JP4644879B2 - Data generator for articulation parameter interpolation and computer program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an interpolation data creating apparatus capable of creating data for articulatory parameter interpolation which reflects shape change of a speech organ by speech action, by simpler processing than transmissive imaging method. <P>SOLUTION: The interpolation data creating apparatus of articulatory parameters for creating interpolation data when the articulatory parameters for synthesizing voice which continuously changes from a first phoneme to a second phoneme are created by interpolation between known articulatory parameters for synthesizing the first and the second phoneme includes: a change rate calculation means 60 for calculating transition of a change rate of a predetermined sound feature amount from the first phoneme to the second phoneme, from an input voice signal including the first phoneme and the second phoneme which are continuous; and an interpolation data creating means 64 for creating the interpolation data using the change rate. <P>COPYRIGHT: (C)2007,JPO&amp;INPIT

Description

  The present invention relates to a technique for performing modulation on a sound source and generating a parameter for smoothing synthesized speech when speech synthesis is performed.

  In recent years, various interfaces have been proposed as an interface between a human and a mechanical system represented by a computer system. Among them, voice is one of the most frequently used recently. By using voice, communication between humans and mechanical systems can be realized in a form close to that between humans.

  There are the following two methods for voice synthesis technology for realizing communication by voice. One is a method of synthesizing a voice by taking out phoneme pieces from a voice waveform of a voice recorded in advance and connecting them. The other is a method of synthesizing speech by simulating changes in the shape of a human speech organ.

  Since the change in the shape of the speech organ is more gradual than the change in the speech waveform, the latter method can synthesize a smoother speech. For this reason, this method has attracted particular attention in recent years.

  In this type of speech synthesis method, a signal from a sound source is converted into an electric signal whose characteristics change depending on parameters representing the shape of the speech organ (for example, vocal tract cross-sectional area function, vocal tract length, opening area, etc .; hereinafter referred to as “articulation parameters”) The audio signal is synthesized by passing through a filter composed of a circuit. For example, the articulation parameter differs between when a sound “a” is uttered and when a sound “i” is uttered. Therefore, various voices can be synthesized by changing the articulation parameters. Of course, this method is currently realized by digital technology using a computer and software.

  In order to synthesize a more natural continuous speech, the articulation parameter when speaking a certain speech and the articulation parameter when speaking the next speech are changed to the shape change of the speech organ during continuous speech. It is necessary to interpolate in a manner consistent with the temporal transition of. For that purpose, it is necessary to calculate the physical quantity reflecting the speed of the shape change of the uttering organ when actually speaking continuous speech and its temporal transition by some method.

Non-Patent Document 1 discloses a prior art that uses the speed of the shape change of a speech organ. This is obtained by temporally interpolating the vocal tract cross-sectional area function in order to synthesize continuous speech, and parameters for temporal interpolation are temporally the average luminance value of an MRI (Magnetic Resonance Imaging) moving image. It is obtained from change. According to this method, since the synthesis is based on an actual articulation operation, there is an advantage that a natural continuous sound can be obtained.
Speech synthesis by temporal interpolation of vocal tract cross section function, Hironori Takemoto, Kiyoshi Honda, Tatsuya Kitamura, Parham Moktari, Hiroyuki Hirai. Proceedings of the Acoustical Society of Japan, pp157-158, March 2005.

  According to the technique disclosed in Non-Patent Document 1, there is a great advantage that a natural continuous sound based on an actual articulation operation can be obtained. However, with this method, it is necessary to take an MRI video while letting humans speak. As is well known, it takes time to capture an MRI moving image. Also, the equipment necessary for taking MRI moving images is large, and there are complicated procedures for its use. Furthermore, the image processing of MRI moving images is difficult and expensive.

  It is desirable that the same effect as that disclosed in Non-Patent Document 1 can be obtained more easily and with a simple process.

  Accordingly, an object of the present invention is to provide an interpolation data generating apparatus capable of generating articulation parameter interpolation data reflecting the shape change of an utterance organ due to an actual utterance operation by a simpler process than a transmission imaging method such as an MRI moving image. It is to provide.

  The articulation parameter interpolation data generation device according to the first aspect of the present invention provides an articulation parameter for synthesizing speech that continuously changes from a first phoneme to a second phoneme. An apparatus for generating data for interpolation when generating by interpolating between known articulation parameters for speech synthesis of a speech, comprising: an input speech signal including a continuous first phoneme and a second phoneme; Change rate calculation means for calculating a change rate change rate of a predetermined acoustic feature amount from the first phoneme to the second phoneme, and interpolation data creation means for creating interpolation data using the change rate Including.

  According to this interpolation data generating apparatus, it is possible to generate interpolation data from speech so as to continuously connect the first phoneme and the second phoneme independent of each other. Unlike MRI moving images, audio can be easily acquired. In addition, it is considered that the change rate of the acoustic feature amount well reflects the change in the shape of the speech organ. Accordingly, it is possible to generate interpolation data that reflects a change in the shape of the uttered organ by an actual utterance operation by a simpler method than using a transmission imaging method such as an MRI moving image.

  Preferably, the rate-of-change calculating means calculates a predetermined spectrum change rate for each frame from the framing means for framing the input voice signal at predetermined time intervals and the voice signal framed by the framing means. Spectrum change rate calculating means.

  According to this interpolation data generation apparatus, the spectrum change rate for each frame can be calculated as the above-described change rate. The spectrum change rate is considered to well reflect the speed of the shape change of the speech organ, and an established technique can be used for the calculation. Therefore, it is possible to generate interpolation data that reflects a change in the shape of the utterance organ due to the actual utterance operation using the characteristics of the sound, which is simpler than using a transmission imaging method such as an MRI moving image.

  More preferably, the spectrum change rate calculating means is the adjacent minimum value of the predetermined spectrum change rate among the frames of the audio signal framed by the framing means from the audio signals from the first phoneme to the second phoneme. Frame combination for combining a plurality of frames, including first and second two frames that provide a second frame and a third frame that exists between the two frames and that provides a maximum value of a predetermined spectral change rate And function approximating means for approximating the transition of a predetermined spectrum change rate with a predetermined continuous function determined by the maximum value and the minimum value of the frames combined by the frame combination means.

  According to this interpolation data generation apparatus, the transition of a predetermined spectrum change rate is approximated by a continuous function having two minimum values corresponding to the sound center as both ends. Therefore, when generating interpolation data, an approximate function having a data amount smaller than the spectrum change rate itself can be used. Therefore, it is possible to simplify the resources required for generating the interpolation data. Therefore, using a simpler method, it is possible to generate interpolation data that reflects changes in the shape of the uttered organ due to the actual utterance operation.

