JP2017090856A - Voice generation device, method, program, and voice database generation device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a technique for generating accent information added to phoneme piece data so as to obtain correct accent information by taking the blank timing of voice data into consideration.SOLUTION: In an association process for associating the phoneme string of segment data with that of a morpheme label data, determination is made as to whether the phoneme label of the segment data side matches the phoneme level of the morpheme label data side sequentially from the collection head of the segment data and the collection head of the morpheme label data (S1205), and when the phoneme label of the segment data is "#" corresponding to a blank segment (YES in S1204) and the phoneme label of the morpheme label data side is not "#" corresponding to a punctuation mark (NO in S1208), the morpheme data of a comma is inserted next to current morpheme data, and thus the association is appropriately carried out.SELECTED DRAWING: Figure 12

Description

  The present invention relates to a voice creation device, method, program, and voice database creation device for creating an accent information voice given to phoneme data.

  In the creation of a speech database for speech synthesis, a segment unit segment in units of phonemes is extracted from speech data for learning into phoneme segment data, and the phoneme label corresponding to the segment is used as a key to the speech database. be registered. Accent information is added to the registered phoneme piece data in order to improve the synthesis quality. At this time, first, morphological analysis processing is executed on the text data corresponding to the speech data, and accent information is assigned to each phoneme obtained as a result. Next, for each segment, the phoneme of the segment is associated with the phoneme obtained by the morphological analysis process, and the accent information attached to the phoneme is acquired and registered in the speech database together with the phoneme piece data of the segment Is done.

  Here, the speech data for learning that is the basis of the phoneme piece data is not necessarily uttered with an accent based on the language information obtained from the corresponding text data. Therefore, conventionally, there is known a technique for correcting accent information generated based on language information obtained from text data based on fundamental frequency information obtained from speech of speech data (for example, described in Patent Documents 1 and 2). Technology).

JP 2012-189703 A JP-A-6-337691

  As described above, in order to give accent information to phoneme piece data, correspondence between phonemes for each segment obtained from speech data and phonemes obtained by morpheme analysis processing for text data corresponding to the speech data Need to do.

  Here, the breathing is performed in the actual speech of the voice data, and the value of the voice data is silent or close to the value at the timing when the breathing is performed. Hereinafter, this timing is referred to as “blank timing”. When the audio data is segmented, the blank timing is detected as a segment having a phoneme indicating silence. On the other hand, this blank timing generally corresponds to a morpheme break and there are many punctuation marks at this position, but there are cases where punctuation marks are not detected. In addition, when the blank timing occurs due to stagnation, the blank timing may be a position of a phoneme in the middle of one morpheme on the language information, for example. In these cases, the correspondence between the phonemes for each segment and the phonemes obtained by the morphological analysis processing is not performed well.

  However, in the above-described conventional technology, since the occurrence of blank timing is not taken into consideration, there is a problem in that correct correspondence information may not be obtained for each phoneme piece data because the correspondence relationship is shifted. In addition, the correspondence between the position of the accent information obtained from the linguistic information and the position of the fundamental frequency obtained from the speech of the speech data also shifts, so there is a problem that the correction of the accent information based on the fundamental frequency may not be successful. was there.

  It is an object of the present invention to obtain correct accent information by considering the blank timing of audio data.

  In one example, morpheme data including a plurality of morphemes generated from text data corresponding to voice data and acquisition processing for acquiring first position data indicating at least one of an accent position and a break position from input voice data A comparison process for comparing the second position data indicating at least one of the position of the accent given to and the separation position between the plurality of morphemes and the first position data acquired from the speech data; When the 1st and 2nd position data do not correspond, the processing part which performs processing which gives the 1st position data instead of the 2nd position data to morpheme data is provided.

  According to the present invention, correct accent information can be obtained by considering the blank timing of audio data.

It is a block diagram of an embodiment of a voice information creation device according to the present invention. It is a figure which shows the data structural example of segment data. It is a figure which shows the specific example of segment data. It is a figure which shows the data structural example of morpheme data. It is a figure which shows the specific example of morpheme data. It is a figure which shows the data structural example of morpheme label data. It is a figure which shows the specific example of morpheme label data. It is a figure which shows the data structural example of the label data for learning. It is a figure which shows the data structural example of fundamental frequency data. It is a figure which shows the hardware structural example of the computer which can implement | achieve an audio | voice information production apparatus as a software process. It is a flowchart which shows the example of a data provision process. It is a flowchart (the 1) which shows the detailed example of a matching process. It is a flowchart (the 2) which shows the detailed example of a matching process. It is a figure which shows the example of the morpheme data after adjustment. It is a flowchart which shows the detailed example of an accent correction process.

