low, while adjusting the value LF 0 _SA hear the synthesized voice is characterized in that selection of the LF 0 _SA value.

Then, in the step of tuning the tone based on the result of the generated synthesized tone, a pairwise linear transformation using the obtained speaker adaptation parameter and a bilinear transformation for the mel-scale frequency transformation are expressed by a single pair linear transformation,

이고,Bilinear transformation parameter α is _F,

ego,

Where α _SA is a bilinear transform coefficient that minimizes the weighted square error sum, α _M is a bilinear transform coefficient for the Mel-scale frequency transform,

In the case of spectral conversion using the bilinear transformation parameter α _F , if the person with terminal impairment wishes to change the spectral characteristics additionally, he or she can listen to the synthesized sound while adjusting the α _F value and obtain the most satisfactory α _F value .

The system and method for statistical speech synthesis reflecting the personalized tone according to the present invention have the following effects.

First, synthesized speech with the main tone characteristics of the person can be generated by only the vowel voice utterance of the horse disordered patient.

Second, in the statistical speech synthesis system, it is possible to improve the communication by generating a synthetic sound reflecting the tone characteristic of the horse disordered person.

Third, it is possible to generate the synthesized voice reflecting the personality by learning the tone characteristic of the user from the voice data of the patient with the hurdle so that the tone of the synthesized voice provided by the AAC device is not limited.

1 is a flowchart showing a statistical speech synthesis method having a speaker adaptation function according to the prior art
2 is a block diagram of a statistical speech synthesis system reflecting a personal tone color according to the present invention
3 is a detailed configuration diagram of a speaker adaptation parameter extracting unit according to the present invention;
FIG. 4 is a flowchart illustrating a statistical speech synthesis method reflecting a personal tone color according to the present invention.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, preferred embodiments of a statistical speech synthesis system and method according to the present invention will be described in detail.

The features and advantages of the statistical speech synthesis system and method reflecting the personalized tone according to the present invention will be apparent from the following detailed description of each embodiment.

FIG. 2 is a configuration diagram of a statistical speech synthesis system reflecting personal tones according to the present invention, and FIG. 3 is a detailed configuration diagram of a speaker adaptation parameter extraction unit according to the present invention.

The present invention learns personal tone characteristics of a person with only one voice utterance which can be most clearly uttered by him / herself, thereby improving communication by generating a synthetic voice reflecting personal tone characteristics in a statistical speech synthesis system.

As shown in FIG. 2, the statistical speech synthesis system reflecting the personalized tone according to the present invention includes a short vowel sound collection unit 20 for collecting short vowel sound from a horse disordered person, A speaker adaptation parameter extracting unit 30 for extracting a formant-based bilinear transformed speaker adaptation parameter by comparing corresponding short sound model among the models, and a synthesized sound generating unit 30 for generating a synthesized sound reflecting the speaker characteristics by applying the extracted speaker adaptation parameter And a synthesized sound tuning unit 50 for selectively tuning a tone color according to a user's taste based on the result of the generated synthetic sound.

Here, the speaker adaptation parameters, such as the detailed construction is in Figure 3 of the extraction unit 30, a collection of short vowels and the sound data delimited units of frames separated by frame unit 31, a frame energy largest frame, and the front and rear K The first formant F1, the second formant F2, and the third formant F3 frequency are extracted from the 2 K + 1 frames, A median determining unit 33 for determining the median of the formant frequencies of the respective frames extracted by the formant frequency extracting unit 32, ( F 1, F 2, and F 3) extracted from the formant frequencies are represented by a bilinear transform of the formant frequencies ( F 1 _M , F 2 _M, and F 3 _M ) of the acoustic model corresponding to the short vowels when the weighted square error sum is the smallest pair of linear transformation coefficient α _SA the speaker Response includes speaker adaptation parameter to obtain a parameter determination unit 34. The

The speech synthesis process in the statistical speech synthesis system reflecting the individual tone colors according to the present invention is as follows.