  More preferably, the frame combination means is a minimum value detecting means for detecting the first and second frames based on the rate of change, and a maximum value detecting means for detecting the third frame based on the rate of change. A sound sequence assigning means for assigning the first phoneme to the first frame and the second phoneme to the second frame, and a frame sequence combining the frames from the first frame to the second frame. Frame sequence creating means for creating

  According to the interpolation data generation apparatus, specific phonemes are assigned to the first frame and the second frame, and a frame sequence is created so as to include the two frames. Therefore, when generating interpolation data, the interpolation target can be automatically determined and interpolated. Therefore, it is possible to generate interpolation data that reflects a change in the shape of the utterance organ by an actual utterance operation by a simpler method.

  More preferably, the function approximating means takes a predetermined constant in the first frame based on the rate of change in the first frame and the rate of change in the third frame, and the third frame in the third frame. Based on the first approximation means for obtaining a first approximation function that takes a value having a predetermined relationship with the value of the rate of change in, and on the basis of the rate of change in the third frame and the rate of change in the second frame And second approximation means for obtaining a second approximation function that takes a predetermined constant in the second frame and takes the same value as the first approximation function in the third frame.

  According to this interpolation data generation device, the type of approximation function can be changed appropriately based on the rate of change in the corresponding frame. The location where the change rate is maximum is not necessarily the intermediate position between the two phonemes to be interpolated, and often shifts before and after the location. Even in such a case, it is possible to approximate the change rate with high accuracy using a simple approximate function by applying different approximation functions before and after the portion where the change rate is maximum as described above. Therefore, the accuracy of approximation is improved, and the time required for the interpolation data generation processing using this approximation function can be shortened. Therefore, it is possible to generate interpolation data that reflects a change in the shape of the utterance organ by an actual utterance operation by a simpler method.

  More preferably, the interpolation data creating means normalizes to obtain a normalized integral function obtained by normalizing the integral function from the first frame to the second frame of the predetermined function obtained by the function approximating means. Includes integration means.

  According to this interpolation data generation device, since only the normalized integration function is calculated as the interpolation data, the time required to generate the interpolation data is shortened compared to the case where the articulation parameters themselves are interpolated. it can. Further, when the interpolation data is stored in a database (hereinafter referred to as DB), the capacity of the DB can be reduced. Therefore, it is possible to generate interpolation data that reflects the change in shape of the utterance organ due to the actual utterance operation by a simpler process.

  More preferably, the interpolation data creation means further interpolates between the articulation parameter for the first phoneme and the articulation parameter for the second phoneme by a normalization integration function, thereby providing the first phoneme. Means for calculating a mixing ratio at a predetermined time between the first phoneme and the second phoneme.

  According to this interpolation data generation device, only the mixture ratio is calculated as the interpolation data. Thus, the time required for generating the interpolation data can be shortened. Further, when the interpolation data is stored in the DB, the capacity of the DB can be reduced. Furthermore, it is not necessary to calculate the mixing ratio when synthesizing speech, compared to the case where the processing is stopped until the calculation of the normalized integration function. Accordingly, it is possible to generate interpolation data that can reflect the change in the shape of the utterance organ due to the actual utterance operation by simpler processing and can reduce the processing amount at the time of speech synthesis.

  More preferably, the interpolation data creation means further interpolates between the articulation parameter for the first phoneme and the articulation parameter for the second phoneme by a mixing ratio, thereby providing an articulation parameter at a predetermined time. Means for calculating a vocal tract cross-sectional area function.

  According to this interpolation data generation device, a vocal tract cross-sectional area function is calculated as interpolation data from the feature amount of speech. Therefore, when performing later speech synthesis using this vocal tract cross-sectional area function, the time required for speech synthesis can be reduced. Therefore, efficient speech synthesis can be performed. As a result, it is possible to generate interpolation data that reflects a change in the shape of the uttered organ due to the actual utterance operation by a process that is simpler than the process using the MRI moving image or the like.

  When executed by a computer, the computer program according to the second aspect of the present invention causes the computer to operate as any one of the articulation parameter interpolation data generation devices described above. Therefore, the same effect as the above-described articulation parameter interpolation data generation device can be obtained.

<Interpolation method based on spectral change rate>
In a speech synthesis system according to an embodiment of the present invention, a vocal tract cross-sectional area function, which is an interpolation parameter, is calculated from a time change of a speech signal, and speech synthesis is performed using the result. Audio signals can be obtained much more easily than MRI moving images, and signal processing is also easy. Before describing the details of this embodiment, the principle of the interpolation method employed in this embodiment will be described in detail. In the following description and drawings, the same reference numerals are assigned to the same components.

(1) Calculation of mixing ratio In order to synthesize the phoneme in the middle of the transition of the utterance voice from the first phoneme to the second phoneme, the articulation parameter in the middle of the transition is defined as the first phoneme and the second phoneme. The ratio of mixing the articulation parameters of the first phoneme and the second phoneme, which is generated by interpolation between the articulation parameters of the first phoneme. For example, in an actual utterance, when moving from “A” to “I”, the “A” sound element gradually decreases, and instead, the “I” sound element increases, and the sound continuously increases. Transition. By using the mixing ratio, it is possible to smoothly interpolate the articulation parameters for generating a continuous sound between vowel centers. In the present embodiment, the mixture ratio is approximated by a value obtained from the spectrum change rate of the speech voice. The spectrum change rate is a value indicating how fast the spectrum of the speech is changing (the expression will be described later), and reflects the movement of the speech organ during speech. Hereinafter, a calculation method of the mixing ratio using the spectrum change rate will be described.

  FIG. 1 shows a relationship between an example of a speech voice waveform and a spectrum change rate obtained from the voice waveform. Referring to FIG. 1, the frames represented by / a /, / i /, / u /, / e /, and / o / in the figure indicate the positions where the vowel centers of each vowel exist. During the production of vowels, the shape of the uttered organ is stable and the rate of spectral change is small. On the other hand, at the time of transition from one vowel to another, the shape of the speech organ changes. It is known that the rate of this change is not constant, the rate of change is small at the beginning and end of the change, and it reaches a maximum somewhere in the middle. In the example shown in FIG. 1, the shape change rate of the speech organ reaches a maximum at the maximum values 30, 32, 34 and 36 of the spectral change rate.

  As described above, the mixing ratio is a parameter for interpolating between adjacent vowel-centric articulation parameters to synthesize continuous and natural speech. Therefore, when calculating the mixing ratio, the frame sequence spectrum change rate is used so that the frames corresponding to the utterance centers of two adjacent vowels (frames with the minimum spectrum change rate) are arranged at both ends. Then, a mixing ratio for interpolating the sound between them may be calculated. However, in this case, since the last frame of the frame sequence corresponds to the first frame of the next frame sequence, it is not subject to the mixing ratio calculation.

  In addition, the audio signal to be processed often includes noise, and for this reason, the graph of the spectrum change rate obtained from the audio signal is deformed. In order to remove the influence of such noise and simplify the calculation, the curve drawn by the spectrum change rate is approximated by some function. In the present embodiment, the curve drawn by the spectrum change rate is approximated by a Gaussian function.