  Hereinafter, embodiments for carrying out the present invention will be described in detail with reference to the drawings. In speech synthesis, a phoneme data string that most closely matches a phoneme string to be synthesized is searched from a speech database, and speech data is synthesized by combining these phoneme data. The present embodiment relates to a technique for creating accent information given to phoneme data when generating phoneme data to be registered in speech data for speech synthesis.

  Here, “speech synthesis” is a process of synthesizing speech data that can cause a computer to utter speech similar to speech when a human reads out the text data when text data representing a sentence is input. Say. The “phoneme piece data” means time domain speech waveform data representing one phoneme. The “phoneme” refers to the smallest linguistic unit that causes a difference in meaning. For example, when there is a word “everything”, “a” and “r” are decomposed so that each has a difference in meaning. “A” “y” “u” “r” “u” each form a phoneme. In the speech synthesis process, first, morpheme data is obtained by executing a morpheme analysis process on the input text data while referring to the morpheme dictionary. Here, “morpheme” refers to the smallest unit in the language that is responsible for its own meaning. For example, when there is text data with the content that “every reality is twisted towards you,” The content of the text data by the morphological analysis processing is, for example, “everything” “reality” “to” “all” “self” “no” “how” “to” “screw bending” “ta” “no” “da” “. Is broken down into 13 morphemes. As the next stage of speech synthesis, the morpheme data is further decomposed to generate data consisting of a sequence of phonemes (hereinafter referred to as “morpheme label data” in this embodiment). For example, when there is one “morpheme” obtained from the text data, the morpheme corresponds to the phoneme “a” “r” “a” “y” “u” “r” “u”, respectively. Each morpheme label data having a label is generated. As the next stage of speech synthesis, for each morpheme label data, the above-mentioned speech database is searched for each phoneme label, so that phoneme data to which the same phoneme label as that phoneme label is assigned is searched. As the final stage of speech synthesis, speech data corresponding to the input text sentence is synthesized by combining the speech data extracted for each morpheme label data.

  FIG. 1 shows an implementation of a speech information creation device 100 according to the present invention for creating accent information to be given to phoneme data when generating phoneme data registered in a speech database for speech synthesis. It is a block diagram of a form. The speech information creation apparatus 100 includes a speech analysis unit 101, a language analysis unit 102, and a data addition unit 103, and finally outputs learning label data 130. The voice analysis unit 101 includes a voice recognition unit 110, an acoustic model 111, and a fundamental frequency analysis unit 112. The language analysis unit 102 includes a morpheme analysis unit 120 and a morpheme dictionary 121.

  The speech recognition unit 110 performs speech recognition processing on the input speech data 113 for learning time domain while referring to the acoustic model 111, so that each speech data 113 is continued with one phoneme. And segment data 114 is output as a result.

  FIG. 2 is a diagram illustrating a data configuration example of the segment data 114. As shown in this figure, each segment data segment [0], segment [1],..., Segment [L-1] is obtained for each of the total L phonemes obtained by the speech recognition processing by the speech recognition unit 110. Holds variable data of seg_id, phone, start, end, prev, and next, respectively. seg_id holds a segment ID (identifier). phone holds a phoneme label. start holds the start time (sample number). end holds the end time (sample number). prev holds a pointer to the previous segment data 114, and next holds a pointer to the next segment data 114. If the current segment data 114 is, for example, segmen [1], prev holds the start address of segment [0], and next holds the start address of segment [2]. Further, if the current segment data 114 is, for example, the head data segment [0], prev holds a null value that is an undefined value. If the current segment data 114 is, for example, end data segment [L-1], next holds a NULL value. By connecting the segment data 114 using prev and next, even if the segment data 114 is added to the last address on the memory, for example, the segment data 114 can be inserted at an arbitrary location.

  FIG. 3 is a diagram showing a specific example of segment data. For example, in the segment data 114 of seg_id = 3, the phone holds “a” as a phoneme label. start holds “730” (sample) as the start time. End holds “810” (sample) as the end time. prev holds “2” as the seg_id of the previous segment data 114. “next” holds “4” as the seg_id of the next segment data 114. In FIG. 3, the phone of each segment data 114 of the first seg_id = 0 or the 17th seg_id = 16 holds “#”. This “#” indicates that in the segment section corresponding to the segment data 114 in which they are registered, the value of the audio data 113 is determined to be no sound or zero silence.

