JP4617092B2 - Chinese tone classification device and Chinese F0 generator - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To make it easier to determine tones in Chinese speech. <P>SOLUTION: A Chinese language tone classification apparatus 110 includes: a tone model 106 trained to output probabilities of tone classification. The anchor based tone discriminating features include onset gap and offset gap, and preferably, an F0 slope coefficient. The apparatus 110 further includes: phonetic segmentation module 150 for segmenting a Chinese speech 108 into a series of tone nuclei; feature extraction module 152 and anchoring feature extraction module 154 for extracting anchor based tone discriminating features; and a pattern matching module 156 that outputs the tone classification that achieves the highest probability. <P>COPYRIGHT: (C)2005,JPO&amp;NCIPI

Description

  The present invention relates to tone classification of tonal languages, and more particularly to identifying changes in tone with context very accurately in tone languages such as Chinese.

  Chinese basic vocabulary tones (called first voice, second voice, third voice, fourth voice) are usually characterized by their different fundamental frequency ("F0") contour patterns. That is, the first voice has a high level contour, the second voice has a contour that rises from the middle, the third voice has a low contour that falls once, and the fourth voice has a contour that descends from a high location. In order to distinguish these, researchers have distinguished the first voice (high F0) and the third voice (low F0) in previous studies (see Non-Patent Document 1), and the third voice ( It is reported that the height of F0 is decisively important for the distinction between low F0) and the fourth voice (high F0) (see Non-Patent Document 2).

  The F0 gradient is effective for distinguishing between different tones such as the first voice (flat), the second voice (up), and the fourth voice (down) (see Non-Patent Document 2). Please refer.)

  The tone F0 contour varies considerably in continuous speech compared to syllables separated from each other. Since not only the height of F0 but also the gradient of the slope fluctuates considerably, it is not possible to specify the tone of the underlying tone from the F0 contour on the surface. On the other hand, perceptual experiments show that humans can perceive vocabulary tones that appear to be below them with high consistency even if F0 fluctuates considerably, given tonal context. (See Non-Patent Document 3). This indicates that there are clues for distinction in addition to the height of F0 and the slope of F0 in the tonal context.

D. Wallen and Y.W. Shu, “Amplitude contour and Mandarin tone information in short segments”, Phonetica 49, 1992, pp. 25-47. (D. Whalen and Y. Xu, "Information for Mandarin tones in the amplitude contour and in brief segments", Phonetica 49, 1992, pp. 25-47.) E. Guarding et al., “Distinction between Modern Mandarin Fourth and Third Voices” Language and Speech, Vol. 29, Part 3, 1986, pp.281-293. (E. Garding et al., "Tone 4 and tone 3 discrimination in modern standard Chinese", Language and Speech, Vol. 29, Part 3, 1986, pp. 281-293.) Y. Shu, “Generation and perception of simultaneous tone,” JASA (4), 1994, pp. 2240-2253. (Y. Xu, "Production and perception of coarticulated tones", J.A.S.A. (4), 1994, pp. 2240-2253.) Y.R. Chao, “Spoken Chinese Grammar” Berkeley, University of California Press, 1968. (Y.-R. Chao, "A grammar of spoken Chinese", Berkeley: University of California Press, 1968.) T.A. Tsumura et al., "Hearing detection of frequency transitions" JASA Vol. 53, No. 1, 1973, pp. 17-25. (T. Tsumura et al., "Auditory detection of frequency transition", J.A.S.A., Vol. 53, No. 1, 1973, pp. 17-25.) S. Shigeno, H.C. Fujisaki, “Effect of preceding anchor on category judgment of voice and non-voice stimuli”, Japanese Psychological Research, 1979, Vol. 21, No. 4, pp.165-173. (S. Shigeno, H. Fujisaki, "Effect of a preceding anchor upon the categorical judgment of speech and non-speech stimuli", Japanese Psychological Research, 1979, Vol. 21, No. 4, pp. 165-173.) J. et al. S. Chang, K. Hirose, “The Association Hypothesis and Its Application to Tone Recognition of Chinese Continuous Speech” ICASSP2000, Volume III, pp. 1419-1422. (J.-S. Zhang, K. Hirose, "Anchoring hypothesis and its application to tone recognition of Chinese continuous speech", ICASSP2000, Vol. III, pp. 1419-1422.) J. et al. S. Chang, S. Nakamura, K. Hirose, “Distinction of Chinese Vocabulary Tone by Association F0 Feature” ICSLP2000 Proceedings, Volume II, pp. 87-90. (J.-S. Zhang, S. Nakamura, and K. Hirose, "Discriminating Chinese lexical tones by anchoring F0 Features", Proc. Of ICSLP2000, Vol. II, pp. 87-90.)