  FIG. 2 shows an example of a spectrum change rate (spectrum change rate 42) obtained from a voice signal when the voice changes from vowel / a / to vowel / i /, and a Gaussian approximating a curve drawn by this spectrum change rate. Function 44 is shown. Referring to FIG. 2, the vowel center of vowel / a / is arranged at the left end of the graph, and the vowel center of vowel / i / is arranged at the right end of the graph. For the above-described reason, the right end of the graph is the frame immediately before the frame including the vowel center of vowel / i /.

  The Gaussian distribution is a symmetrical distribution. However, the spectral change rate generally has different distributions on the left and right of the local maximum frame 40. Therefore, different Gaussian functions are assigned to the left and right of the local maximum frame. The Gaussian function used for this approximation is determined from the maximum value and the minimum value of the spectrum change rate. Note that the Gaussian function is asymptotic to the x-axis but never becomes zero. On the other hand, in the approximation by the Gaussian function, it is desirable that the articulation parameters coincide with the articulation parameters at both vowel positions at both ends of the frame sequence. For this purpose, the value of the approximation function becomes 0 at the vowel positions at both ends of the frame sequence. There is a need. Therefore, for example, once the above Gaussian function is obtained, a constant is subtracted from the Gaussian function so that the value of the function at both ends of the frame sequence is 0, or the constant is subtracted after multiplying the Gaussian function by a predetermined multiple. In other words, the function value is set to 0 at both ends (vowel positions) of the frame sequence.

  After obtaining a Gaussian function that approximates the spectrum change rate in this way, the mixing ratio is calculated from the Gaussian function as follows.

  The upper part of FIG. 3 shows the curve of the spectral change rate 50 shown in FIG. 2 and the approximation by the corresponding Gaussian function (approximation function 52). The approximate function 52 is integrated, and the mixture ratio during the transition from the vowel / a / to the vowel / i / is obtained by a curve 54 (shown in the lower part of FIG. 3) normalized so that the final integration value becomes 1. Is obtained.

  The calculation from the allocation of the above Gaussian function to the mixture ratio is summarized as follows. It is assumed that there are m frames between adjacent vowel centers, and the maximum value of the spectrum change rate exists at the nth. At this time, if the frame number between vowels is k (0 ≦ k <n) and the function corresponding to the curve 60 (ie, the mixing ratio) is br (k), the mixing ratio br (k) is expressed by the following equation (1). And (2).

また、ｙ（ｋ）は次の式（３）及び（４）で表わされる。

Y (k) is expressed by the following equations (3) and (4).

ここで、ｇｗ_ｎ（ｋ）は、ｎ点のＧａｕｓｓｉａｎ Ｗｉｎｄｏｗのｋ番目の点である。

Here, gw _n (k) is the kth point of the n-point Gaussian Window.

(2) Calculation of vocal tract cross-sectional area function Next, a method for calculating the vocal tract cross-sectional area function using the mixture ratio br will be described.

  4 and 5 show examples of vocal tract cross-sectional area functions. The vocal tract cross-sectional area function is a function of time and indicates the relationship between the distance from the glottis and the vocal tract cross-sectional area at a certain time. 4 and 5, the vertical axis represents the vocal tract cross-sectional area, and the horizontal axis represents the distance from the glottis. The vocal tract cross-sectional area function shown in FIG. 4 is that at time t1. The vocal tract cross-sectional area function shown in FIG. 5 is that at time t2. In order to obtain the vocal tract cross-sectional area function, it is necessary to obtain the distance from the glottis at each time and the cross-sectional area of the vocal tract at that position. In this embodiment, the vocal tract is divided into a predetermined number of sections, and the vocal tract cross-sectional area function is represented by the vocal tract cross-sectional area in each section at each time. In this embodiment, the number of sections is assumed to be constant, and changes in vocal tract length are handled by changes in section length.

  In the following, the mixture ratio br is considered as a function br (t) at time t. At this time, the distance (xpos) from the glottis in the n-th section counted from the glottis is expressed by the following equation (5).

また声道の断面積（ａｒｅａ）は次の式（６）で表わされる。

The cross-sectional area (area) of the vocal tract is expressed by the following equation (6).

以上の式（１）〜（６）を用いる事により、ある母音から他の母音に移行する間の任意の時点における声道の各セクションにおける断面積を、両母音に対する調音パラメータの間の補間により算出する事ができる。

By using the above formulas (1) to (6), the cross-sectional area in each section of the vocal tract at an arbitrary time point during the transition from one vowel to another vowel is obtained by interpolation between the articulation parameters for both vowels. It can be calculated.

<First Embodiment>
Hereinafter, a speech synthesizer according to an embodiment of the present invention using the above-described interpolation method will be described in detail.

[Constitution]
(1) Speech synthesis system 70
FIG. 6 shows a block diagram of the speech synthesis system 70 according to the first exemplary embodiment of the present invention. In the present embodiment, what a continuous speaker utters is recorded, and the articulation parameters of each vowel and the mixing ratio between the vowels are calculated from the recording. It is assumed that the continuous speech text 84 is given in advance.

  Referring to FIG. 6, the speech synthesis system 70 uses the speech signal and the utterance text 84 to generate data used for interpolation to generate data used to interpolate articulation parameters for speech synthesis. And an interpolating DB 88 for holding data generated by the interpolating data generating device 86, and an audio for outputting a synthesized speech signal 94 using the data in the interpolating DB 88 for the input text 90. And a synthesizer 92.

(2) Interpolation data generator 86
FIG. 7 shows a block diagram of the interpolation data generation device 86. The interpolation data generation device 86 includes a spectrum change rate generation unit 60 for generating the spectrum change rate from the audio signal, a spectrum change rate DB 62 for holding the generated spectrum change rate, a spectrum change rate and an utterance. An interpolation data generation unit 64 for generating articulation parameters and interpolation data for each vowel center from the text 84 is included. Since the vowel-centric articulation parameter can also be considered as interpolation data having an interpolation ratio of 0 or 1, in this embodiment, the vowel-centric articulation parameter is also referred to as interpolation data.

(3) Spectrum change rate generation unit 60
FIG. 8 shows a block diagram of the spectrum change rate generation unit 60. The spectrum change rate generation unit 60 includes a framing unit 100 for framing a given audio signal, a spectrum extraction unit 102 for extracting a spectrum of the framed audio signal, and a cepstrum from the extracted spectrum. A cepstrum calculation unit 104 for calculating the delta cepstrum, a delta cepstrum calculation unit 106 for calculating a delta cepstrum that is a difference in voice feature amount from the calculated cepstrum, a spectrum change rate from the calculated delta cepstrum, A spectrum change rate calculation unit 108 for supplying the calculation result to the spectrum change rate DB 62.

