JP2005134685A - Vocal tract shaped parameter estimation device, speech synthesis device and computer program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To estimate a vocal tract shaped parameter as accurate as possible. <P>SOLUTION: A vocal tract shaped parameter estimation device 20 includes an initial value calculation part 40 to calculate the initial value of the vocal tract shaped parameter of a living body from image information 22 that indicates the structure of the inside of a living body during prescribed utterance, a standard spectral parameter calculation part 44 to calculate a spectral parameter from a speech waveform 24 equivalent to this utterance, a spectral parameter conversion part 42 to optimize the vocal tract shaped parameter calculated by the initial value calculation part 40 by using a prescribed optimizing technique with a spectral parameter calculated by the standard spectral parameter calculation part 44 as a standard, a spectral parameter comparison part 46, and a vocal tract shaped parameter correction part 48. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to a speech synthesizer, and more particularly, to a technique for clarifying synthesized speech in a method of synthesizing speech from vocal tract shape parameters.

  Various interfaces have been proposed as interfaces between humans and mechanical systems represented by computer systems. Among them, speech is one that has been regarded as promising for a long time and has been used particularly frequently recently. By using voice, communication similar to that between humans can be realized between humans and mechanical systems.

  Among speech synthesis techniques used in speech communication, there is a method of synthesizing speech by simulating the physiological motion of a human speech voicing organ. In this method, the human speech generation model is divided into a sound source model that generates sound waves and a vocal tract model that represents articulation based on the shape of the vocal tract. The vocal tract model is often described by a vocal tract shape parameter indicating a shape of the vocal tract, particularly a cross-sectional shape.

  In this method, if the characteristics of a certain person's sound source and the vocal tract shape parameters corresponding to the vocal tract shape are accurately known, it is possible to generate clear speech that closely resembles that person's voice. For example, it is possible to synthesize speech that closely resembles the original speech for a person who has lost the speech organ for some reason and can no longer speak.

  Speech synthesis using vocal tract shape parameters is a promising speech synthesis method because of its high parameter compression ratio, high robustness to parameter deformation, and the possibility of being applicable to voice quality conversion. is there.

  In this method, how to estimate the vocal tract shape parameter becomes a problem. With respect to this point, Patent Document 1 refers to a correspondence relationship between a vocal tract shape parameter determined in advance based on a standard speaker's vocal tract model and a formant frequency of an actual speech, and A technique for estimating vocal tract shape parameters is disclosed.

JP-A-11-327592, paragraphs 0020 to 0042

  However, it is generally known that similar speech can be generated from different vocal tract shapes in human speech. For this reason, there is a problem that when the vocal tract shape parameter is obtained from the speech waveform as in the technique described in Patent Document 1, the solution cannot be obtained uniquely. As a result, there is a problem that the result of speech synthesis is not favorable.

  Therefore, an object of the present invention is to provide a vocal tract shape parameter estimation device capable of estimating a vocal tract shape parameter as accurately as possible, and to perform speech synthesis using the vocal tract shape parameter thus estimated. Is to provide a device.

  An apparatus for estimating a vocal tract shape parameter according to a first aspect of the present invention includes a vocal tract shape parameter calculating unit for calculating a vocal tract shape parameter of a living body from image information representing a structure inside the living body at a predetermined utterance. The vocal tract calculated by the vocal tract shape parameter calculation means based on the spectral parameter calculation means for calculating the spectral parameters from the speech waveform equivalent to the predetermined utterance and the spectral parameters calculated by the spectral parameter calculation means Vocal tract shape parameter optimization means for optimizing the shape parameter by a predetermined optimization method.

  The vocal tract shape parameter is calculated from the image information representing the actual internal structure of the living body, and the initial value is used as an initial value to optimize the spectral parameter obtained from the actual speech waveform. Since an initial value is a vocal tract shape parameter obtained from actual image information, an optimal value can be obtained as a vocal tract shape parameter for generating a speech waveform close to an actual speech waveform by optimization processing. As a result, if speech synthesis is performed using the vocal tract shape parameters thus estimated, clear speech close to the speech of the speaker can be synthesized.