4 is a flowchart illustrating a statistical speech synthesis method reflecting a personal tone color according to the present invention.

The statistical speech synthesis method reflecting the individual tone color according to the present invention includes a step of collecting short vowel sound from a patient with terminal impairment (S401), comparing the collected short vowel speech with a corresponding short vowel sound model among the acoustic model S402 of the statistical speech synthesis system A step S403 of extracting a formant-based bilinear transformed speaker adaptation parameter, and a step S404 of generating a synthesized sound reflecting speaker characteristics by applying the extracted speaker adaptation parameter. (S405) of tuning a tone according to the taste of the user based on the result of listening to the synthesized sound.

Hereinafter, the present invention will be described in detail in each step.

First, the step S401 of collecting a short vowel voice from a patient with a horse disorder is performed by a patient with a horse disorder among short vowel sounds (/ / / / / / / / / / / / / / / / /, etc.) It emits shortest voices that can be uttered most clearly and collects it through a digital recording device including a personal computer.

Listen to the recorded voice and if it is unclear or unnatural, repeat the recording and listen to the most clear vocalizations.

The step S403 of extracting the formant-based bilinear transformed speaker adaptation parameter by comparing the short monophonic sound and the acoustic model of the statistical speech synthesis system and comparing the corresponding short-

First, the collected short vowel sound data is divided into frames in the same manner as that used in the normal statistical speech synthesis method. For 2 K + 1 frames having the largest frame energy and K frames before and after the frame, The first formant F1, the second formant F2, and the third formant F3 frequency are extracted from the three formants.

At this time, the value of K is experimentally determined, and the unit of the formant frequency is Hz.

These formant frequencies are obtained by a conventional linear prediction (LP) analysis method. 2 K + 1 of the formant frequency of the k-th frame of the frames F 1 (k), F 2 (k) and F 3 when said (k), the formant frequency of the collected short vowels, F 1, F 2 And F3 are determined as a median of the formant frequencies of each frame as shown in the following Equation (1).

Here, Median {} is the median of the values in {}.

The acoustic model of the statistical speech synthesis system is obtained from the speech of a specific speaker or the speech of many speakers based on a normal statistical speech synthesis method.

The LSP parameter, that is, the MGC-LSP parameter, which is converted from the mean MGC parameter of the most central position of the triphone model corresponding to the short vowels collected in this acoustic model, i.e., the mute- Second, and third formant frequencies are extracted in the same manner, and these are called formants frequencies F 1 _M , F 2 _M, and F 3 _M of the acoustic model.

In the speaker adaptation process according to the present invention, the formants frequency ( F 1, F 2 and F 3) extracted from a specific short vowel sound of a patient with a horse disorder is compared with the formant frequencies ( F 1 _M , F 2 _M , F 3 _M ) is expressed as a bilinear transform, a bilinear transform coefficient? _SA is obtained as a speaker adaptation parameter that minimizes the weighted square error sum.

This process will be described in more detail as follows.

First, the bilinear transformation in the frequency (Hz) region is as follows.

와

는 각각 쌍선형 변환 이전 및 이후의 주파수를 의미하며, α는 쌍선형 변환 파라미터이다. 수학식 2를

에 대해 정리한 함수

는 다음 수학식 3과 같다. here

Wow

Respectively denote the frequencies before and after the bilinear transformation, and a is the bilinear transformation parameter. Equation (2)

Functions that are summarized for

Is expressed by the following equation (3).

를 이용하여 말장애 환자의 특정 단모음으로부터 추출한 포먼트 주파수(F1,F2 및 F3)들과 이 단모음에 해당하는 음향모델의 포먼트 주파수(F1_M,F2_M 및 F3_M)의 쌍선형 변환 값들과의 가중제곱오차합은 다음 수학식 4에서와 같다.The bilinear transform function of equation (3)

( F 1, F 2, and F 3) extracted from a specific short vowel of a horse disordered person and formant frequencies ( F 1 _M , F 2 _M, and F 3 _M ) of the acoustic model corresponding to the short vowel, And the weighted squared error sum with the binarized conversion values of Equation (4) is as shown in Equation (4).