  Here, since the delta cepstrum is a parameter representing a difference in cepstrum, which is an audio feature amount between frames, it can be said that the larger the value of the delta cepstrum, the larger the spectrum change between frames. The delta cepstrum dcep (n, m) is expressed by the following equation (7). Here, m represents the frame number, and n represents the order of the cepstrum.

このデルタケプストラムを使用して、スペクトル変化率ｓ（ｍ）は次の式（８）で表わされる。

Using this delta cepstrum, the spectral change rate s (m) is expressed by the following equation (8).

（４）補間用データ生成部６４
図９に、補間用データ生成部６４のブロック図を示す。補間用データ生成部６４は、スペクトル変化率ＤＢ６２に保持されたスペクトル変化率を参照して、母音中心を示す調音運動極小値を含むフレームを検出するための極小値検出部１１０と、スペクトル変化率ＤＢ６２に保持されたスペクトル変化率を参照して、発話母音間の境界を示す調音運動極大値を含むフレームを検出するための極大値検出部１１２と、発話テキストを比較参照して、調音運動が極小値になるフレームに順に母音を割当てるための母音割当部１１４と、全フレームのうち、第１の音素の母音中心であるスペクトル変化率最小値を含むフレームから始まって、スペクトル変化率最大値を含むフレームを経て、第２の音素の母音中心であるスペクトル最小値を含むフレームの一つ前のフレームで終わる様に複数のフレームを一組のフレームシーケンスとして組合せるためのフレームシーケンス作成部１１６とを含む。

(4) Interpolation data generation unit 64
FIG. 9 shows a block diagram of the interpolation data generation unit 64. The interpolation data generation unit 64 refers to the spectral change rate held in the spectral change rate DB 62, and detects a frame including the articulation minimum value indicating the vowel center, and the spectral change rate. By referring to the spectrum change rate stored in the DB 62 and comparing and referring to the utterance text, the maximal value detecting unit 112 for detecting a frame including the articulatory maximal value indicating the boundary between utterance vowels, A vowel assignment unit 114 for assigning vowels in order to the frame having the minimum value and a spectrum change rate maximum value starting from a frame including the spectrum change rate minimum value that is the vowel center of the first phoneme among all frames. Multiple frames so that they end with the frame preceding the frame containing the minimum spectral value that is the vowel center of the second phoneme. And a frame sequence creation unit 116 for combining a set of frame sequence.

  The interpolation data generation unit 64 further has a minimum value in the first frame in the frame sequence combined by the frame sequence creation unit 116, a maximum value of the Gaussian function in the maximum value frame of the spectrum change rate, and A Gaussian function allocating unit 118 for allocating a Gaussian function having a local minimum value of 0, a local maximum value in the frame of the maximum value of the spectral change rate, and a local minimum value in the first frame of the next frame sequence. A Gaussian function assigning unit 120 for assigning a Gaussian function having a value of 0. The first half (from minimum value to maximum value) of the Gaussian functions assigned by the Gaussian function assignment unit 118 and the latter half (one frame from the maximum value to the minimum value) of the Gaussian functions assigned by the Gaussian function assignment unit 120. Is connected to the local maximum part to obtain a function representing the movement of the speech organ. Hereinafter, this function is referred to as “speech organ change rate function”.

  Further, the interpolation data generation unit 64 integrates this speech organ change rate function from the first frame to the last frame (the frame immediately before the next frame sequence) and normalizes it so that the maximum value becomes 1. An integration / normalization unit 122 for outputting a function (hereinafter referred to as “speech organ shape function”), and a mixture ratio in each frame is calculated from the speech organ shape function output by the integration / normalization unit 122. A mixing ratio calculation unit 124 for holding the sound, an articulation parameter DB 128 for holding pre-calculation parameters for various phonemes, and an articulation parameter corresponding to the phoneme assigned by the vowel assignment unit 114. To calculate the vocal tract cross-sectional area function between each phoneme at each time by interpolation using the corresponding mixing ratio, and the result is used for interpolation. And a vocal tract area function calculation unit 126 for providing the B88. This vocal tract cross-sectional area function is the interpolation data in the present embodiment.

(5) Speech synthesizer 92
FIG. 10 shows a block diagram of the speech synthesizer 92. Referring to FIG. 10, text 90 is attached with time information for uttering each phoneme. The speech synthesizer 92 divides the input text 90 into phonemes, extracts data for interpolation between the two phonemes from the interpolation DB 88 and outputs the data as an interpolated articulation parameter 132 for each two adjacent phonemes. Interpolation parameter extraction unit 130, a clock unit 140 for generating a clock signal of a predetermined period, and an interpolation articulation parameter 132 at a timing determined by the clock from the clock unit 140 in accordance with the sound generation length of the continuous voice to be synthesized. Are sequentially output and provided to the filter 138, a sound source 136 for outputting a sound source signal to be a sound source for speech, and a sound source 136 with characteristics that change in accordance with articulation parameters provided by the output unit 134. And a filter 138 for modulating the sound source signal from and outputting a synthesized speech signal 94.

  Note that when it is necessary to add intonation to the sound or adjust the pitch, it is necessary to change the frequency of the sound source signal from the sound source 136.

[Operation]
There are two aspects to the operation of the speech synthesis system 70 according to the present embodiment. That is, in the first aspect, data for articulation parameter interpolation (articulation parameter after interpolation) is generated from the speech signal and the corresponding utterance text, and an interpolation DB is created (operation of the interpolation data generation device 86). Equivalent to The second aspect is an aspect (corresponding to the operation of the speech synthesizer 92) that synthesizes continuous speech of the input text 90 using the data in the interpolation DB 88. Hereinafter, it demonstrates in order.

(1) Operation of Interpolation Data Generation Device 86 The interpolation data generation device 86 according to the present embodiment operates as follows.

  Referring to FIG. 8, when an audio signal is given, framing section 100 frames the audio signal. The spectrum extraction unit 102 extracts a speech spectrum from each frame of the framed voice signal. The cepstrum calculation unit 104 calculates a cepstrum for each frame based on the speech spectrum extracted from each frame. The delta cepstrum calculation unit 106 calculates a delta cepstrum from the calculated cepstrum according to the above-described equation (7). The spectrum change rate calculation unit 108 calculates the spectrum change rate from the calculated delta cepstrum according to the above-described equation (8). The calculated spectrum change rate is given to the spectrum change rate DB 62 and held.

  Next, referring to FIG. 9, the minimum value detection unit 110 detects a frame having a minimum change rate from the spectrum change rate held in the spectrum change rate DB 62. Similarly, the maximum value detection unit 112 detects a frame having a maximum change rate from the spectrum change rate held in the spectrum change rate DB 62.

  The vowel assignment unit 114 assigns the sound corresponding to the vowel center that is the minimum value by sequentially matching the sounds included in the utterance text 84 to the detected articulation minimum value.

  The frame sequence creation unit 116 creates a frame sequence composed of a pair of frames that give adjacent minimum values among frames in which the spectrum change rate is minimum, and all frames in between. This includes frames in which the spectral change rate is maximized. By this process, a set of one frame sequence is created. By repeating this process, a set of all frame sequences is created.