  In particular, at the time of registration in a speech database (not shown), waveform data corresponding to the sample section is cut out from the speech data 113 for each sample section of the segment indicated by the start and end of the segment data 114 described above and used as phoneme piece data. The The phoneme piece data is registered in the speech database using the phoneme label representing the phoneme corresponding to the segment as a search key.

  At the time of speech synthesis, for each morpheme label data, it is not possible to obtain high-quality synthesized speech data simply by searching phoneme segment data corresponding to each phoneme label from the speech database. Ideally, the synthesized speech data with the best quality can be obtained if the phoneme piece data generated from the speech data when the human actually reads the input text sentence is registered and searched in the speech database. However, since the storage capacity of the speech database is limited, phoneme piece data corresponding to all text sentences cannot be registered. Therefore, accent information is added to the phoneme piece data and registered in the speech database. On the other hand, at the time of speech synthesis, accent information is also generated at the time of morpheme analysis processing for text data to be synthesized, and is given to each morpheme label data. Then, for each morpheme label data, when searching for phoneme piece data corresponding to each phoneme label from the speech database, phoneme piece data having accent information having a value close to the accent information given to the morpheme label data is extracted. And combine. As a result, it is possible to select phoneme piece data generated in an accent state close to the text data to be synthesized from the speech database, and it is possible to improve the quality of the synthesized speech data.

  Specifically, in the present embodiment shown in FIG. 1, the morphological analysis unit 120 constituting the language analysis unit 102 first inputs the utterance text 122 that is text data corresponding to the speech data 113 for learning, A morpheme analysis process is executed on the utterance text 122 while referring to the morpheme dictionary 121. As a result, the morpheme analyzer 120 outputs morpheme data 123 and morpheme label data 124. The morpheme analyzer 120 detects an accent position for each accent block determined based on the morpheme indicated by the morpheme data 123. Here, the term “accent” refers to the relative level of the fundamental frequency of a phoneme or the relative strength of the signal strength of a phoneme in a word or word combination (a range that can be expressed in one word). This is called “block”. The “accent position” refers to the position of the mora having the relatively highest fundamental frequency or the position of the mora having the relatively strongest signal strength in the accent block. “Mora” refers to a segmental unit of sound having a certain length of time in phonological theory, and is a convenient unit for expressing the position of an accent in an accent block. Next, the morpheme analysis unit 120 adds, for each morpheme label data 124, accent information indicating the positional relationship with the accent position corresponding to the accent block to which the phoneme indicated by the data belongs, to the morpheme label data 124. The “accent information” is, for example, the number of mora of the phoneme corresponding to the morpheme label data 124 from the accent position corresponding to the accent block to which the phoneme belongs to the mora to which the phoneme belongs (a negative value before the accent position and a plus after the accent position) Value) and a mora number indicating the number of mora included in the accent block from the first mora.

  FIG. 4 is a diagram illustrating a data configuration example of the morpheme data 123. As shown in this figure, each morpheme data morph [0], morph [1],..., Morph [M-1] is obtained for each of the total M morphemes obtained by the morpheme analysis processing by the morpheme analyzer 120. Holds morph_id, original, read, pronounce, group, accent, prev, and next variable data, respectively. morph_id holds a morpheme ID. original holds the morpheme on the notation. read holds the reading on the notation. pronounce holds the pronunciation. group holds part-of-speech information. Accent holds accent information. prev holds a pointer (null head) to the previous morpheme data 123. “next” holds a pointer to the next morpheme data 123 (terminal is NULL). As in the case of the segment data 114, even if the morpheme data 123 is added to the last address on the memory by connecting the morpheme data 123 using prev and next, the morpheme data 123 is stored at an arbitrary location. Can be inserted into.

  FIG. 5 is a diagram illustrating a specific example of the morpheme data 123. For example, when there is data of the utterance text 122 for learning with the content “every reality is twisted towards me”, the content is, for example, “everything” “reality” “ ”“ All ”“ My ”“ No ”“ How ”“ To ”“ Screw Bending ”“ Ta ”“ No ”“ Da ”“. ”The morpheme corresponding to each notation, reading, The morpheme data 123 shown in FIG. 5 in which the pronunciation is registered as original, read, and promise is generated. For example, in the first morpheme data 123 with morph_id = 0, original holds “any” as a morpheme on the notation. “read” holds “Arayul” as a notation reading. pronounce holds “Ayur” as a pronunciation. The group holds “combinations” as part-of-speech information. The Accent variable holds “3” as accent information. This indicates that there is an accent position at the third mora in the morpheme. The mora will be described later. prev holds “NULL” indicating that there is nothing in front as the morph_id of the previous morpheme data 123. “next” holds “1” as the morph_id of the next morpheme data 123.