  Preferably, the spectrum parameter calculation means includes first spectrum envelope calculation means for calculating a spectrum envelope of a speech waveform equivalent to a predetermined utterance, and the vocal tract shape parameter optimization means includes the given vocal tract shape parameter Calculated by the second spectral envelope calculating means for calculating the spectral envelope of the speech synthesized based on the first spectral envelope, the spectral envelope calculated by the first spectral envelope calculating means, and the second spectral envelope calculating means. A spectral envelope comparison means for comparing the spectral envelope and calculating the difference thereof, and correcting the vocal tract shape parameter by an optimization method so that the difference calculated by the spectral envelope comparison means approaches zero; The vocal tract shape parameter correcting means for giving to the spectrum envelope calculating means and a predetermined end condition are established. In response, the second condition detection means for outputting the vocal tract shape parameter modified by the vocal tract shape parameter modification means, and the second spectral envelope calculation means includes a first calculation of the vocal tract shape parameter. The spectrum envelope is calculated by using the vocal tract shape parameter calculated by the vocal tract shape parameter calculation means for the first time, and the vocal tract shape parameter given from the vocal tract shape parameter correction means in other cases.

  In the optimization, optimization based on a spectrum envelope reflecting the vocal tract shape is performed. The influence of the fine structure on the spectrum by the sound source can be removed, and the vocal tract shape parameters can be estimated accurately.

  Preferably, the vocal tract shape parameter correcting means includes means for correcting the vocal tract shape parameter by a simulated annealing method so that the difference approaches zero.

  By using the simulated annealing method, there is a high possibility that a global optimum solution can be obtained instead of a local solution.

  More preferably, the end condition detecting means is a vocal tract corrected by the vocal tract shape parameter correcting means in response to the difference calculated by the spectrum envelope comparing means being smaller than a predetermined threshold value. Means for outputting the shape parameters.

  The end condition detection means is configured such that the difference calculated by the spectrum envelope comparison means is smaller than a predetermined threshold value, or the second spectrum envelope calculation means, spectrum envelope comparison means, and vocal tract shape parameter correction Means may be included for outputting the vocal tract shape parameter modified by the vocal tract shape parameter modifying means in response to the repetition of the processing by the means being completed a predetermined number of times.

  By terminating the process according to a predetermined termination condition, an estimated value can be obtained within a certain amount of time.

  More preferably, the vocal tract shape parameter calculating means calculates a vocal tract cross-sectional area function by calculating a cross-sectional area of a cross section at a plurality of locations of the biological vocal tract from a plurality of stereoscopic cross-sectional images of the biological body at a predetermined utterance. A vocal tract cross-sectional area function specifying means for determining, and a means for extracting a time-varying pattern of the vocal tract cross-sectional area function specified by the vocal tract cross-sectional area function specifying means and outputting it as a vocal tract shape parameter.

  A speech synthesizer according to a second aspect of the present invention performs speech synthesis using one of the above vocal tract shape parameter estimation devices and a vocal tract shape parameter estimated by the vocal tract shape parameter estimation device. Voice synthesis means.

  With the above vocal tract shape parameters, the vocal tract shape parameters for synthesizing speech close to actual speech can be accurately estimated. Since speech synthesis is performed using the vocal tract shape parameters, the synthesized speech is naturally close to the speech of the speaker and clear.

  The computer program according to the third aspect of the present invention, when executed by a computer, causes the computer to operate as any of the above vocal tract shape parameter estimation devices.

  A computer program according to the fourth aspect of the present invention, when executed by a computer, causes the computer to operate as the above-described speech synthesizer.