  To date, none of the studies have clearly shown clues for discrimination in tonal contexts other than the height of F0 and the slope of F0. If such cues are available, it should be a great help to classify tones in tonal languages such as Chinese, and if there is an appropriate model trained with such cues, a speech recognition or speech synthesizer The performance should be improved further.

  Accordingly, one of the objects of the present invention is to facilitate the determination of the Chinese tone.

  Another object is to synthesize more natural quality Chinese speech.

The Chinese tone classification device according to the present invention provides training data so as to output a probability of tone classification used in Chinese when a set of feature amounts including a feature amount for identifying a tone is given by association. Means for storing a tone model trained with the set is included. Feature quantity identifying the tone by association includes the head gap and tail gap. Tone classifier further includes means for segmenting the audio data is input Chinese into a series of tone nuclei, means for extracting a feature value identifying the tone by association from each tone nuclear extract Pattern matching means for applying the acoustic feature extracted by the means to the tone model and selecting a tone classification that achieves the highest probability output by the tone model.

Preferably, the feature to identify the tone by association further comprises an inclined slope of the fundamental frequency (F0) contours of tone nucleus.

More preferably, the feature to identify the tone by associating further comprises a head F0 and end F0 tones nucleus.

More preferably, the feature to identify the tone by associating further comprises the normalized power of the tone nucleus.

The Chinese tone classification device according to another aspect of the present invention outputs a probability of tone classification used in Chinese when a set of feature amounts including a feature amount for identifying a tone is given by association. A means for storing a tone model trained using the training data set, a means for segmenting input Chinese speech data into a series of tone nuclei, and an association from each of the tone nuclei and means for extracting a feature value to identify the tone. Feature quantity identifying the tone by association includes the head gap and tail gap. The apparatus further includes a tone classification means for determining a tone classification according to the combination of the leading gap and trailing gap codes extracted by the extracting means and outputting the determined classification.

  Preferably, the tone classification means includes means for determining that the tone is the first Chinese voice when the tone gap is positive and the tail gap is positive.

  The tone classification means may further include means for determining that the tone is a Chinese second voice when the tone head gap is negative and the tail gap is positive.

  Preferably, the tone classification means further includes means for determining that the tone is a third Chinese voice when the tone gap is negative and the tail gap is negative.

  More preferably, the tone classification means further includes means for determining that the tone is the fourth voice of Chinese when the leading gap of the tone is positive and the trailing gap is negative.

The Chinese F0 generation device according to still another aspect of the present invention outputs a probability of each tone classification used in Chinese when a set of feature amounts including a feature amount for identifying a tone is given by association. Means for storing a tone model trained using the training data set. Feature quantity identifying the tone by association includes the head gap and tail gap. The apparatus further includes means for storing a probabilistic F0 model that, given a parsed Chinese text, outputs the probabilities of possible Chinese tones for each of the speech units in the Chinese text. The means for generating a sequence of F0 that matches the input Chinese text according to the output of the stochastic F0 model, and whether the F0 sequence output by the means for generating is consistent with the tone model Means for determining whether or not.