  The Gaussian function assigning unit 118 assigns a Gaussian function to the spectrum change rate between the first frame and the frame including the maximum value among the combined frame sequences. Similarly, the Gaussian function assignment unit 120 assigns a Gaussian function to the spectrum change rate included in the frame including the maximum value and the frame immediately before the last frame of the frame sequence. The integration / normalization unit 122 integrates the speech organ change rate function obtained by connecting the two Gaussian functions from the beginning to the end of the frame sequence (the frame immediately before the frame that gives the last minimum value), Further, by normalizing so that the maximum value becomes 1, a speech organ shape function is generated.

  The mixing ratio calculation unit 124 calculates the mixing ratio from the utterance organ shape function obtained in this way by the above-described equations (1), (2), (3), and (4). The vocal tract cross-sectional area function calculation unit 126 reads out the articulation parameters corresponding to the phonemes allocated by the vowel allocation unit 114 from the articulation parameters held in the articulation parameter DB 128. Then, using the read articulation parameters, the vocal tract cross-sectional area function between each phoneme in each frame included in the frame sequence is calculated from the calculated mixing ratio by the above-described equations (5) and (6). . The calculated vocal tract cross-sectional area function is given to the interpolation DB 88 and held.

  This process is executed for all frame sequences.

(2) Operation of Speech Synthesizer 92 The speech synthesizer 92 according to the present embodiment operates as follows.

  Referring to FIG. 10, when text 90 is input, interpolation parameter extraction unit 130 divides input text 90 into phonemes. Further, the interpolation parameter extraction unit 130 interpolates a vocal tract cross-sectional area function, which is interpolation data (articulation parameters after interpolation) for interpolating the two phonemes for each pair of adjacent two phonemes in the input text 90. Extract from the DB88. This extraction operation is performed for all combinations of adjacent two phonemes in the input text 90 and output as an interpolated articulation parameter 132.

  The output unit 134 sequentially reads the output interpolated articulation parameters 132, and from the position information between the two phonemes of the parameters, the length of the speech to be synthesized, and the like attached to the interpolated articulation parameters 132, from the clock unit 140. Each interpolation articulation parameter is given to the filter 138 at an appropriate time according to the clock. The filter 138 modulates the sound source signal from the sound source 136 by changing its characteristic according to the given interpolation articulation parameter, and outputs a synthesized speech signal 94. By applying this synthesized voice signal 94 to a speaker via an amplifier (not shown), continuous voice is generated.

  Note that the process of creating a set of frame sequences performed by the frame sequence creation unit 116 (see FIG. 9) may be performed manually.

[Effect of the first embodiment]
In this way, according to the speech synthesis system 70 according to the first embodiment of the present invention, the vocal tract cross-sectional area function that is the articulatory parameter in a manner that matches the movement of the speech organ in the actual human speech. Is interpolated. Therefore, it is possible to synthesize a continuous sound that is smooth and natural in terms of hearing. Further, in the first embodiment, the calculation of the vocal tract cross-sectional area function is performed in advance, and this vocal tract cross-sectional area function need only be read out during actual speech synthesis. As a result, there is an effect that the amount of calculation at the time of actual speech synthesis is reduced. Also, unlike the prior art, interpolation parameters are extracted from the audio signal. As a result, it is possible to reduce labor and cost associated with taking MRI moving images.

[Realization by computer]
The speech synthesis system 70 according to the first embodiment of the present invention can be realized by a computer and a computer program executed on the computer. The control structure of the computer program that implements the speech synthesis system 70 will be described below with reference to FIGS.

(1) Program for Implementing Interpolation Data Generation Unit 64 FIG. 11 shows a detailed flowchart of processing for calculating interpolation parameters. Referring to FIG. 11, when the interpolation parameter calculation process is started, an initial process is first performed at step 150. In other words, the work area is cleared, variables to be used are cleared, and the like. Here, 1 is assigned to a variable i which is a subscript of the array of minimum values. Subsequently, in step 152, a value obtained by adding 1 to the variable i is substituted into the variable j, and the process proceeds to step 154.

  In step 154, it is determined whether or not the jth data is set in the array of local minimum values. If it is set, the process proceeds to step 156; otherwise, the interpolation parameter calculation process is terminated.

  In step 156, the above-described equations (1) to (6) are used between the i-th articulation parameter (referred to as articulation parameter (i)) in the array of local minimum values and the articulation parameter (j). To interpolate. Specifically, first, the frame number (i) and the frame number (j) are referred to by referring to the i-th frame number (referred to as frame number (i)) and the frame number (j) of the array of minimum values. Two types of Gaussian functions are assigned to the spectral change rates of the frames between and. That is, the spectrum change rate between the frame number (i) and the frame number (j), the maximum value of the spectrum change rate therebetween, and the frame number that gives the maximum value are specified, and the frame number (i) A Gaussian function up to the frame giving the maximum value and a Gaussian function from the frame giving the maximum value to the frame number (j) -1 are assigned. These Gaussian functions are connected at the local maximum part to obtain the speech organ change rate function. Based on this speech organ change rate function, the mixing ratio in each frame is calculated using the above-described equations (1) to (4). Furthermore, using the articulation parameter (i), the articulation parameter (j), and the calculated mixture ratio, the vocal tract cross-sectional area function in the frame is calculated by the above-described equations (5) and (6). In this way, the interpolated vocal tract cross-sectional area function is calculated in all frames between the frame number (i) and the frame number (j).

  Subsequently, in step 158, 1 is added to the value of the variable i, and the process returns to step 152 again.

  In this way, the process of calculating the interpolation data is repeated for the frames between all the minimum values.

(2) Program for Implementing Speech Synthesizer 92 FIG. 12 shows a flowchart of a computer program for realizing speech synthesizer 92. Referring to FIG. 12, when the speech synthesis process is started, in step 160, a process for extracting a vocal tract cross-sectional area function, which is an interpolation parameter, from the interpolation DB is performed in accordance with input speech text 90. Subsequently, in step 162, the vocal tract cross-sectional area function extracted in step 160 is output to the filter according to the clock, and output processing for generating a synthesized speech signal is performed.

  FIG. 13 shows a detailed flowchart of the process of extracting the interpolation parameter in step 160. Referring to FIG. 13, when the interpolation parameter extraction process is started, an initial process is first performed at step 170. In other words, the work area is cleared, variables to be used are cleared, and the like. Here, 1 is assigned to a variable i which is a subscript of the phoneme array described later. In step 172, the input text 90 is read. In step 174, the text 90 is divided into phonemes, and these phonemes are sequentially set in the array. Processing proceeds to step 176.