  As described above, according to the present invention, an optimal value can be obtained as a vocal tract shape parameter for generating a speech waveform close to an actual speech waveform. As a result, by performing speech synthesis using the estimated vocal tract shape parameters, it is possible to synthesize clear speech that is close to the actual speech of the speaker.

  In the vocal tract shape parameter estimation device and the speech synthesizer according to the present embodiment, when the vocal tract shape parameter is obtained, a raw resonance signal such as MRI (Magnetic Resonance Imaging) or CT (Computed Tomography) is used. Using the vocal tract shape of the speaker when speaking in a predetermined task, captured by a device capable of capturing a three-dimensional shape in the body, and the voice uttered in the same task as when imaging, The vocal tract shape parameters are obtained by simulating the propagation of sound waves. As an image having a three-dimensional shape, for example, an image obtained by capturing a plurality of cross-sectional images of a living body and restoring the images can be used as in an embodiment described later. Other than that, for example, it can be used if it can directly image a three-dimensional shape.

  In this method, when obtaining the vocal tract shape parameter, the vocal tract shape parameter is obtained from the vocal tract shape at the time of utterance imaged by MRI or CT. A spectrum parameter is obtained from a speech waveform obtained by performing speech synthesis from the vocal tract shape parameter. The difference between this spectral parameter and the spectral parameter obtained by analyzing the actual utterance equivalent to the utterance at the time of imaging by MRI or the like is used as an evaluation value, and the vocal tract shape parameter is determined by an optimization method such as a simulated annealing method. To correct. This process is repeated, and correction is terminated when a predetermined termination condition is satisfied, and the vocal tract shape parameter at that time is finally set as the estimated vocal tract shape parameter. Using this vocal tract shape parameter, speech synthesis is performed by a conventionally known method. When the simulated annealing method is used, there is a high possibility that a global optimum solution that is not a local solution can be obtained.

  In the present embodiment, as will be described later, a time variation pattern of the vocal tract cross-sectional area function is used as the vocal tract shape parameter. However, as the vocal tract shape parameter, various parameters can be adopted according to the speech synthesis method as long as the vocal tract shape can be expressed for speech synthesis.

  FIG. 1 shows a block diagram of a speech synthesizer 20 according to the present embodiment. The voice synthesizer 20 uses the MRI imaging data 22 relating to the in vivo of the speaker captured by MRI when the speaker is speaking in a certain task, and the voice waveform of the voice actually spoken by the speaker in the same task. 24, the vocal tract shape parameter 26 of the speaker is obtained, and a synthesized speech signal 32 is generated using the vocal tract shape parameter 26. The details of the MRI imaging data 22 will be described later.

  Referring to FIG. 1, the speech synthesizer 20 receives an initial value calculation unit 40 for calculating a vocal tract shape parameter of a living body from MRI imaging data 22 and the vocal tract shape parameter, and receives the value of the vocal tract shape parameter And a spectral parameter conversion unit 42 for calculating a spectral parameter of speech synthesized based on the above. As will be described later, the spectral parameter conversion unit 42 uses the vocal tract shape parameter output from the initial value calculation unit 40 at the beginning of the vocal tract shape parameter estimation process, but the corrected vocal tract shape parameter is used in the subsequent processing. To calculate the spectral parameters. In the present embodiment, a spectrum envelope is used as a spectrum parameter.

  The speech synthesizer 20 further includes a reference spectrum parameter calculation unit 44 for calculating a reference spectrum parameter serving as a reference in estimating the vocal tract shape parameter from the speech waveform 24 obtained from the utterance of the speaker, and a spectral parameter conversion. The spectral parameter calculated by the unit 42 and the reference spectral parameter obtained by the reference spectral parameter calculating unit 44 are compared, and the spectral parameter comparing unit 46 for outputting the difference is compared with the spectral parameter comparing unit 46. Based on the output difference, the vocal tract shape parameter for applying to the spectral parameter conversion unit 42 by correcting the vocal tract shape parameter so that the difference becomes small (approaching 0) by the simulated annealing method. And a correction unit 48. In the present embodiment, the spectral parameter comparison unit 46 corrects the vocal tract shape parameter when the magnitude of the difference obtained as a result of the comparison is smaller than a predetermined threshold value or a predetermined number of times. When the end condition is satisfied, the process is terminated, and the vocal tract shape parameter at that time (the vocal tract shape parameter finally output by the vocal tract shape parameter correction unit 48) is used as the optimized vocal tract shape parameter 26. Has a function to output.

