JP6756607B2 - Accent type judgment device and program - Google Patents

Description

The present invention relates to a device and a program for determining a voice accent type based on a fundamental frequency extracted from voice waveform data.

Conventionally, there is known a text-to-speech synthesis technique for synthesizing speech for arbitrary text using statistical information. In order to realize text-to-speech synthesis, it is necessary to prepare a large amount of information such as voice waveform data, text of the utterance content, reading, accent, etc. as a preliminary preparation.

In order to make advance preparations for correctly assigning readings and accents to a large amount of voice waveform data, work by an announcer or other specialist who can distinguish accents is required. However, there is a problem that it is costly to perform such advance preparation.

In order to solve this problem, a method for automatically determining the accent type for voice waveform data has been proposed (see, for example, Non-Patent Document 1).

The method described in Non-Patent Document 1 is a pitch representing each mora in the accent phrase constituting the voice waveform data in units of the mora (consonant + vowel, vowel, sokuon or syllabary) constituting the accent phrase. Ask for. Then, a feature vector of (mora number -1) dimension is constructed by the difference value of the pitches of adjacent mora, and the feature vector of the accent phrase having the same accent type is used for each accent type (mora number -1). Learn the model by finding the normal distribution of dimensions.

Then, this method calculates a (mora number-1) -dimensional feature vector for the accent phrase to be determined, and uses the learned model to determine the accent type corresponding to the feature vector of the accent phrase to be determined.

Further, as a method for determining the type of accent of voice waveform data, a method for supporting text-speech synthesis has also been proposed (see, for example, Patent Document 1).

In the method described in Patent Document 1, the voice waveform data corresponding to a predetermined phrase is divided into a plurality of accent phrases, and the accent phrase to be determined is changed with time in pitch with the mora constituting the accent phrase as a unit. Calculate the slope and final pitch of. Then, the values of four types of evaluation functions, H (High), L (Low), U (Up), and D (Down), are calculated in mora units using the slope of the time change of the pitch and the final pitch. Further, the value of the evaluation function of the entire mora constituting the accent phrase is added, the combination in which the added value is maximized is obtained, and the type of accent is determined.

Japanese Patent No. 4129899

Toshinori Ishii, Nobuaki Minematsu, Keikichi Hirose, "Accented Judgment of Japanese Continuous Speech Considering Pitch Perception", IEICE Technical Report, Vol.101, No.270, 23-30, 2001

However, in the method of Non-Patent Document 1, since the difference value of the pitches of adjacent mora is used, only (number of mora-1) types of accent types can be determined. Further, the method of Patent Document 1 combines evaluation functions of H, L, U, and D in mora units for accent phrases, and determines the type of accent according to the combination, and does not determine the accent type. Absent.

Further, in both the methods of Non-Patent Document 1 and Patent Document 1, the accent type is determined and the accent type is determined based on the pitch of only the individual accent phrases constituting the voice waveform data. There is.

In order to correctly determine the accent type of an accent phrase, it is necessary to consider the fluctuation of the pitch in the whole sentence of the text longer than the interval of the accent phrase or in the phrase (phrase) unit. It is not always possible to obtain a highly accurate accent type by focusing only on individual accent phrases as in the methods of Non-Patent Document 1 and Patent Document 1.

As described above, the conventional method has a problem that the accuracy of the accent type determination is lowered. The determined speech accent type is stored as a part of the acoustic model data (prosody data) at the time of pre-learning of speech synthesis, for example. Then, this acoustic model is used in text-to-speech synthesis for synthesizing speech for arbitrary text. If the determination accuracy of the accent type is lowered, the accent of the voice becomes unnatural when the text voice is synthesized using the acoustic model in which the accent type is learned, and it becomes impossible to realize the voice synthesis with high accuracy.

Therefore, in order to determine the accent type with high accuracy, a new method that considers the entire sentence of the text longer than the interval of the accent phrase or the entire voice waveform data of each phrase has been desired.

Therefore, the present invention has been made to solve the above problems, and an object of the present invention is to provide an accent type determination device and a program capable of accurately determining an accent type from voice waveform data.