  In step 176, a value obtained by adding 1 to the variable i is substituted into the variable j. In step 178, the i-th phoneme (referred to as phoneme (i)) and phoneme (j) are referenced from the phoneme array set in step 174. At this time, it is determined whether or not a phoneme is set in the phoneme (j) (step 180). If the phoneme (j) has no value (that is, if it is terminated), the interpolation parameter extraction process is terminated. Otherwise, the process proceeds to Step 182.

  In step 182, all interpolation articulation parameters between phonemes (i) and phonemes (j) are extracted from the interpolation DB and stored in the work area in order. Subsequently, at step 184, 1 is added to the variable i, and the process returns to step 176.

  In this way, all the interpolation articulation parameters related to the input text are extracted in order, and are sequentially output and stored in the work area.

  In the output process of step 162 shown in FIG. 12, a synthesized speech signal is generated using the interpolation articulation parameters extracted in the interpolation parameter extraction process of step 160 and accumulated in the work area. Note that the details of this processing are obvious from the description of the configuration and operation of the speech synthesizer 92 described above, and therefore detailed description thereof will not be repeated here.

[Computer hardware configuration]
FIG. 14 shows an example of the external appearance of a computer system that executes the above-described computer program, and FIG. 15 shows an example of a block diagram thereof.

  Referring to FIG. 14, a computer system 330 includes a computer 340 having an FD (flexible disk) drive 352 and a CD-ROM (compact disk read only memory) drive 350, a keyboard 346, a mouse 348, and a monitor 342. And a speaker 372.

  Referring to FIG. 15, in addition to the FD drive 352 and the CD-ROM drive 350, the computer 340 includes a CPU (central processing unit) 356 and a bus 366 connected to the CPU 356, the FD drive 352, and the CD-ROM drive 350. And a read only memory (ROM) 358 for storing a boot-up program and the like, and a bus 366, and a random access memory (RAM) 360 for storing a program command, a system program, work data and the like, and a bus 366. And a sound board 368 connected to the speaker 372. The computer system 330 further includes a printer (not shown).

  Although not shown here, the computer 340 may further include a network adapter board that provides a connection to a local area network (LAN).

  A computer program for causing the computer system 330 to operate as the interpolation data generator 86 shown in FIG. 7 or the speech synthesizer 92 shown in FIG. 10 is a CD-ROM 362 inserted into the CDROM drive 350 or the FD drive 352. Alternatively, it is stored in the FD 364 and further transferred to the hard disk 354. Alternatively, the program may be transmitted to the computer 340 through a network (not shown) and stored in the hard disk 354. The program is loaded into the RAM 360 when executed. The program may be loaded directly into the RAM 360 from the CD-ROM 362, from the FD 364, or via a network.

  This program includes a plurality of instructions that cause the computer 340 to operate as the interpolation data generation device 86 or the speech synthesis device 92 of this embodiment. Some of the basic functions required to perform this method are provided by an OS running on the computer 340 or a third party program, or various toolkit modules installed on the computer 340. Therefore, this program does not necessarily include all functions necessary for realizing the interpolation data generation device 86 or the speech synthesis device 92 of this embodiment. This program is a command that implements the interpolation data generation device 86 or the speech synthesizer 92 by calling an appropriate function or “tool” in a controlled manner so as to obtain a desired result. Only need to be included. The operation of computer system 330 is well known and will not be repeated here.

<Second Embodiment>
In the first embodiment described above, the interpolation DB 88 calculates the vocal tract cross-sectional area function, which is an interpolation parameter, by actually applying the interpolation formula, accumulates it as interpolation data, and uses it to continuously Synthesized speech. However, the present invention is not limited to such an embodiment. For example, the interpolating DB 88 stores not a vocal tract cross-sectional area function but a mixing ratio (calculated by the above-described equations (1) to (4)) for calculating the vocal tract cross-sectional area function. A method of using it for continuous speech synthesis is also conceivable. Hereinafter, a second embodiment to which this method is applied will be described.

[Constitution]
The speech synthesis system according to the second embodiment of the present invention has a configuration similar to the speech synthesis system 70 according to the first embodiment shown in FIG. However, the interpolation data generation apparatus of the system according to the second embodiment has a configuration similar to that of the interpolation data generation apparatus 86 in the first embodiment, but the interpolation data generation unit 64 (FIG. 9). 16), the difference is that an interpolation data generation unit 190 having the configuration shown in FIG. 16 is included. The system according to the second embodiment includes a speech synthesizer 210 having the configuration shown in FIG. 17 in place of the speech synthesizer 92 (see FIG. 10) in the first embodiment.

(1) Interpolation data generation unit 190
FIG. 16 is a block diagram of the interpolation data generation unit 190 according to the present embodiment. Referring to FIG. 16, interpolation data generation unit 190 has the same configuration as interpolation data generation unit 64 according to the first embodiment, but includes vocal tract cross-sectional area function calculation unit 126 shown in FIG. 9. It is different from the interpolation data generation unit 64 in that the output of the mixture ratio calculation unit 124 is not stored and is stored in the interpolation DB 208 as it is. In other words, in the second embodiment, unlike the first embodiment, interpolation of the vocal tract cross-sectional area function, which is an articulatory parameter, is not performed, and only the mixing ratio for each frame is calculated. Is.

(2) Speech synthesizer 210
FIG. 17 shows a block diagram of speech synthesis apparatus 210 according to the present embodiment. Referring to FIG. 17, speech synthesizer 210 has a configuration similar to speech synthesizer 92 according to the first embodiment shown in FIG. The difference is that the speech synthesizer 210 interpolates a vocal tract cross-sectional area function between two phonemes from the interpolating DB 208 for each adjacent phoneme group instead of the interpolation parameter extraction unit 130 shown in FIG. A mixing ratio extraction unit 212 for extracting the mixing ratio of the above, an articulation parameter DB 216 for holding articulation parameters calculated in advance for various phonemes, and an articulation parameter corresponding to the phoneme given by the text 90 It includes a vocal tract cross-sectional area function calculation unit 214 that is extracted from the articulation parameter DB 216, calculates a vocal tract cross-sectional area function between phonemes by interpolation using a corresponding mixing ratio, and outputs the inter-phoneme cross-sectional area function as an interpolated articulation parameter 132. .

[Operation]
In the second embodiment, the spectrum change rate generation unit 60 operates in the same manner as the spectrum change rate generation unit 60 included in the speech synthesis system according to the first embodiment. Therefore, detailed description thereof will not be repeated.

(1) Operation of Interpolation Data Generation Unit 190 The interpolation data generation unit 190 shown in FIG. 16 includes a vocal tract cross-sectional area function calculation unit 126, which is different from the interpolation data generation unit 64 included in the first embodiment. The only difference is whether or not. Therefore, detailed description of the operation will not be repeated. However, since the mixing ratio calculation unit 124 only calculates the mixing ratio as interpolation data here, the interpolation DB 208 holds the mixing ratio instead of the vocal tract cross-sectional area function.