  FIG. 6 is a diagram illustrating a data configuration example of the morpheme label data 124. As shown in this figure, a total of N morpheme label data morph_label [0], morph_label [1],... Determined based on the morpheme data 123 obtained by the morpheme analysis processing by the morpheme analysis unit 120. Each morph_label [N-1] holds variable data of mlabel_id, phone, accent, mola, morph_id, prev, and next. mlabel_id holds a morpheme label ID. phone holds a phoneme label. accent holds the number of mora up to the accent position. mola holds the mora number. morph_id holds the morpheme ID (see FIG. 4) of the morpheme data 123 to which the phoneme indicated by the phoneme label held in the phone belongs. prev holds a pointer (null head) to the previous morpheme label data 124. “next” holds a pointer to the next morpheme label data 124 (null end). The meanings of prev and next are the same as in the case of the segment data 114 (FIG. 2) or the morpheme data 123 (FIG. 4).

  FIG. 7 is a diagram illustrating a specific example of the morpheme label data 124. Each morpheme data having each morpheme ID of morph_id = 0,1,2,3 shown in FIG. 5 corresponding to, for example, four consecutive morphemes “every”, “reality”, “to”, and “all” by morpheme analysis processing. From 123, 21 phonemes “a” “r” “a” “y” “u” “r” “u” “g” “e” “N” “j” “i” “ts” “u” “ “o”, “s”, “u”, “b”, “e”, “t”, and “e” are extracted, and 21 morpheme label data 124 in which phoneme labels corresponding to the respective phonemes are registered in the phone is generated. For convenience of processing, data having a phoneme label “#” indicating silence is registered as the morpheme label data 124 having the first mlabel_id = 0 morpheme label ID.

  Here, for example, with respect to the morpheme “any”, “real”, “to”, and “all” that generated the morpheme label data 124 by the morpheme analysis process, the accent block “every” having the same delimiter as the morpheme delimiter is used. “Reality”, “O” and “All” are determined. Taking the accent block “every” as an example, four mora “a”, “ra”, “yu”, and “ru” are recognized in the accent block by the morphological analysis process. Then, by the morphological analysis processing, for example, the third mora “yu” (“yu”) is detected as the accent position in the accent block “every”. That is, in this example, the accent position is “3”.

  Furthermore, mlabel_id = 1,2,3,4,5 having each phoneme label corresponding to each of phonemes “a”, “r”, “a”, “y”, “u”, “r”, and “u” by morpheme analysis processing , 6, 7 in each morpheme label data 124, each mola variable includes mora numbers “1”, “2”, “2”, “3”, “3” of each mora including each phoneme in the accent block “every”. “3”, “4” and “4” are retained. That is, in the morpheme label data 124 of mlabel_id = 1 and phone = “a”, the phoneme of the phoneme label “a” is the first of the four mora “a” “ra” “yu” “ru”. Therefore, the mora number “1” is held in the mola. Further, in the morpheme label data 124 of mlabel_id = 2 and phone = “r”, the phoneme of the phoneme label “r” is included in the second mora “ra” of the four mora. Holds the mora number “2”. Similarly, in the morpheme label data 124 of “a” of mlabel_id = 3 and phone =, the phoneme of the phoneme “a” is included in the second mora “ra” of the four mora, Holds the mora number “2”. The same applies to other phonemes.

  In addition, mlabel_id = 1,2,3,4, each phoneme label corresponding to each of phonemes “a”, “r”, “a”, “y”, “u”, “r”, and “u” is obtained by morphological analysis processing. In each of the morpheme label data 124 of 5, 6, and 7, each accent variable includes the mora number “1” “2” “2” “3” “3” “4” “4” held in each mola variable. "And the difference between the accent position" 3 "detected in the accent block" 1-3 = -2 "," 2-3 = -1 "," 2-3 = -1 "," 3-3 = “0”, “3-3 = 0”, “4-3 = 1”, and “4-3 = 1” are registered respectively.