[Prerequisites]
(F0 change phenomenon due to tonal context)
The contextual changes considered by the inventors here are two well-known phenomena: “step-down” and “contextual assimilation”. Strictly speaking, the staircase descent phenomenon is a kind of contextual assimilation in the specific context where a high pitch target follows a low pitch target.

-Effect of stair-step descent The effect of stair-step descent is that the height of F0 of the second “H” tone is the first “H” because “L” exists in the tone sequence of “HLH”. This is known as the phenomenon of lowering. If H and L act alternately in a sequence where H and L appear alternately, the contour of F0 will be a stepped function. The F0 height of the H tone at a later position in a certain utterance may be lower than the L tone in front of the same utterance.

  Three of the four basic vocabulary tones in Chinese are low in either the first (second voice), last (fourth voice), or both the first and last (third voice) as shown in FIG. Have a target.

  FIG. 2 shows an example of a step-like descent phenomenon. The spoken text is "you3 qing1 wei1 e4 hua4 xian4 xiang4" ("There is a phenomenon that gets worse little by little."). Two interesting phenomena can be seen from the F0 sequence indicated by the dotted line 20 in FIG.

  1. The effect of the staircase-like descent occurs three times in succession in the third half of the fourth voice, making it a downward staircase.

  2. Furthermore, the head of H of the last fourth voice is at the same level as the F0 height of the third voice at the beginning of the utterance.

  It is well known that the height of the vocabulary tone F0 is significantly affected by a step-down, but does not interfere with human pitch perception. That is, the plurality of fourth voices shown in FIG. 2 are not only perceived as having the same tone, but are also perceived as having substantially the same height at the beginning. Furthermore, in speech synthesis research, if the same F0 value is assigned to two adjacent tones that are likely to be affected by a staircase descent at a high pitch point, unnatural stress is applied to the second lexical tone in synthesized speech. It feels like there is.

-Tonal change by contextual assimilation The concept of tonal change by contextual assimilation used here is that the influence of the change by assimilation is so severe that the direction of the F0 slope of the tone changes to that of another tone. It refers to the phenomenon of being. FIG. 3 shows an example in which the spoken text is “si1 zhu2 guan3 xian2” (traditional stringed and woodwind instruments). In the example shown in FIG. 3, the F0 contour of the portion surrounded by the line 30 is the second voice, and the characteristic pattern is an ascending gradient. Here, however, the contour of F0 has changed to a flat one, which is that of the first voice.

  One important issue is whether the tone has changed to another tone when the F0 contour changes. The problem was so confusing that it used to show a very different view. In Non-Patent Document 4, to explain the case as shown in FIG. In that proposal, if the second voice is preceded by a higher tone, followed by one of the four basic tones, the second voice sandwiched in between is a normal speed voice. It is suggested that it will change to the first voice.

  However, recent research has revealed that as long as there is a tone context following F0 changes due to context, the listener still perceives that tone as that of the original tone (Non-Patent Document 3). This means that the second voice that has undergone the change in F0 in FIG. 3 is still perceived as the second voice rather than the first voice. This has been confirmed by some Chinese-speaking speakers.

-Two findings on perception experiments Non-Patent Document 3 conducts a more systematic investigation on F0 changes due to tone context and its effect on human pitch perception. One of the interesting experiments in Non-Patent Document 3 is how humans perceive a tone that is sandwiched in a sequence of three syllables. It is examined whether it is perceived similarly in the context where the syllable and the third syllable are exchanged. The exchange of the first and third syllables provides a special way to test the effect of tone context on pitch perception, as the tone environment is quite different for the tones caught between them Expected.

  FIG. 4 illustrates two findings obtained from the experiment of Non-Patent Document 4.

  1. When the original tone context is natural and the context after the exchange is unnatural, the tone between them tends to be perceived as the same tone as the original context (see the first line in FIG. 4).

  2. If the original tone context is unnatural and the post-exchange context is natural, the tone between the two is perceived as a tone with a contour in the opposite direction to the original tone in the exchanged context There is a trend (see the bottom row in FIG. 4).