(2) Operation of Speech Synthesizer 210 In speech synthesizer 210, output unit 134, sound source 136, filter 138, and clock unit 140 operate in the same manner as in speech synthesizer 92 according to the first embodiment. To do. Therefore, detailed description thereof will not be repeated.

  Referring to FIG. 17, text 90 includes text composed of phoneme strings to be uttered and time information for uttering each phoneme. When the text 90 is input, this text and time information are given to the mixture ratio extraction unit 212 and the vocal tract cross-sectional area function calculation unit 214.

  The mixing ratio extraction unit 212 extracts all the mixing ratios, which are interpolation data for interpolating the two phonemes, from the interpolation DB 208 for each combination of adjacent two phonemes in the phoneme sequence given by the text 90. . The vocal tract cross-sectional area function calculation unit 214 reads the articulation parameter corresponding to the phoneme given by the text 90 from the articulation parameters held in the articulation parameter DB 216. Then, using the read articulation parameters, a vocal tract cross-sectional area function between the phonemes is calculated from the mixture ratio extracted by Equation (5) and Equation (6). Then, the interpolated articulation parameter 132 is output by interpolating between the two phonemes included in the text 90 with the vocal tract cross-sectional area function. This operation is performed for all combinations of adjacent two phonemes in the phoneme string in the text 90, and is given to the output unit 134 as an interpolated articulation parameter 132. The subsequent operation is the same as that of the speech synthesizer 92.

  Note that the process of creating a frame sequence set performed by the frame sequence creation unit 116 (see FIG. 16) may be performed manually.

[Effect of the second embodiment]
Also in the speech synthesis system according to the present embodiment, the vocal tract cross-sectional area function that is the articulation parameter is interpolated by a method that matches the movement of the speech organ in the actual human speech. Therefore, it is possible to synthesize a continuous sound that is smooth and natural in terms of hearing. Further, the interpolation data generation apparatus calculates only the mixture ratio corresponding to each frame between two phonemes, and does not calculate the vocal tract cross-sectional area function. Therefore, the time required for generating the interpolation data can be shortened, and the capacity required for the interpolation DB 208 can be reduced. Also in the present embodiment, unlike the prior art, interpolation parameters are extracted from the audio signal. As a result, it is possible to reduce labor and cost associated with taking MRI moving images.

[Realization by computer]
Similarly to the first embodiment, the speech synthesis system according to the second embodiment of the present invention can also be realized by a computer and a computer program executed on the computer. Note that the control structure of the computer program that implements the speech synthesis system according to the present embodiment can be easily realized by those skilled in the art based on the description of the first embodiment. The hardware configuration of the computer is the same as that described in the first embodiment. Therefore, detailed description thereof will not be repeated here.

<Third Embodiment>
In the second embodiment described above, the interpolating DB 208 stores the mixing ratio for interpolating the vocal tract cross-sectional area function between two phonemes, and uses this during actual interpolation to use the vocal tract cross-sectional area. The function was calculated and synthesized continuous speech. However, the present invention is not limited to such an embodiment. For example, in the interpolation DB, not the vocal tract cross-sectional area function or the mixing ratio used for the calculation, but a Gaussian function is assigned to the spectral change rate, and the speech organ shape function obtained by integration and normalization is stored. A method of calculating a vocal tract cross-sectional area function, which is an articulation parameter, using these during continuous speech synthesis is also conceivable. Hereinafter, a third embodiment to which this method is applied will be described.

[Constitution]
The speech synthesis system according to the third embodiment of the present invention has a configuration similar to the speech synthesis system 70 according to the first embodiment shown in FIG. However, the interpolation data generation apparatus of the system according to the third embodiment has a configuration similar to that of the interpolation data generation apparatus 86 in the first embodiment, but the interpolation data generation unit 64 (FIG. 9). 18) in that an interpolation data generation unit 230 shown in FIG. 18 is included. The system according to the third embodiment includes a speech synthesizer 250 having the configuration shown in FIG. 19 in place of the speech synthesizer 92 (see FIG. 10) in the first embodiment.

(1) Interpolation data generation unit 230
FIG. 18 is a block diagram of the interpolation data generation unit 230 according to the present embodiment. Referring to FIG. 18, the interpolation data generation unit 230 includes a local minimum value detection unit 110, a local maximum value detection unit 112, and a vowel assignment unit 114, which are the same as the interpolation data generation unit 64 in the first embodiment. A frame sequence creation unit 116, a Gaussian function allocation unit 118, a Gaussian function allocation unit 120, and an integration / normalization unit 124. In the third embodiment, unlike the first embodiment, the calculation of the mixing ratio and the vocal tract cross-sectional area function are not performed, but the integration / normalization processing of the speech organ change rate function is performed. Only the utterance organ shape function is calculated, and this utterance organ shape function is output to the interpolation DB 246 as interpolation data.

(2) Speech synthesizer 250
FIG. 19 shows a block diagram of speech synthesis apparatus 250 according to the present embodiment. Referring to FIG. 19, speech synthesizer 250 has a configuration similar to speech synthesizer 92 according to the first embodiment shown in FIG. The difference is that instead of the interpolation parameter extraction unit 130 shown in FIG. 10, the integrated and normalized speech organ shape function is extracted from the interpolating DB 246 for each set of adjacent phonemes in the input text 90. Interpolation data extraction unit 252, a mixture ratio calculation unit 254 for calculating a mixture ratio from the extracted interpolation data, and an articulation parameter for holding articulation parameters calculated in advance for various phonemes The articulation parameters corresponding to the phonemes given by the DB 216 and the text 90 are extracted from the articulation parameters DB 216, the vocal tract cross-sectional area function between the phonemes is calculated by interpolation using the mixture ratio, and output as the interpolated articulation parameters 132. This includes the same vocal tract cross-sectional area function calculation unit 214 (see FIG. 17) as that used in the second embodiment.

[Operation]
Of the speech synthesis system according to the third embodiment, the spectrum change rate generation unit 60 is the same as the spectrum change rate generation unit 60 included in the speech synthesis systems according to the first embodiment and the second embodiment. To work. Therefore, detailed description thereof will not be repeated.

(1) Operation of Interpolation Data Generation Unit 230 The interpolation data generation unit 230 shown in FIG. 18 is different from the interpolation data generation unit 64 included in the first embodiment in the mixture ratio calculation unit 124 and the vocal tract cross-sectional area. The only difference is whether or not the function calculation unit 126 (see FIG. 9) is included. Therefore, detailed description of the operation will not be repeated. However, since only the speech organ shape function obtained by integrating / normalizing the speech organ change rate function by the integration / normalization unit 124 is calculated as interpolation data here, the interpolating DB 246 can replace the mixing ratio. The utterance organ shape function is held.

(2) Operation of Speech Synthesizer 250 Referring to FIG. 19, text 90 includes text composed of phoneme strings to be uttered and time information for uttering each phoneme. When the text 90 is input, the text and time information are provided to the interpolation data extraction unit 252 and the vocal tract cross-sectional area function calculation unit 214.