In order to solve the above problem, the accent type determination device according to claim 1 cuts out a plurality of accent phrases from the voice waveform data, and in the accent type determination device for determining the accent type of the accent phrase, a frame from the voice waveform data. A voice waveform pitch calculation unit that calculates each pitch, a plurality of accent phrases cut out from the voice waveform data, a plurality of mora cut out from the accent phrase, and a pitch for each frame calculated by the voice waveform pitch calculation unit. Corresponds to the first pitch representative value calculation unit that calculates the pitch representative value for each mora and outputs the pitch representative value for each mora included in the accent phrase of the voice waveform data, and the voice waveform data. Based on a phonetic label creation unit that creates a context-dependent phonetic label that includes information on phrases, accent phrases, mora, and accent positions contained in the text, and the context-dependent phonetic label created by the phonetic label creation section. Then, a plurality of accent phrases are cut out from the text, a plurality of mora are cut out from the accent phrase, and the accent clause is based on the number of the mora contained in the accent phrase so as to express all accent types. The pitch for each frame is set for each accent type label creation unit that changes the accent position and creates a plurality of accent type context-dependent phonetic labels, and for each of the accent type context-dependent phoneme labels created by the accent type label creation unit. The pitch representative value for each mora is calculated from the pitch estimation unit obtained as the pitch estimation value reflecting the pitch fluctuation in the entire sentence or the phrase of the text and the pitch estimation value for each frame obtained by the pitch estimation unit. For each accent type context-dependent phonetic label, a second pitch representative value calculation unit that outputs a pitch representative value for each mora included in the accent phrase of the text and the first pitch representative value calculation unit The pitch representative value for each mora included in the accent clause of the output voice waveform data and the accent of the text output for each accent type context-dependent phonetic label by the second pitch representative value calculation unit. A distance calculation unit that calculates the distance between the pitch representative value for each mora included in the clause and outputs the distance for each accent type context-dependent phonetic label, and the accent type context by the distance calculation unit. The distance output for each dependent accent label is the minimum The accent type context-dependent phoneme label is specified, and the accent type determination unit for determining the accent type of the accent type context-dependent phoneme label as the accent type of the accent phrase is provided.

Further, in the accent type determination device according to claim 2, in the accent type determination device according to claim 1, the first pitch representative value calculation unit determines a pitch value for each frame included in the mora. The pitch value of the final frame in the predetermined formula is obtained as the pitch representative value of the mora, and the second pitch representative value calculation unit determines the pitch estimation value for each frame included in the mora. Approximate to the equation, the value of the pitch estimated value of the final frame in the predetermined equation is obtained as the pitch representative value of the mora.

Further, in the accent type determination device according to claim 3, in the accent type determination device according to claim 1, the pitch estimation unit obtains the pitch estimation value by using a prosody model learned in advance. It is characterized by.

Further, the accent type determination device according to claim 4 is the accent type determination device according to claim 3, and the prosody model is learned by using the voice emitted by the same speaker as the speaker of the voice waveform data. It is characterized by being a model.

Further, the program of claim 5 is characterized in that the computer functions as the accent type determination device according to any one of claims 1 to 4.

As described above, according to the present invention, it is possible to accurately determine the accent type from the voice waveform data.

It is a block diagram which shows the structural example of the accent type determination apparatus by embodiment of this invention. It is a flowchart which shows the processing example of the accent type determination apparatus. It is a figure explaining the pitch, a mora, an accent phrase, a text, a frame, and a pitch representative value of a mora of voice waveform data.

Hereinafter, embodiments for carrying out the present invention will be described in detail with reference to the drawings. FIG. 1 is a block diagram showing a configuration example of an accent type determination device according to an embodiment of the present invention, and FIG. 2 is a flowchart showing a processing example of the accent type determination device. Further, FIG. 3 is a diagram for explaining the pitch, mora, accent phrase, text, frame, and pitch representative value of the mora of the voice waveform data. In FIG. 3, the pitch curve of the voice waveform data shows most of the trends including breathing in the text section corresponding to the voice waveform data.

With reference to FIG. 1, the accent type determination device 1 includes a voice text analysis unit 10, a voice waveform pitch calculation unit 11, a pitch representative value calculation unit 12, a text analysis unit (phoneme label creation unit) 13, and an accent type label creation. It includes a unit 14, a prosody model 15, a pitch estimation unit 16, a pitch representative value calculation unit 17, a distance calculation unit 18, and an accent type determination unit 19. It is assumed that the voice waveform data is sampled with a sampling frequency of 16 kHz and a conversion bit number of 16 bits, and there is a transcribed text corresponding to the voice waveform data.