(F0 feature quantity by association hypotheses and associated)
For the purpose of finding a more effective feature amount for distinguishing lexical tone, the present inventors adopt the knowledge on psychological acoustic perception of Non-Patent Document 5 or Non-Patent Document 6, and It made a such association hypothesis as (see non-Patent Document 7 and non-Patent Document 8).

  -The relative F0 difference between the end of the first lexical tone and the beginning of the second lexical tone is to identify high or low pitches, as well as a direct clue to the down slope with respect to the F0 contour. It can be an important clue.

  -There must be a timing allocation mechanism for competitive effects (see Non-Patent Document 6).

Based on this hypothesis, the vocabulary tone of the continuous speech uses the association pattern shown by the table shown in FIG. 5 in addition to using the F0 pattern of flat, up, down, or down. Can be acoustically characterized. The terms used in FIG. 5 are as follows.

-Leading gap: The difference between the leading F0 and the trailing F0 of the tone of the preceding vocabulary .
-End gap: difference between the end F0 and the beginning F0 of the tone of the following vocabulary.

(Identification by association of change by tone context)
The inventors of the present invention have found that it is possible to consistently predict a tone that changes the tone described above by using this association hypothesis.

-Identification by associating voices that fall in a staircase pattern FIG. 6 illustrates a method for estimating a leading gap and a trailing gap. In particular, the estimation of the leading and trailing gaps of the tone that goes down in a staircase pattern of FIG. 2 is illustrated.

  In FIG. 6, thin vertical lines indicate syllable boundaries. The feature quantity “r” represents the leading and trailing gaps, and the feature quantity “d” represents the F0 slope with the duration normalized. The leading and trailing points are points corresponding to the tone kernel.

The fourth voice has a pitch pattern of “HL”. According to the tone discrimination hypothesis by association , when the following conditions are satisfied, it is determined as the fourth voice.

  1. The leading gap is positive. That is, r ≧ 0. Since the fourth voice starts with H, if the end of the preceding tone is L, r> 0. On the other hand, if the end of the preceding tone is H, r≈0.

  2. By the same mechanism as above, the tail gap is negative.

  3. The slope of the F0 slope is negative, that is, d <0 (down slope pitch).

From FIG. 6, it can be seen that the four fourth voices are hardly different in that the above-mentioned conditions are satisfied, although their absolute F0 heights differ greatly. Every fourth of the four fourth voices has a similar level of positive leading gap, a similar level of negative trailing gap, and a similar F0 contour downslope. These three features are very important to distinguish the tone according to the pitch association hypothesis, so these four fourth voices should be perceived as similar tones, which was demonstrated by listening .

Distinguishing by association of tones that have undergone contextual assimilation FIG. 7 illustrates the estimation of leading and trailing gaps r and F0 slope gradient d for the second voice. The following can be seen from FIG.

  1. The leading gap is rather negative, i.e. r1 <0.

  2. The tail gap is rather positive, i.e. r2> 0.

  3. The F0 contour slope is flat, ie d≈0.

Since it can be seen from FIG. 5 that the second voice is most applicable to the above-described feature amount, it is predicted that this tone will be the second voice. It is the negative leading gap and the positive trailing gap that determine the prediction of the second voice based on the association hypothesis.

(Prediction of tone in exchanged context)
Based on the associated tone discrimination hypothesis, it is also possible to easily and accurately predict what tone is perceived in experiments in the above-described exchanged context. 8 and 9 show this prediction procedure and the prediction results for the tones sandwiched in the exchanged context. The target is first predicted whether it is a high (+) or low (-) target based on the leading gap, trailing gap, and downhill pitch, and then based on the two targets at the beginning and end Which tone is predictable.

  From FIGS. 8 and 9, it can be seen that the predicted results are consistent with the two findings reported.