  As described above, in FIG. 1, when the morpheme label data 124 in which the phoneme string including the accent information is registered is obtained by the morpheme analysis unit 120 in the language analysis unit 102, the data adding unit 103 performs the speech analysis unit 101. A correspondence relationship between the phoneme string of the segment data 114 generated by the voice recognition unit 110 and the phoneme string of the morpheme label data 124 is generated. For example, in the segment data 114 illustrated in FIG. 3 and the morpheme label data 124 illustrated in FIG. 7, the data adding unit 103 sequentially determines the phoneme label and the morpheme label of the phone variable of the segment data 114 from the top of each. It is checked whether or not the phoneme label of the phone variable of the data 124 matches. When the phoneme labels of the segment data 114 and the morpheme label data 124 are the same, the data adding unit 103 adds the data structure of FIG. 6 to the contents of the segment data 114 having the data structure example of FIG. Accent information registered in the morpheme label data 124 having an example, that is, the number of mora up to the accent position and the mora number in the accent block are given and output as the learning label data 130.

  FIG. 8 is a diagram illustrating a data configuration example of the learning label data 130. The number of learning label data 130 is basically the same as the L pieces of segment data 114 illustrated in FIG. L pieces of learning label data train_label [i] (0≤i≤L) corresponding to L pieces of segment data segment [i] (0≤i≤L) are tlabel_id, phone, start, end, accent, Holds each variable data of mola, pitch, vowel, prev, next. tlabel_id is a learning label ID. phone, start and end are respectively copied from phone, start and end (see FIG. 2) of the segment data segment [i]. Accent and mola are copied from the accent and mola (see FIG. 6) of the morpheme label data morph_label [j] (0 ≦ j ≦ N) associated with the segment data segment [i]. vowel is a flag indicating whether or not the phoneme label indicated by the phone is a vowel, and is set to “1” if it is a vowel and “0” if it is a consonant. prev holds a pointer to the previous learning label data 130, and next holds a pointer to the next learning label data 130. The pitch will be described later.

  Using the learning label data 130 obtained in this way, the data of the section corresponding to the start and end (see FIG. 8) of the learning label data 130 is cut out from the speech data 113 of FIG. This phoneme piece data is registered in the speech database together with the phoneme label registered in the phone of the learning label data 130 of FIG. 8 and the accent information registered in the accent and mola. During speech synthesis, the accent position is detected for each accent block even during morphological analysis of text data, and for each phoneme in the synthesis target phoneme string, a value indicating the positional relationship with the accent position of the accent block to which the phoneme belongs is obtained. Is done. Then, when searching for phoneme data corresponding to the phoneme label from the speech database, phoneme data to which accent information having a value closest to the value indicating the positional relationship acquired for each phoneme is given. Is extracted. As a result, it is possible to select phoneme piece data that is pronounced in the same accent state as the text sentence data, and it is possible to improve the quality of the synthesized speech data.

  Here, as described in the section “Problems to be Solved by the Invention”, breathing occurs in the actual speech of the voice data 113, and the timing of the voice data 113 is a silence value or a silence value. It is blank timing that takes a value. When the audio data 113 is segmented, the blank timing is associated with the segment data 114 having the phoneme label “#” indicating silence. For example, in the phoneme label “#” held in the phone variable of each segment data 114 of the first seg_id = 0 or the 17th seg_id = 16 illustrated in FIG. 3, the segment corresponding to the segment data 114 is blank. Indicates a segment. On the other hand, this blank timing generally corresponds to a delimiter of language information. In this case, when the morphological analysis process is executed, the blank timing is detected as a punctuation morpheme. When the morpheme label data 124 is extracted from such punctuation mark morphemes, the phoneme label “#” indicating blank timing is registered in the phone variable of the morpheme label data 124 (see FIG. 6). In this state, when the data assigning unit 103 associates the phoneme string of the segment data 114 with the phoneme string of the morpheme label data 124, the phoneme labels “#” are well matched and correct. A correspondence is generated.