With reference to FIGS. 1 and 2, the accent type determination device 1 inputs the voice waveform data and the transcript text corresponding to the voice waveform data (step S201).

The voice text analysis unit 10 inputs voice waveform data and text. Then, the voice text analysis unit 10 sets the delimiter position of the accent phrase with respect to the text by analyzing the text (step S202). Since the technique for setting the delimiter position of the accent phrase is known, the description thereof is omitted here.

The voice text analysis unit 10 analyzes the voice waveform data and the text to perform phoneme segmentation and sets the phoneme delimiter position (step S203).

The processing of phoneme segmentation is known and, for example, the technique of forced alignment is used. For details on the forced alignment technology, refer to the following URL.
"The HTK Book (for HTK Version 3.4)", Internet <URL: http://htk.eng.cam.ac.uk/>

The voice text analysis unit 10 outputs data regarding the accent phrase delimiter position and data regarding the phoneme delimiter position to the pitch representative value calculation unit 12.

The voice waveform pitch calculation unit 11 inputs voice waveform data, extracts a fundamental frequency from the voice waveform data for each frame in units of 5 ms (step S204), and calculates the logarithm of the fundamental frequency (step S205). Hereinafter, the logarithm of the fundamental frequency is referred to as pitch.

The process of extracting the fundamental frequency for each frame is known, and for example, an audio signal processing technique (pitch command) is used. For details of this technology, refer to the following URL.
"REFERENCE MANUAL for Speech Signal Processing Toolkit Ver. 3.9", Internet <URL: http://sp-tk.sourceforge.net/>

The voice waveform pitch calculation unit 11 outputs the pitch of each frame in the voice waveform data to the pitch representative value calculation unit 12.

The pitch representative value calculation unit 12 inputs the pitch for each frame in the voice waveform data from the voice waveform pitch calculation unit 11, and also inputs the data regarding the accent phrase break position and the phoneme break position from the voice text analysis unit 10. To do.

The pitch representative value calculation unit 12 identifies (cuts out) a plurality of accent phrases constituting the voice waveform data based on the data relating to the delimiter positions of the accent phrases. Then, the pitch representative value calculation unit 12 identifies (cuts out) a plurality of moras constituting the accent phrase based on the data relating to the phoneme delimiter positions (step S206).

As shown in FIG. 3, the accent phrase is a section (arrow section) separated by the delimiter position (tip position of the arrow) of the accent phrase. Further, the mora specified from the phoneme dividing position is a section (arrow section) separated by the tip position of the arrow.

The pitch representative value calculation unit 12 calculates the pitch representative value for each mora from the pitch for each frame in the mora section (mora section) (step S207). As a result, the pitch representative value of the mora constituting the accent phrase is calculated for each of the accent phrases constituting the voice waveform data.

Specifically, the pitch representative value calculation unit 12 cuts out the mora constituting the accent phrase, and specifies the pitch for each frame included in the mora section from the start time to the end time of the mora. The pitch representative value calculation unit 12 treats the interpolated value as a pitch by interpolating the pitch for each frame in the adjacent voiced section in the unvoiced section where the pitch does not exist. Further, the pitch representative value calculation unit 12 interpolates the average value calculated from the pitch for each frame in the voiced section of the entire sentence when the unvoiced section portion is adjacent to the beginning, end of sentence, or pause (interval) in the sentence. Interpolate to the value of the endpoint of the powerless interval. As shown in FIG. 3, the mora is composed of a plurality of frames, and the pitch of each frame included in the mora section is specified.

Next, the pitch representative value calculation unit 12 approximates the pitch value for each frame included in the mora section as a linear function formula by the least squares method, and calculates the pitch value of the final frame in the linear function formula. , Set as the pitch representative value of the mora. Then, the pitch representative value calculation unit 12 outputs the pitch representative value of each mora in the accent phrase section (accent phrase section) to the distance calculation unit 18.