  The Chinese tone distinction pattern indicates a pattern that is an acoustic feature amount of tone in speech. A good pattern enables consistent and reliable tone identification from the acoustic features of the actual speech, regardless of whether the utterance consists of separated syllables or continuous natural speech. Should be.

The tone identification pattern by association is the first identification pattern proposed for the Chinese tone in continuous speech. Compared to traditional Chinese tone patterns, this new pattern provides the ability to better identify tone in continuous speech. This proposal should have many applications. The following is a brief description of the application of this to Chinese automatic tone classification and Chinese speech synthesis.

[First embodiment]
FIG. 10 illustrates a tone classification system 100 according to the first embodiment of the present invention. The system 100 is based on the tone discrimination feature quantity based on the association described above.

Referring to FIG. 10, the system 100 performs tone model training for training the tone model 106 using the training data 102 so as to output the tone classification with the highest probability given a feature value of a set of associations. A unit 104 and a tone classification unit 110 that outputs a tone classification 112 using the tone model 106 trained by the tone model training unit 104 in response to the input speech 108.

  The tone model 106 includes a number of Gaussian mixture models (GMMs) prepared for each tone. If only context-independent GMMs are used, the total number of GMMs is 5. That is, four for basic four voices and one for light voices. If a context-dependent GMM is used, a maximum of 175 GMMs can be used (5 * 5 * 5 = 125 for three tone contexts, 4 * 5 for left boundary tone, right boundary 5 * 5 for tone and 5 for isolated tone).

The tone model training unit 104 includes an acoustic segmentation module 130 for acoustically segmenting the tone kernels of each utterance of the training data 102, and a tone kernel of the speech signal segmented by the acoustic segmentation module 130. A feature extraction module 132 for extracting an acoustic feature quantity including F0 and power, and a feature quantity by associating a head gap, a tail gap, an F0 contour gradient, etc. from the acoustic feature quantity extracted by the feature extraction module 132 a feature extraction module 134 by associating to extract, by using a feature amount associated extracted by the feature extraction module 134 by associating, and a tone model estimation module 136 for estimating the model parameters of the tone models 106 .

  The acoustic segmentation of tonal nuclei in the acoustic segmentation module 130 can be achieved by statistical methods or simply using the speech segmentation of the speech recognizer.

The tone classification unit 110 segments the input speech 108 and outputs an acoustic segmentation module 150 for outputting the tone kernel of the segmented speech signal, and the speech signal segmented by the acoustic segmentation module 150. feature extraction and feature extraction module 152 for extracting an acoustic feature quantity including F0 and power of tone nucleus, by association for extracting a feature amount by the association from the acoustic feature amount extracted by the feature extraction module 152 Searching for a feature pattern that best matches the feature quantity by the association extracted by the module 154 and the feature extraction module 154 by the association among the tone models 106, and outputting the tone classification 112 corresponding to the best matching pattern Pattern matching And an Lumpur 156.

  System 100 operates as follows. This system has two operating phases. A training aspect and an operational aspect.

  In the training aspect, utterances in training data 102 are acoustically segmented at tonal boundaries by acoustic segmentation module 130. The tone kernel of the segmented speech signal is provided to the feature extraction module 132, and the acoustic features of the segmented speech signal including F0 and power are extracted.

The feature extraction module 134 by association extracts a feature amount for tone discrimination based on association from the acoustic feature amount of the tone nucleus. Identifying feature according to the two most important association is the top gap and bottom gap, these by 11, is calculated as r _{1, 1} and r _{2, 2.} Another important feature amount is the slope of the F0 contour, the head F0, and the tail F0. If desired, the normalized power of tone kernels may be included. In this case, the feature quantity of one tone is a five-dimensional or six-dimensional vector.

  Through the tone model training process, the parameters of the Gaussian mixture model in the tone model 106 for each tone are estimated from the training data 102.