  The speech synthesizer 20 further uses the sound source 30, the optimized vocal tract shape parameter 26 obtained by the above processing, the sound source 30, and the optimized vocal tract shape parameter 26 to generate a synthesis target phoneme sequence. 50 and a voice synthesizer 28 for generating a synthesized voice signal 32 in accordance with 50. The synthesis method used by the speech synthesizer 28 may be any method as long as it uses a vocal tract shape parameter. The sound source 30 is preferably optimized for the voice of the speaker, but other sound sources may be used.

  FIG. 2 shows a cross-sectional image of the mid-sagittal plane of the human body at a certain point during speech obtained by MRI as an example. The MRI imaging data 22 includes the cross-sectional shape of the human body on a plurality of planes parallel to the plane other than the median sagittal plane at the same time point. As a result, the internal shape of the living body can be three-dimensionally restored from the plurality of images.

  Based on the three-dimensional shape inside the living body restored in this way, the initial value calculation unit 40 calculates the cross-sectional area of the vocal tract at a certain point in time and the distance along the vocal tract from a predetermined position (eg glottis) to the cross-section. Seeking a relationship between. This relationship is called a vocal tract cross-sectional area function. The vocal tract cross-sectional area function varies with time depending on the speech. Therefore, as described above, it is possible to obtain a time change pattern of the vocal tract cross-sectional area function when the speaker speaks in a certain task.

  In the present embodiment, as shown by the cross sections A to L in FIG. 2, the cross-sectional areas of the vocal tract are calculated at 12 locations and set as vocal tract shape parameters. In this case, the position of the vocal cords which is the origin of the vocal tract cross-sectional area function is shown as a point 50 in FIG.

  FIG. 3 is a cross-sectional view of the vocal tract and its surroundings in cross sections A to L. These sectional views are obtained from a three-dimensional shape obtained by synthesizing a plurality of MRI sectional shapes as described above. In FIG. 3, the vocal tract cross section is indicated by reference numerals 60 to 82. Thus, since the cross-sectional shape image of the vocal tract can be synthesized from the MRI image, the vocal tract area in each cross section can be obtained from the image from these drawings. At this time, the vocal tract is imaged in a form that can be distinguished from other parts as a space having nothing other than air.

  A technique for restoring a three-dimensional image from a plurality of MRI images, a technique for obtaining a cross-sectional image including a vocal tract cross section at each position of the vocal tract from the obtained three-dimensional image, and a vocal tract from each vocal tract cross-sectional image As a technique for obtaining the cross-sectional area, there are known image processing techniques, and in the present embodiment, these image processing techniques are used.

  FIG. 4 is a graph showing the time variation pattern of the vocal tract cross-sectional area function obtained from the MRI image. This graph is a graph of vocal tract cross-sectional area functions obtained at each time when a speaker pronounces “Aiueo” along the time axis. In FIG. 4, the horizontal axis represents the vocal tract cross section, the distance from the vocal cord (point 50 in FIG. 2), and the vertical axis represents the vocal tract cross sectional area of each vocal tract cross section. The axis labeled “Iueo” is the time axis, and the time axis indicates the elapsed time from the start of the utterance in milliseconds. The cross-sectional area shape and its time change pattern expressed in this way are the vocal tract shape parameters extracted by the initial value calculation unit 40.

  An example of the vocal tract cross-sectional area function extracted from the MRI imaging data 22 at a certain time is shown in FIG. Such a vocal tract cross-sectional area function is obtained for each time.