As shown in FIG. 3, the number of frames in the mora section is N, the frame number of the start time is x ₁ , the pitch value is y ₁ , the frame number of the end time is x _N , the pitch value is y _N , and the least squares method. Let f (x) = ax + b be the equation of the linear function approximated by. a is the slope of the linear function f (x), and b is the intercept.

The slope a and the intercept b of the linear function f (x) are represented by the following equations.

In this way, the pitch representative value calculation unit 12 sets the pitch value f (x _N ) = y _N of the final frame x _{N in} the approximate linear function equation f (x) as the pitch representative value of the mora. The pitch representative value S _m of each mora in the accent clause section is calculated. m is the mora number.

The text analysis unit 13 inputs the text corresponding to the voice waveform data and analyzes the text to create a context-sensitive phoneme label for the text (step S208). A context-sensitive phoneme label is a phoneme-based label that depends on the context of the entire text. Context-sensitive phoneme labels are, for example, the position, number and number of clauses contained in the entire text, the position, number and number of accent phrases contained in the phrase, the position, number and number of mora contained in the accent phrase, and within the accent phrase. Contains information to identify the context of the entire text, such as the location of accents in.

The process of creating context-sensitive phoneme labels is known. For details, refer to, for example, the technology of the Japanese speech synthesis system provided at the following URL.
"OPEN JTalk", Internet <URL: http://open-jtalk.sourceforge.net/>

The text analysis unit 13 outputs the context-sensitive phoneme label of the text to the accent type label creation unit 14.

The accent type label creation unit 14 inputs the context-sensitive phoneme label of the text from the text analysis unit 13. Then, the accent type label creation unit 14 identifies the delimiter positions of the plurality of accent phrases constituting the text based on the information contained in the context-sensitive phoneme label, and identifies (cuts out) the plurality of accent phrases constituting the text. ). The accent phrase specified by the accent type label creation unit 14 is the same as the accent phrase specified before the pitch representative value is calculated by the pitch representative value calculation unit 12.

The accent type label creation unit 14 identifies (cuts out) a plurality of moras constituting the accent phrase based on the information contained in the context-sensitive phoneme label. Then, the accent type label creation unit 14 changes the position of the accent core (the position of the accented mora) so as to express all the accent types for the accent phrase, and the context-dependent phoneme label (accent type context-dependent phoneme label). ) Is created (step S209). The number of context-sensitive phoneme labels created by the accent type label creation unit 14 is determined by the number of moras constituting the accent phrase. If the text text is composed of L accent phrases and the number of mora in the _l- th accent phrase is M _l , different context-sensitive phoneme labels of M _l types are created. Here, l is an integer of 1 or more and L or less, and M _l is an integer of 1 or more.

The accent type label creation unit 14 outputs M _l context-sensitive phoneme labels (context-dependent phoneme labels for each accent type) to the pitch estimation unit 16 for the l-th accent phrase.

The pitch estimation unit 16 inputs M _l context-sensitive phoneme labels for the lth accent phrase from the accent type label creation unit 14. Then, the pitch estimation unit 16 estimates the pitch for each frame in the mora section for each of the M _l context-sensitive phoneme labels using the pre-learned prosody model 15, and obtains the pitch estimation value (step S210). ).

The process of estimating the pitch for context-sensitive phoneme labels is known, and for details, refer to, for example, the technology of the Japanese speech synthesis system provided at the above URL or the speech synthesis technology provided at the following URL. ..
"HTS", Internet <URL: http://hts-engine.sourceforge.net/>

The pitch estimation unit 16 transmits the pitch estimation value for each frame in the mora section to the pitch representative value calculation unit 17 for each accent phrase for each of the M _l context-sensitive phoneme labels (context-dependent phoneme labels for each accent type). Output.

The prosody model 15 is a DB (database) created in advance by a model learning unit (not shown). The model learning unit creates context-dependent phoneme labels for voice waveform data such as sentences or phrases of various texts and text corresponding to the voice waveform data, calculates the pitch from the voice waveform data, and corresponds to this. The model learned so that the pitch can be estimated by using the context-dependent phoneme label is stored in the prosody model 15.

Therefore, the pitch estimation value calculated using the prosody model 15 is a value that reflects the fluctuation of the pitch (in the section longer than the accent phrase) in the whole sentence of the text or in the section such as the phrase. That is, the pitch estimation value is a value that reflects the accent position in the entire text sentence or phrase.