In the next tone recognition / classification phase, the input speech 108 is segmented by the acoustic segmentation module 150, and the acoustic feature quantity is extracted from the segmented speech signal by the feature extraction module 152. Further, from the feature quantity extracted by the feature extraction module 152, a feature quantity by association is extracted by the feature extraction module 154 by association .

The pattern of the feature amount by the association is compared with the tone model. The tone sequence with the highest probability is output as the recognized tone sequence (tone classification 112).

[Second Embodiment]
The pattern by association is not only used directly for generating the F0 pattern from the text as in the conventional rule-based synthesis system, but also integrated with the stochastic F0 generation system to improve the tonal intelligibility of the synthesized speech. It can also be improved.

In a rule-based intonation generation system, a text intonation F0 pattern to be synthesized is predicted using a number of rules developed from a training corpus. The associated tone identification pattern can be integrated with this set of rules to easily obtain a clue from the generated F0 that is sufficient to identify the tone. In other words, as compared with the standard tone pattern, the association pattern has a high degree of freedom with respect to the arrangement of the locus of F0. This should improve the naturalness of the synthesized speech.

  In the current probabilistic F0 generation system, a training period is provided prior to the synthesis process. What is characteristic of this training process is that there are too many factors related to the context and sufficient training data cannot be obtained to build a robust intonation model.

  Taking Chinese as an example, there are about 1300 tonal syllables, with more than 2.2 billion possible three syllable connections. Considering additional factors such as emphasis and position in the prosodic phrase (start point, middle point, end point), question or plain sentence, etc., possible combinations can be in trillions of units. It is impossible to collect enough data to train a model for such a large number of contextual combinations. Therefore, the generated intonation F0 often sounds strange due to serious data sparseness problems.

However, by using a tone distinction pattern based on association as an auxiliary, a post-check module for examining a stochastically synthesized F0 contour can be easily used.

  FIG. 12 shows an F0 generation unit 202 according to the second embodiment of the present invention for generating an F0 sequence 204 from input text 200. The F0 generation unit 202 is used together with the probabilistic F0 model 206 used in the prior art F0 generator and the tone model 208 used in the first embodiment of the present invention.

  The F0 generation unit 202 generates a text parsing module 220 for parsing the input text 200 and a text intonation F0 parsed by the text parsing module 220 using the probabilistic F0 model 206. Check the F0 sequence output by the intonation F0 generation module 222 and the tone model 208 and check the F0 sequence output by the intonation F0 generation module 222, correct the F0 that does not match the pattern, and output the corrected F0 sequence 204 Module 224.

As described above, the feature pattern by association makes it possible to reliably identify the tone when correcting the synthesized F0 contour according to the style of speaking at a high level such as a change in speaking speed and a change in emotion. This problem can hardly be realized with the prior art Chinese TTS (text-to-speech synthesis) system.

  The above-described embodiments are merely examples and should not be construed as limiting. 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 figure which shows a Chinese tone pattern. It is a figure which shows the example of the effect which descends in the step shape observed by the 4th continuous 4th voice. It is a figure which shows the example of assimilation on a context. It is a figure which shows the influence of context exchange with respect to the tone perception of the vocabulary tone pinched | interposed. It is a figure which shows the feature-value by the correlation of the four basic vocabulary tone in a continuous sound. It is a figure which shows the estimation of a head gap and a tail gap about the tone which falls in the step shape of FIG. It is a figure which shows the estimation of a head gap and a tail gap about the 2nd tone of FIG. It is a figure which shows prediction of a tone change in the perception experiment of an exchange context. It is a figure which shows prediction of a tone change in the perception experiment of an exchange context. 1 is a block diagram showing a tone classification system 100 according to a first embodiment of the present invention. It is a figure which shows extraction of the tone distinction feature-value by correlation . FIG. 7 is a diagram showing an F0 generation unit 202 according to the second embodiment of the present invention for generating an F0 sequence checked and corrected by a tone pattern by associating a tone model.