  When the vocal tract cross-sectional area function is obtained in this way, the spectral distribution of speech obtained as a result of speech synthesis based on the function can be calculated using a known speech processing tool. An example of the result is shown in FIG. This conversion is performed by the spectral parameter converter 42 shown in FIG. Note that the spectral parameter conversion unit 42 converts the vocal tract shape parameter from the initial value calculation unit 40 into a spectral parameter at the beginning of the iterative processing, but after the correction given from the vocal tract shape parameter correction unit 48 at subsequent iterations. Are converted into spectral parameters and supplied to the spectral parameter comparison unit 46.

  FIG. 7 shows an example of a spectrum obtained by analyzing the actual speech sound corresponding to the speech when the MRI image is captured. FIG. 8 shows the spectrum envelope of this spectrum. In general, the spectrum of an uttered speech has a fine structure determined by a vibration waveform of a sound source on a spectrum envelope determined by a vocal tract shape. Therefore, it is considered that the spectrum envelope shown in FIG. 8 corresponds to the vocal tract shape at the time of actual speech.

  The function of the reference spectrum parameter calculation unit 44 is to extract a spectrum envelope as shown in FIG. 8 from the speech waveform 24 as a reference spectrum parameter.

  The spectral parameter comparison unit 46 has a function of comparing the spectral parameter output from the spectral parameter conversion unit 42 with the spectral parameter output from the reference spectral parameter calculation unit 44 and calculating the difference. When the difference becomes smaller than a predetermined threshold value or when a predetermined number of iterations are completed, the vocal tract shape parameter at that time is output as the optimized vocal tract shape parameter 26. FIG. 9 shows a spectrum when optimized so that the envelope of the spectrum converted by the spectrum parameter converting unit 42 approaches the envelope of the reference spectrum shown in FIG. The vocal tract cross-sectional area function at this time is shown as a graph 100 in FIG. For comparison, the initial value of the vocal tract cross-sectional area function shown in FIG.

  The vocal tract shape parameter correction unit 48 has a function of optimizing the vocal tract shape parameter by the simulated annealing method based on the difference between the spectral parameters calculated by the spectral parameter comparison unit 46. Not only the simulated annealing method but also other optimization methods can be used.

  The speech synthesizer 20 according to the present embodiment operates as follows. Prior to the following processing, it is assumed that the MRI imaging data 22 and the voice waveform 24 shown in FIG. 1 have already been prepared.

  When the MRI imaging data 22 is given, the initial value calculation unit 40 calculates an initial value of a vocal tract shape parameter (a vocal tract cross-sectional area function) from the MRI imaging data 22 and gives it to the spectrum parameter conversion unit 42. The spectral parameter conversion unit 42 converts the initial value of the vocal tract shape parameter into a spectral parameter and provides it to the spectral parameter comparison unit 46.

  On the other hand, the reference spectrum parameter calculation unit 44 calculates a reference spectrum parameter from the speech waveform 24 and gives it to the spectrum parameter comparison unit 46.

  The spectral parameter comparison unit 46 compares the spectral parameter given from the spectral parameter conversion unit 42 with the spectral parameter given from the reference spectral parameter calculation unit 44, and if the difference is smaller than a predetermined threshold, Are output as optimized vocal tract shape parameters 26 (ie, the initial vocal tract shape parameters). Otherwise, the spectral parameter comparison unit 46 gives the difference obtained as a result of the comparison to the vocal tract shape parameter correction unit 48.

  When the vocal tract shape parameter correction unit 48 receives the spectral parameter difference data from the spectral parameter comparison unit 46, the vocal tract shape parameter correction unit 48 corrects the initial value of the vocal tract shape parameter by the simulated annealing method so that the difference approaches zero. To the spectral parameter converter 42.

  The spectral parameter conversion unit 42 calculates a spectral parameter from the vocal tract shape parameter given from the vocal tract shape parameter correction unit 48 and gives it to the spectral parameter comparison unit 46.