Here, the prosody model 15 uses the voice emitted by the same speaker as the speaker of the voice waveform data whose accent type is determined by the accent type determination device 1 as learning data, and uses the voice waveform data of the speaker as learning data. It is desirable that the DB is learned by using it. As a result, the accent type determination device 1 can determine the accent type from the voice waveform data with higher accuracy.

The pitch representative value calculation unit 17 inputs the pitch estimation value for each frame in the mora section for each accent phrase for each of the M _l context-sensitive phoneme labels from the pitch estimation unit 16. Then, the pitch representative value calculation unit 17 calculates the pitch representative value for each mora from the pitch estimation value for each frame in the mora section by the same processing as the pitch representative value calculation unit 12 (step S211). As a result, for each of the M _l context-sensitive phoneme labels, for each accent phrase that constitutes the text corresponding to the speech waveform data, the pitch representative value of the mora that constitutes the accent phrase is calculated.

Specifically, the pitch representative value calculation unit 17 cuts out the mora constituting the accent phrase based on the context-sensitive phoneme label, and specifies the pitch for each frame included in the mora section from the start time to the end time of the mora. To do. The pitch representative value calculation unit 17 performs interpolation processing and the like in the same manner as the pitch representative value calculation unit 12 for the silent section where the pitch does not exist. Then, the pitch representative value calculation unit 17 approximates the pitch value for each frame included in the mora section as an equation of the linear function by the least squares method, and sets the pitch of the final frame in the equation of the linear function to the mora. It is set as the pitch representative value of.

Of the M _l context-sensitive phoneme labels in the l-th accent phrase, the pitch representative value of the m-th mora is T _{l, n, m} for the n-th context-sensitive phoneme label.

The pitch representative value calculation unit 17 outputs the pitch representative value of each mora in the accent clause section to the distance calculation unit 18 for each of the M _l context-sensitive phoneme labels (context-dependent phoneme labels for each accent type).

The distance calculation unit 18 inputs the pitch representative value of each mora in the accent phrase section for the voice waveform data from the pitch representative value calculation unit 12. Further, the distance calculation unit 18 inputs the pitch representative value of each mora in the accent clause section for each of the M _l context-sensitive phoneme labels (context-dependent phoneme labels for each accent type) from the pitch representative value calculation unit 17. To do.

The distance calculation unit 18 sets the pitch representative value of each mora in the l-th accent clause section of the voice waveform data and the pitch representative value of each mora in the accent phrase section for each of the M _l context-dependent phoneme labels. The distance between them is calculated (step S212).

ここで、Ｍ_l個のモーラからなるアクセント句の場合、アクセント型の数はＭ_lであるから、ｎ＝１，・・・，Ｍ_l、ｍ＝１，・・・，Ｍ_lである。 As described above, the pitch representative value of the m-th mora in the l-th accent clause section of the voice waveform data is S _{l, m,} and the n-th of the M _l context-dependent phoneme labels in the l-th accent clause. _Let T _{l, n, m} be the pitch representative values of the m-th mora in the accent clause for the context-dependent phoneme label of. The distances C _{l and n} between the two are calculated by the following formula.

Here, in the case of an accent phrase consisting of M _l mora, since the number of accent types is M _l , n = 1, ..., M _l , m = 1, ..., M _l .

The distance calculation unit 18 calculates the distances (C _{l, 1} , ..., C _{l, Ml} ) for each of the M _l context-sensitive phoneme labels (context-sensitive phoneme labels for each accent type) for each accent phrase. Output to the accent type determination unit 19.

The accent type determination unit 19 inputs the distances (C _{l, 1} , ..., C _{l, Ml} ) for each of the M _l context-sensitive phoneme labels for each accent phrase. Then, the accent type determination unit 19 identifies the context-dependent phoneme label having the minimum distance among the M _l context-sensitive phoneme labels, and determines the accent type of the specified context-dependent phoneme label (step S213). Then, the accent type determination unit 19 outputs the determined accent type as the accent type of the accent phrase.

The accent type n _lmin that minimizes the distance in the l-th accent phrase is calculated by the following formula.