100 tone classification system, 102 training data, 104 tone model training unit, 106 tone model, 108 input speech, 110 tone classification unit, 112 tone classification, 130, 150 acoustic segmentation module, 132, 152 feature extraction module, 134, 154 association feature extraction module, 136 tone model estimation module, 156 pattern matching module, 200 input text, 202 F0 generation unit, 204 F0 sequence, 206 stochastic F0 model, 208 tone model, 220 text parsing module, 222 intonation F0 Generation module, 224 post-check module

Claims (9)

A training data set that outputs a probability of tone classification used in Chinese given a set of features including features that identify the tone by the fundamental frequency (F0) at the tone kernel boundary of the Chinese speech. A tonal classifier for Chinese, including means for storing a tonal model trained using, wherein the features identifying the tone by F0 at the tone kernel boundary are a head gap, a tail gap, and a tone kernel A combination of signs of the slope of the F0 contour of
The tone classification device further includes:
Means for segmenting input Chinese speech data into a series of tone nuclei;
Means for extracting from each of the tone nuclei a feature that identifies the tone by F0 at the tone nucleus boundary ;
Pattern matching means for applying to the tone model the acoustic feature extracted by the means for extracting and selecting a tone classification that achieves the highest probability output by the tone model; Tone classification device. 2. The Chinese tone classification apparatus according to claim 1, wherein the feature amount for identifying a tone by F0 at a tone nucleus boundary further includes a head F0 and a tail F0 of the tone nucleus. 3. The Chinese tone classification device according to claim 1, wherein the feature quantity for identifying a tone by F0 at a tone nucleus boundary further includes a normalized power of the tone nucleus. 4. A tone model trained using a training data set to output the probability of tone classification used in Chinese given a set of features that include a feature that identifies the tone by F0 at the tone kernel boundary Means for storing
Means for segmenting input Chinese speech data into a series of tone nuclei;
Means for extracting a feature quantity for identifying a tone by F0 at a tone kernel boundary from each of the tone nuclei, wherein the feature quantity identifies a tone by F0 at a tone kernel boundary. Includes a combination of a leading gap, a trailing gap, and a sign of the slope gradient of the F0 contour of the tone kernel,
The tone classification device further includes:
The tone for judging the tone classification according to the combination of the leading gap, the trailing gap extracted by the means for extracting, and the sign of the slope gradient of the F0 contour of the tone nucleus, and outputting the judged category Chinese tone classification device including classification means. 5. The Chinese character according to claim 4, wherein the tone classification means includes means for determining that the tone is the first Chinese voice when the tone head gap is positive and the tail gap is positive. Tone classification device. The Chinese tone classification means according to claim 5, wherein the tone classification means further includes means for determining that the tone is a second Chinese voice when the tone start gap is negative and the tail gap is positive. Tonal classification device for words. The Chinese tone classification unit according to claim 6, wherein the tone classification means further includes means for determining that the tone is a third Chinese voice when the leading gap of the tone is negative and the trailing gap is negative. Tonal classification device for words. The Chinese tone classification unit according to claim 7, wherein the tone classification means further includes means for determining that the tone is the fourth voice of Chinese when the leading gap of the tone is positive and the trailing gap is negative. Tonal classification device for words. A Chinese F0 generator,
A tone model trained using a training data set to output the probability of tone classification used in Chinese given a set of features that include a feature that identifies the tone by F0 at the tone kernel boundary A Chinese F0 generation device including means for storing the feature, wherein the feature quantity for identifying a tone by association is a combination of a leading gap, a trailing gap, and a sign of an inclination gradient of a tone kernel F0 contour. Including
The F0 generator further includes
Means for storing a probabilistic F0 model that, given a parsed Chinese text, outputs the probabilities of possible Chinese tones for each of the speech units in the Chinese text;
Means for generating a sequence of F0 matching the input Chinese text according to the output of the probabilistic F0 model;
Means for generating Chinese F0, comprising: means for determining whether the F0 sequence output by the means for generating is consistent with the tone model;

Images