  Thereafter, the optimization process of the vocal tract shape parameters by the spectral parameter comparison unit 46, the vocal tract shape parameter correction unit 48, and the spectral parameter conversion unit 42 is repeated. Then, when the difference between the spectrum parameter calculated by the spectrum parameter conversion unit 42 and the reference spectrum parameter calculated by the reference spectrum parameter calculation unit 44 becomes smaller than a threshold value, the spectrum parameter comparison unit 46 The road shape parameter is output as the optimized vocal tract shape parameter 26. Even if the difference does not become smaller than the threshold value, the spectral parameter comparison unit 46 sets the vocal tract shape parameter at that time as the optimized vocal tract shape parameter 26 even when the above-described repetitive processing is executed a predetermined number of times. Output.

  Thus, an optimized vocal tract shape parameter 26 is obtained for a speaker. At the time of speech synthesis, when the speech synthesizer 28 is provided with a phoneme string 50 that is a synthesis target, the speech synthesizer 28 performs a vocal tract filtering process defined by the optimized vocal tract shape parameter 26 on the speech signal from the sound source 30. The synthesized voice signal 32 is output.

  As described above, according to the speech synthesizer 20 according to the present embodiment, the actual vocal tract shape parameter of the speaker is set as an initial value, and the reference obtained from the actual speech waveform is obtained from the vocal tract shape parameter. The vocal tract shape parameters are optimized so that speech having a spectrum close to the spectrum can be synthesized. Therefore, a value close to the vocal tract shape parameter at the time of actual speech of the speaker can be estimated, and clear speech synthesis that is close to the speech of the speaker can be realized.

  According to the speech synthesizer 20 of the present embodiment, speech close to the speaker's speech can be synthesized from the captured image of the cross section of the living body. Therefore, it is possible to synthesize natural speech that resembles the original speech of the patient before excision of the speech organ based on the captured image of the cross-section of the patient whose speech organ has been removed for some reason. It becomes.

  The speech synthesizer 20 described above can be realized by computer hardware and a computer program executed on the computer hardware. Any computer hardware can be used as long as it includes an apparatus capable of receiving image information such as an MRI image and audio waveform data.

  Furthermore, for each component for realizing the speech synthesizer 20 described above, a tool kit for digital signal processing for realizing them can be easily obtained. In this embodiment, there is a particular feature in the way of combining these parts. Therefore, based on the above description of the present embodiment, a person skilled in the art can easily create a computer program for realizing the speech synthesizer 20 according to the present embodiment.

  The embodiment disclosed herein is merely an example, and the present invention is not limited to the above-described embodiment. The scope of the present invention is indicated by each of the claims after taking into account the description of the detailed description of the invention, and all modifications within the meaning and scope equivalent to the wording described therein are intended. Including.

It is a block diagram of the speech synthesizer concerning one embodiment of the present invention. It is a figure which shows an example of an MRI image. It is a figure which shows an example of the vocal tract cross section obtained from the MRI image. It is a graph which shows the time change pattern of a vocal tract cross-sectional area function. It is a graph which shows an example of the vocal tract cross-section function obtained from the MRI image. It is a graph which shows an example of the spectrum acquired from the vocal tract cross-sectional area function. It is a graph which shows an example of the spectrum acquired from the actual speech waveform. It is a graph which shows the spectrum envelope of the spectrum shown in FIG. It is a graph which shows an example of the spectrum acquired from the vocal tract shape parameter optimized by the speech synthesizer concerning one embodiment of the present invention. It is a graph which compares and shows the initial value of a vocal tract cross-sectional area function, and the vocal tract cross-sectional area function after optimization.