  The interpolation data extraction unit 252 refers to the text 90 and extracts a speech organ shape function held in the interpolation DB 246 for each pair of adjacent phonemes in the text. The mixing ratio calculation unit 254 calculates the mixing ratio of the articulation parameters corresponding to the two phonemes from the data calculated by the above formulas (1) to (4), and supplies the calculated mixture ratio to the vocal tract cross-sectional area function calculation unit 214. The subsequent operation is the same as that of the speech synthesizer 210 in the second embodiment.

  Note that the process of creating a frame sequence set performed by the frame sequence creation unit 116 (see FIG. 18) may be performed manually.

[Effect of the third embodiment]
Also in the speech synthesis system according to the present embodiment, the vocal tract cross-sectional area function that is the articulation parameter is interpolated by a method that matches the movement of the speech organ in the actual human speech. Therefore, it is possible to synthesize a continuous voice that is smoother and more natural for hearing. Interpolation data generation section 230 only calculates the speech organ shape function by integrating and normalizing the speech organ change rate function obtained by approximating the spectrum change rate with a Gaussian function, and the mixture ratio or vocal tract cross-sectional area is calculated. The function is not calculated. Therefore, the time required for generating the interpolation data can be shortened, and the capacity required for the interpolation DB 246 can be reduced. However, since both the calculation of the mixing ratio and the calculation of the vocal tract cross-sectional area function are performed in the speech synthesizer 250, the amount of calculation is larger than in the first embodiment or the second embodiment. Also in the present embodiment, unlike the prior art, interpolation parameters are extracted from the audio signal. As a result, it is possible to reduce labor and cost associated with taking MRI moving images.

[Realization by computer]
Similarly to the first embodiment, the speech synthesis system according to the third embodiment of the present invention can be realized by a computer and a computer program executed on the computer. Note that the control structure of the computer program that implements the speech synthesis system according to the present embodiment can be easily realized by those skilled in the art based on the description of the first embodiment. The hardware configuration of the computer is the same as that described in the first embodiment. Therefore, detailed description thereof will not be repeated here.

  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 claim in 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 graph which shows an example of an utterance voice waveform, and an example of a spectrum change rate. It is a graph which shows an example of the spectrum change rate for mixture ratio calculation. It is a graph which shows an example of the Gaussian function by which the spectrum change rate 50, the normal distribution 52 by the Gaussian function allocation, and the integration process for calculation of a mixture ratio were carried out. It is a figure which shows the example of a vocal tract cross-sectional area function. It is a figure which shows the example of a vocal tract cross-sectional area function. 1 is a block diagram of a speech synthesis system 70 according to a first embodiment of the present invention. 3 is a block diagram of an interpolation data generation device 86. FIG. 4 is a block diagram of a spectrum change rate generation unit 60. FIG. It is a block diagram of the data generation part for interpolation 64 concerning a 1st embodiment of the present invention. It is a block diagram of the speech synthesizer 92 which concerns on the 1st Embodiment of this invention. It is a detailed flowchart of the process which calculates an interpolation parameter. 3 is a flowchart of a computer program that implements a speech synthesizer 92. It is a detailed flowchart of the process which extracts the interpolation parameter of step 160. FIG. It is a figure which shows an example of the external appearance of a computer system. It is a block diagram of the computer system shown in FIG. It is a block diagram of the data generation part 190 for interpolation which concerns on the 2nd Embodiment of this invention. It is a block diagram of the speech synthesizer 210 which concerns on the 2nd Embodiment of this invention. It is a block diagram of the data generation part 230 for interpolation which concerns on the 3rd Embodiment of this invention. It is a block diagram of the speech synthesizer 250 which concerns on the 3rd Embodiment of this invention.

Explanation of symbols

60 Spectral change rate generation unit 64 Interpolation data generation unit 100 Framing unit 108 Spectral change rate calculation unit 110 Minimum value detection unit 112 Maximum value detection unit 116 Frame sequence creation unit 118 Gaussian function allocation unit 120 Gaussian function allocation unit 122 Normalizer 124 Mixing ratio calculator 126 Vocal tract area function calculator

Claims (5)

Articulation parameters for synthesizing speech that varies continuously from the first phoneme to the second phoneme are generated by interpolation between known articulation parameters for speech synthesis of the first and second phonemes. A data generation device for articulation parameter interpolation for generating interpolating data at the time,
Change rate calculation means for calculating a transition of a predetermined spectrum change rate from the first phoneme to the second phoneme from an input speech signal including the first phoneme and the second phoneme that are continuous; ,
Look including an interpolation data generating means for generating said interpolation data by using the rate of change,
The rate of change calculation means includes:
Framing means for framing the input audio signal every predetermined time;
Spectrum change rate calculating means for calculating the predetermined spectrum change rate for each frame from the voice signal framed by the framing means,
The spectrum change rate calculating means includes:
First and second for giving adjacent minimum values of the predetermined spectrum change rate in the frames of the voice signal framed by the framing means from the voice signal from the first phoneme to the second phoneme. Frame combining means for combining a plurality of frames including two frames and a third frame that exists between the two frames and that gives a maximum value of the predetermined spectral change rate;
Function approximating means for approximating the transition of the predetermined spectral change rate with a Gaussian function defined by the maximum value and the minimum value of the frames combined by the frame combination means;
The interpolation data creation means includes:
A normalized integral function obtained by integrating the Gaussian function obtained by the function approximating means from the first frame to the second frame to obtain an integral function, and further normalizing the integral function with the maximum value. A data generation apparatus for articulation parameter interpolation , including integration / normalization means for obtaining The function approximating means is:
  Means for obtaining a first Gaussian function approximating the predetermined spectral change rate between the first frame and the third frame among the frames combined by the frame combination means;
  Means for obtaining a second Gaussian function approximating the predetermined spectral change rate between the third frame and the second frame among frames combined by the frame combination means;
  By connecting the first Gaussian function and the second Gaussian function in the third frame, the transition of the predetermined spectral change rate is approximated between the maximum value and two minimum values. The long sound parameter interpolation data generation apparatus according to claim 1, further comprising means for obtaining a Gaussian function. The interpolation data creation means further interpolates between the articulation parameter for the first phoneme and the articulation parameter for the second phoneme by the normalized integration function, thereby the first phoneme. The articulation parameter interpolation data generation device according to claim 1 or 2 , further comprising means for calculating a mixture ratio between a phoneme and the second phoneme at a predetermined time. The interpolation data creating means further interpolates between the articulation parameter for the first phoneme and the articulation parameter for the second phoneme by the mixing ratio, thereby providing articulation at the predetermined time. 4. The articulation parameter interpolation data generation device according to claim 3 , further comprising means for calculating a vocal tract cross-sectional area function as a parameter. A computer program that, when executed by a computer, causes the computer to operate as the articulation parameter interpolation data generation device according to any one of claims 1 to 4 .

Images