는, 말장애 환자의 음성의 경우 일반인의 음성에 비해 포먼트 주파수 추출의 신뢰도가 떨어짐을 감안하여, 추정된 i번째 포먼트 주파수의 신뢰도를 고려한 가중치이다.Here,

Is a weight considering the reliability of the estimated i- th formant frequency considering that the reliability of the formant frequency extraction is lower than the voice of a general person in the case of speech of a horse disordered person.

계산 방법으로는 다음과 같은 수학식 5를 사용할 수 있다.Weights that take into account the reliability of individual formant frequencies

As a calculation method, the following Equation (5) can be used.

와

는 수학식 1에 사용된 2K + 1개 프레임의 i번째 포먼트 주파수

들의 평균과 표준편차 값을 의미한다.here

Wow

Is the i- th formant frequency of the 2 &lt; K &gt; + 1 frame used in equation

Means the mean and standard deviation of the mean.

And the function g ( x ) is a function with monotone increasing property for the range of x ≥ 0.

From Equation (5), as the standard deviation is relatively smaller than the mean value of the i- th formant frequency, it is determined that the reliability of the formant estimation is high and the weight increases. The specific g ( x ) function setting is determined experimentally.

를 n -1개의 구간으로 균등 분할한 n개의 α값 The value &lt; RTI ID = 0.0 &gt; a &lt; / RTI &gt; that minimizes the weighted squared error sum of Equation (4)

The equal dividing the α value of n as n -1 of period

을 각각 수학식 4에 대입하여 그 중에서 WSE(α)를 최소화하는 α값을 α_SA로 정한다.

Are substituted into Equation (4), and among them, the value of? That minimizes the WSE (?) Is defined as? _SA .

At this time, α _min , α _max And n values are determined experimentally. The obtained? _SA value is determined as a single speaker adaptation parameter expressing the tone color characteristic of the horse disordered patient.

The step S404 of generating a synthesized sound reflecting speaker characteristics by applying the extracted speaker adaptation parameter is as follows.

The main feature parameters that constitute the acoustic model of the statistical speech synthesis system are the Mel-Generalized Cepstrum (MGC) parameter, the log F0 parameter, and the state duration parameter.

In the conventional speech synthesis step using MGC parameters, a bilinear transform using a value of α _M = 0.42 is used for the mel-scale frequency conversion.

Therefore, this pair with the speaker adaptation parameters obtained in step 2 of the invention, linear transform and Mel-when the a bilinear transformation for scaling the frequency converted serial connection represented by a single bilinear transformation, a pair of case linear transformation parameter α _F is the following mathematical Expression 6 is expressed as follows.

Equation (6) is the same as in the case where a normal VTLN-based speaker adaptation technique is applied to statistical speech synthesis.

In the speech synthesis step using the MGC parameter of the statistical speech synthesis system, instead of the existing bilinear conversion using? _M = 0.42, a bilinear conversion using? _F in Equation 6 is performed, A synthetic sound is generated.

The step S405 of tuning the tone according to the taste of the user based on the result of the generated synthesized sound can be selectively performed as follows.

It is possible to generate a synthetic sound reflecting the tone characteristic of the horse disordered person in the statistical speech synthesis system only by the steps S401 to S404.

In the embodiment of the statistical speech synthesis system and method according to the present invention, the tone color changing process using only the spectrum characteristic conversion of the synthesized tone is performed, and the rhythm characteristics of the synthesized tone are changed without performing the tone color change, The rhyme characteristics are used as it is because it is difficult for the horse disordered person to express the natural rhyme because the rhyme characteristics directly obtained from the speech of the impaired person are very different from the rhyme characteristics of the synthesized sound actually desired by the person with impairment.