Explanation of symbols

  20 speech synthesizer, 22 MRI imaging data, 24 speech waveform, 26 optimized vocal tract shape parameter, 28 speech synthesizer, 30 sound source, 32 synthesized speech signal, 40 initial value calculator, 42 spectral parameter converter, 44 Reference spectral parameter calculation unit, 46 spectral parameter comparison unit, 48 vocal tract shape parameter correction unit

Claims (9)

Vocal tract shape parameter calculating means for calculating a vocal tract shape parameter of the living body from image information representing a structure inside the living body at a predetermined utterance;
Spectral parameter calculation means for calculating a spectral parameter from a speech waveform equivalent to the predetermined utterance;
Vocal tract shape parameter optimization means for optimizing the vocal tract shape parameter calculated by the vocal tract shape parameter calculation means with reference to the spectral parameter calculated by the spectral parameter calculation means Estimating device. The spectrum parameter calculation means includes first spectrum envelope calculation means for calculating a spectrum envelope of a speech waveform equivalent to the predetermined utterance,
The vocal tract shape parameter optimization means includes:
A second spectral envelope calculating means for calculating a spectral envelope of a voice synthesized based on a given vocal tract shape parameter;
A spectrum envelope comparing means for comparing the spectrum envelope calculated by the first spectrum envelope calculating means and the spectrum envelope calculated by the second spectrum envelope calculating means and calculating the difference thereof;
A vocal tract shape parameter correcting means for correcting the vocal tract shape parameter by the optimization method so that the difference calculated by the spectral envelope comparison means approaches 0, and providing the second spectral envelope calculation means;
Ending condition detecting means for outputting the vocal tract shape parameter corrected by the vocal tract shape parameter correcting means in response to establishment of a predetermined ending condition,
The second spectral envelope calculation means calculates the vocal tract shape parameter calculated by the vocal tract shape parameter calculation means for the first time of the calculation of the vocal tract shape parameter, and otherwise the vocal tract shape parameter. The vocal tract shape parameter estimation apparatus according to claim 1, wherein a spectrum envelope is calculated using each of the vocal tract shape parameters given from the parameter correction means. The estimation of the vocal tract shape parameter according to claim 2, wherein the vocal tract shape parameter correction means includes means for correcting the vocal tract shape parameter by a simulated annealing method so that the difference approaches zero. apparatus. The termination condition detection means is a vocal tract corrected by the vocal tract shape parameter correction means in response to the magnitude of the difference calculated by the spectrum envelope comparison means being smaller than a predetermined threshold value. 4. The vocal tract shape parameter estimation apparatus according to claim 2, further comprising means for outputting a shape parameter. The termination condition detection means is configured such that the magnitude of the difference calculated by the spectrum envelope comparison means is smaller than a predetermined threshold, or the second spectrum envelope calculation means, the spectrum envelope comparison means, and 3. A means for outputting the vocal tract shape parameter modified by the vocal tract shape parameter modifying means in response to the repetition of the processing by the vocal tract shape parameter modifying means being completed a predetermined number of times. Or the estimation apparatus of the vocal tract shape parameter of Claim 3. The vocal tract shape parameter calculating means includes:
Vocal tract cross-sectional area function specification for determining a vocal tract cross-sectional area function by calculating cross-sectional areas of cross-sections at a plurality of locations of the vocal tract of the living body from a plurality of stereoscopic cross-sectional images of the living body at the time of the predetermined utterance Means,
6. A means for extracting a time-varying pattern of the vocal tract cross-sectional area function specified by the vocal tract cross-sectional area function specifying means and outputting it as the vocal tract shape parameter. The vocal tract shape parameter estimation apparatus according to claim 1. The vocal tract shape parameter estimation device according to any one of claims 1 to 6,
A speech synthesizer including speech synthesis means for performing speech synthesis using the vocal tract shape parameter estimated by the vocal tract shape parameter estimation device. A computer program that, when executed by a computer, causes the computer to operate as the vocal tract shape parameter estimation device according to any one of claims 1 to 6. A computer program that, when executed by a computer, causes the computer to operate as the speech synthesizer according to claim 7.

Images