As described above, according to the accent type determination device 1 of the embodiment of the present invention, the pitch representative value calculation unit 12 sets the pitch value for each frame included in the mora section by the least squares method for the voice waveform data. Approximate to the formula of the next function, the pitch of the final frame in the formula of the linear function is set as the pitch representative value of the mora. As a result, the pitch representative value of the mora constituting the accent phrase is calculated for the voice waveform data.

The accent type label creation unit 14 changes the position of the accent nucleus for each accent phrase based on the context-sensitive phoneme label of the text generated by the text analysis unit 13, and M _l context-sensitive phoneme labels for each accent type. To create.

The pitch estimation unit 16 estimates the pitch for each frame in the mora section for each of the M _l context-sensitive phoneme labels using the prosody model 15, and obtains the pitch estimation value. The pitch representative value calculation unit 17 calculates the pitch representative value of the mora from the pitch estimation value for each frame in the mora section by the same processing as the pitch representative value calculation unit 12. As a result, for each of the M _l context-sensitive phoneme labels, the pitch representative value of the mora constituting the accent phrase is calculated for the text corresponding to the speech waveform data.

The distance calculation unit 18 includes the pitch representative value of each mora in the accent clause section of the voice waveform data calculated by the pitch representative value calculation unit 12, and M _l context-dependent phonemes calculated by the pitch representative value calculation unit 17. Calculate the distance between each mora's pitch representative value in the accent clause interval for each of the labels. The accent type determination unit 19 determines the accent type that minimizes the distance between M _l context-sensitive phoneme labels for each accent phrase, and outputs the determined accent type as the accent type of the accent phrase.

In the prior art, the accent type of the accent phrase is determined based on the pitch of only each accent phrase constituting the voice waveform data. On the other hand, in the embodiment of the present invention, the pitch estimation unit 16 has a pitch estimation value based on the prosody model 15 learned by using various texts and corresponding speech waveform data for the context-sensitive phoneme label for each accent type. Is calculated, and the accent type determination unit 19 determines the accent type using the distance calculated from this pitch estimation value. This pitch estimated value is a value considering the fluctuation of the pitch in the section such as a sentence or phrase of the text, that is, the fluctuation of the pitch in the section longer than the accent phrase. The determined accent type is also the result of considering the entire sentence or phrase that is longer than the accent phrase. Therefore, it is possible to determine the accent type with high accuracy from the voice waveform data.

Further, the determined accent type is stored as a part of the data (prosody data) of the acoustic model at the time of pre-learning of speech synthesis, and this acoustic model is used at the time of text speech synthesis. Therefore, if the determination accuracy of the accent type is improved, the accent of the voice becomes natural when the text voice is synthesized using the acoustic model in which the accent type is learned, and the voice synthesis with high accuracy can be realized.

As the hardware configuration of the accent type determination device 1 according to the embodiment of the present invention, a normal computer can be used. The accent type determination device 1 is composed of a computer provided with a volatile storage medium such as a CPU and RAM, a non-volatile storage medium such as a ROM, and an interface. Voice text analysis unit 10, voice waveform pitch calculation unit 11, pitch representative value calculation unit 12, text analysis unit 13, accent type label creation unit 14, prosody model 15, pitch estimation unit 16, pitch provided in the accent type determination device 1. Each function of the representative value calculation unit 17, the distance calculation unit 18, and the accent type determination unit 19 is realized by causing the CPU to execute a program describing these functions. These programs are stored in the storage medium, read by the CPU, and executed. In addition, these programs can be stored and distributed in storage media such as magnetic disks (floppy (registered trademark) disks, hard disks, etc.), optical disks (CD-ROM, DVD, etc.), semiconductor memories, etc., and can be distributed via a network. You can also send and receive.

Although the present invention has been described above with reference to embodiments, the present invention is not limited to the above-described embodiments, and various modifications can be made without departing from the technical idea. For example, in the above embodiment, the pitch estimation unit 16 of the accent type determination device 1 uses the prosody model 15 for each of the M _l context-sensitive phoneme labels to change the pitch of the entire text or a phrase. The pitch estimate value was calculated in consideration. In the present invention, it is not always necessary to use the prosody model 15. The pitch estimation unit 16 may calculate a pitch estimation value in consideration of pitch fluctuations of the entire text sentence or a phrase or the like by a known method without using the prosody model 15.