However, in case that the tone color change using only the spectrum conversion does not sufficiently satisfy the preference of the horse disorder person, a step that can be tuned according to the user's preference can be selectively added to both the rhythm characteristic and the spectrum characteristic of the synthetic tone.

라고 하면, 다음 수학식 7과 같이 음색 변환을 수행할 수 있다.That is, among the prosodic features of the synthetic sound, the intonation characteristics representing the high and low sound are represented by the log F0 parameter of the acoustic model. For the jth frame of the synthetic sound, the log F0 value is LF 0 ( j ) The F0 value

, It is possible to perform tone color conversion as shown in Equation (7).

Here, LF 0 _SA is a user-specified parameter for converting the intonation characteristics of the synthetic sound. If LF 0 _SA > 0, the sound becomes higher and LF 0 _SA &Lt; 0, the sound becomes lower.

Adjust the value of LF 0 _SA and listen to the synthesized sound and select the most satisfactory LF 0 _SA value for the person with speech impairment.

The process of Equation (7) can be implemented by modifying the log F0 model parameter among the acoustic model parameters instead of implementing each frame in the synthetic sound generating step.

In the case of spectral conversion using the bilinear transformation parameter α _F obtained through Equation (6), if the person with terminal impairment wishes to further change the spectral characteristics, the synthesized sound is adjusted while adjusting the α _F value, Select a satisfactory α _F value.

The system and method of statistical speech synthesis that reflects the individual tone color according to the present invention generates a synthesized tone having the main tone characteristics of the person only by the vocalization of the short vowel of the horse disordered person and the short vowel voice and the corresponding short tone voice among the acoustic models of the statistical speech synthesis system And extracts the speaker adaptation parameters in the form of a bilinear transformation parameter based on the formant frequency to generate a synthetic sound reflecting the individual tone characteristics in the statistical speech synthesis system to improve the communication.

As described above, it will be understood that the present invention is implemented in a modified form without departing from the essential characteristics of the present invention.

It is therefore to be understood that the specified embodiments are to be considered in an illustrative rather than a restrictive sense and that the scope of the invention is indicated by the appended claims rather than by the foregoing description and that all such differences falling within the scope of equivalents are intended to be embraced therein It should be interpreted.

20. Short vowel sound collection section 30. Speaker adaptation parameter extraction section
40. Synthetic sound generating section 50. Synthetic sound tuning section

Claims (10)

A short vowel voice collection unit for collecting short vowel voice from a horse disordered patient;
And a second formatting unit for dividing the short vowel sound data collected by the short vowel sound collection unit into a frame unit and a 2K + 1 frame including a frame having the largest frame energy and K frames before and after the frame, A formant frequency extracting unit for extracting a frequency of a first formant F1, a second formant F2 and a third formant F3; and a median of formant frequencies of each frame extracted by the formant frequency extracting unit and determining a median determining unit which, formant frequency (F 1, F 2 and F 3) is extracted from a particular vowel of the end disorder frequency of formant of the acoustic model corresponding to the vowel (F 1 _M, F 2 _M and F 3 _M ) as a speaker adaptation parameter, and a speaker adaptation parameter determiner for obtaining a bilinear transform coefficient &lt; RTI ID = 0.0 &gt; _SA &lt; / RTI &gt; that minimizes the weighted squared error sum when expressed as a bilinear transform, A speaker adaptation parameter extraction unit for extracting a conversion speaker adaptation parameter;
A synthesized voice generating unit for generating a synthesized voice reflecting speaker characteristics by applying the speaker adaptation parameter extracted by the speaker adaptation parameter extracting unit;
And a synthetic tone tuning unit for selectively tuning a tone color of the synthesized voice generated by the synthetic voice generating unit. delete Collecting short vowel speech from a horse disordered patient;
And the step to separate the collected vowel sound data in units of frames, the first formant (F1) for the 2 K + 1 frames consisting of a frame energy as the largest frame, and that after K frames, the second formant ( F2), third formant (F3) and a step of extracting frequency, extracted from the step, and a particular short vowels in speech disorder patient for determining a median value (median) of the formant frequency of the frames extracted formant frequency (F 1 , F 2, and F 3) are expressed as a bilinear transform of the formant frequencies ( F 1 _M , F 2 _M, and F 3 _M ) of the acoustic model corresponding to this short vowel, the weighted squared error sum is minimized is a bilinear transform coefficients comprising: a step to obtain the α _SA with speaker adaptation parameters, extracts the speaker adaptation parameters Four pairs of treatments based on a linear transformation;
And generating a synthesized voice that reflects the speaker characteristics by applying the extracted speaker adaptation parameter. The method according to claim 3, wherein in the step of generating a synthesized sound reflecting the speaker characteristics by applying the extracted speaker adaptation parameter,
And selectively tuning the tone color based on the result of the generated synthetic tone. delete 4. The method according to claim 3, wherein, in the step of obtaining the bilinear transform coefficient alpha _SA with the weighted squared error sum as a speaker adaptation parameter,
The bilinear transformation in frequency (Hz)

ego,
here

Wow

Characterized in that each of the plurality of frequency bands represents a frequency before and after the bilinear transformation, and a is a bilinear transformation parameter. The method according to claim 6, wherein the binarization conversion in the frequency (Hz)

Functions that are summarized for

The

ego,
Binary Linear Conversion Functions

( F 1, F 2, and F 3) extracted from a specific short vowel of a horse disordered person and formant frequencies ( F 1 _M , F 2 _M, and F 3 _M ) of the acoustic model corresponding to the short vowel, The weighted squared error sum with the binarized transformed values of &lt; RTI ID = 0.0 &gt;

ego,
Here,

Is a weight in consideration of the reliability of the estimated i- th formant frequency considering the reliability of the formant frequency extraction is lower than that of the voice of the general person in the case of speech of a horse disordered person. Way. 8. The method of claim 7, wherein the weighting factor that takes into account the reliability of the individual formant frequencies

The calculation,

ego,
here

Wow

Is the i- th formant frequency of the 2 K + 1 frames used

And a function g ( x ) is a function having a monotone increasing characteristic for a range of x & gt ; = 0. The statistical speech synthesis method reflecting the personal tone. 5. The method according to claim 4, wherein in tuning the tone color based on the result of the generated synthesized tone,
Among the prosodic features of the synthetic sound, the intonation characteristics representing the high and low sound are represented by the log F0 parameter of the acoustic model. For the jth frame of the synthetic sound, the log F0 value is LF0 ( j )

In the tone color conversion,

Lt; / RTI &gt;
Here, LF 0 _SA is a user-specified parameter for converting the intonation characteristics of the synthetic sound. If LF 0 _SA > 0, the sound becomes higher and LF 0 _SA &Lt; 0, the sound is lowered, and the LF 0 _SA value is adjusted, and the LF 0 _SA value is selected by listening to the synthesized sound while adjusting the LF 0 _SA value. 5. The method according to claim 4, wherein in tuning the tone color based on the result of the generated synthesized tone,
If the bilinear conversion using the speaker adaptation parameters and the bilinear conversion for the mel-scale frequency conversion are expressed as a single bilinear conversion by serial connection,
Bilinear transformation parameter α is _F,

ego,
Where α _SA is a bilinear transform coefficient that minimizes the weighted square error sum, α _M is a bilinear transform coefficient for the Mel-scale frequency transform,
In the case of spectral conversion using the bilinear transformation parameter α _F , if the person with terminal impairment wishes to change the spectral characteristics additionally, he or she can listen to the synthesized sound while adjusting the α _F value and obtain the most satisfactory α _F value And selecting a statistical speech synthesis method based on the personalized tone.

Images