Further, in the above-described embodiment, the pitch representative value calculation unit 12 approximates the pitch value for each frame included in the mora section as a linear function equation by the least squares method, and determines the final frame in the linear function equation. The pitch value is set as the pitch representative value of the mora. On the other hand, the pitch representative value calculation unit 12 may approximate the pitch value for each frame included in the mora section as an equation of a linear function by a method other than the least squares method. Further, the pitch representative value calculation unit 12 may approximate the pitch value for each frame included in the mora section as a predetermined expression other than the expression of the linear function by a method such as the least squares method. Further, the pitch representative value calculation unit 12 may set the pitch value of a frame other than the final frame in a predetermined formula such as a linear function as the pitch representative value of the mora. The same applies to the pitch representative value calculation unit 17.

1 Accent type judgment device 10 Voice text analysis unit 11 Voice waveform pitch calculation unit 12 Pitch representative value calculation unit 13 Text analysis unit 14 Accent type label creation unit 15 Prosody model 16 Pitch estimation unit 17 Pitch representative value calculation unit 18 Distance calculation unit 19 Accent type judgment unit

Claims (5)

In an accent type determination device that cuts out a plurality of accent phrases from voice waveform data and determines the accent type of the accent phrase.
A voice waveform pitch calculation unit that calculates the pitch for each frame from the voice waveform data,
A plurality of accent phrases are cut out from the voice waveform data, a plurality of mora are cut out from the accent phrase, and a pitch representative value for each mora is calculated from the pitch for each frame calculated by the voice waveform pitch calculation unit. A first pitch representative value calculation unit that outputs a pitch representative value for each mora included in the accent phrase of the voice waveform data, and a first pitch representative value calculation unit.
A phoneme label creation unit that creates a context-dependent phoneme label that includes information on phrases, accent phrases, mora, and accent positions included in the text for the text corresponding to the speech waveform data.
Based on the context-dependent phoneme label created by the phoneme label creation unit, a plurality of accent phrases are cut out from the text, a plurality of mora are cut out from the accent phrase, and all accent types are expressed for the accent phrase. As described above, the accent type label creation unit that changes the accent position based on the number of the mora included in the accent clause and creates a plurality of accent type context-dependent phoneme labels.
A pitch estimation unit that obtains the pitch for each frame for each accent-type context-sensitive phoneme label created by the accent-type label creation unit as a pitch estimation value that reflects pitch fluctuations in the entire sentence or phrase of the text. ,
A pitch representative value for each mora is calculated from the pitch estimation value for each frame obtained by the pitch estimation unit, and the pitch for each mora included in the accent phrase of the text for each accent type context-sensitive phoneme label. A second pitch representative value calculation unit that outputs representative values, and
The pitch representative value for each mora included in the accent clause of the voice waveform data output by the first pitch representative value calculation unit, and the accent type context-dependent phoneme label by the second pitch representative value calculation unit. A distance calculation unit that calculates the distance between the pitch representative value for each mora included in the accent clause of the text output for each, and outputs the distance for each accent type context-dependent phoneme label. ,
The distance calculation unit identifies the accent-type context-dependent phoneme label that minimizes the distance output for each accent-type context-dependent phoneme label, and sets the accent type of the accent-type context-dependent phoneme label as the accent phrase. Accent type judgment unit that judges as accent type,
An accent type judgment device characterized by being equipped with. In the accent type determination device according to claim 1,
The first pitch representative value calculation unit is
The pitch value for each frame included in the mora is approximated to a predetermined formula, and the pitch value of the final frame in the predetermined formula is obtained as the pitch representative value of the mora.
The second pitch representative value calculation unit is
Accent type determination characterized in that the pitch estimation value for each frame included in the mora is approximated to a predetermined formula, and the value of the pitch estimation value of the final frame in the predetermined formula is obtained as the pitch representative value of the mora. apparatus. In the accent type determination device according to claim 1 or 2.
The pitch estimation unit
An accent type determination device characterized in that the pitch estimation value is obtained using a prosody model learned in advance. In the accent type determination device according to claim 3,
An accent type determination device, characterized in that the prosody model is a model learned by using a voice uttered by the same speaker as the speaker of the voice waveform data. A program for causing a computer to function as the accent type determination device according to any one of claims 1 to 4.

Images