  On the other hand, in actual speech of voice data, punctuation marks may not be detected in the language information at the timing when silence timing occurs due to breathing. In this case, the phoneme label “#” indicating the blank timing is registered in the segment data 114, but the punctuation morpheme corresponding to the blank timing is not output, and the phoneme label “#” indicating the blank timing in the phone variable is output. The morpheme label data 124 in which “is registered” is not output. For example, in the example of the segment data 114 in FIG. 3, the phoneme string “arayurugeNjitsuo” (notation: “every reality”) and the phoneme string “subete” (notation: “all”) are connected in the pronunciation of the speech data 113. Between the segment data 114 of seg_id = 15 and phoneme label phone = “o” and the segment data 114 of seg_id = 17 and phoneme label phone = “s”, seg_id = 16, phoneme label phone = Segment data 114 of a blank segment “#” is generated. On the other hand, in the example of the morpheme data 123 shown in FIG. 5, no punctuation mark is detected between the morpheme “O” and the morpheme “all”, and therefore the example of the morpheme label data 124 shown in FIG. Also, between the morpheme label data 124 of mlabel_id = 15 and phoneme label phone = “o” and the morpheme label data 124 of mlabel_id = 16 and phoneme label phone = “s”, the phoneme label phone = “#” The blank segment morpheme label data 124 is not generated. In such a state, the data adding unit 103 has the phoneme string of the segment data 114 illustrated in FIG. 3 and the phoneme string of the morpheme label data 124 illustrated in FIG. 7 without considering the presence of the blank segment. , Segment data 114 of seg_id = 15 and phoneme label phone = “o” is matched with morpheme label data 124 of mlabel_id = 15 and phoneme label phone = “o”. Later, the segment data 114 of the blank segment with seg_id = 16 and phoneme label phone = “#” is compared with the morpheme label data 124 with mlabel_id = 17 and phoneme label phone = “s”. Since the phoneme labels do not match and the correspondence is not established, the result is that the phoneme piece data based on the voice data 113 cannot be used for registration in the voice database.

  Therefore, in the present embodiment, the data adding unit 103 generates a correspondence relationship between the phoneme string of the segment data 114 and the phoneme string of the morpheme label data 124 while determining the positional relationship between the blank segment and the morpheme separator, Control is performed so that there is no deviation in the correspondence. Details of this processing will be described later with reference to the flowcharts of FIGS.

  Here, the speech data 113 that is the basis of the phoneme piece data is not necessarily uttered at the accent position based on the morpheme data 123 obtained from the corresponding utterance text 122. Therefore, in the present embodiment, the data assigning unit 103 acquires the accent obtained from the morpheme label data 124 based on the correspondence between the phoneme string of the segment data 114 correctly generated as described above and the phoneme string of the morpheme label data 124. The new accent position for each block is recalculated as the position corresponding to the segment with the highest fundamental frequency extracted based on the speech data 113. Then, the data adding unit 103 corrects the accent information of the learning label data 130 corresponding to the segment belonging to the accent block based on the newly calculated accent position.

  Specifically, first, the fundamental frequency analysis unit 112 in the speech analysis unit 101 of FIG. 1 extracts the fundamental (pitch) frequency of the speech data 113 every predetermined frame period (for example, 256 milliseconds), and the fundamental frequency data 115 is output.

  FIG. 9 is a diagram illustrating a data configuration example of the basic frequency data 115. The K fundamental frequency data pitch [i] (0 ≦ i ≦ K−1) obtained as a result of the analysis of the fundamental frequency for each predetermined frame period with respect to the audio data 113 are time, pitch, prev, and next variables, respectively. Retain data. time holds the time (sample number at the beginning, center, or end of the current frame period) corresponding to the current frame period. The pitch holds the fundamental frequency [Hz] (Hertz) obtained as a result of the analysis. prev holds a pointer to the previous fundamental frequency data 115, and next holds a pointer to the next fundamental frequency data 115.

  Using the fundamental frequency data 115 generated by the fundamental frequency analysis unit 112 as described above, the data adding unit 103 generates a new accent position for each accent block, and for learning corresponding to the segment belonging to the accent block The accent information of the label data 130 is corrected based on the newly calculated accent position. Details of this processing will be described later with reference to the flowchart of FIG.

  FIG. 10 is a diagram illustrating a hardware configuration example of a computer that can realize the functions of the speech analysis unit 101, the language analysis unit 102, and the data addition unit 103 of the speech information creation apparatus 100 of FIG. 1 as software processing. In the computer shown in FIG. 10, a CPU 1001, a ROM (Read Only Memory) 1002, a RAM (Random Access Memory) 1003, an input device 1004, an output device 1005, an external storage device 1006, and a portable recording medium 1010 are inserted. A portable recording medium driving device 1007 and a communication interface 1008, which are connected to each other by a bus 1009. The configuration shown in the figure is an example of a computer that can implement the above system, and such a computer is not limited to this configuration.

  The ROM 1002 is a speech analysis processing program that realizes the function of the speech analysis unit 101 in FIG. 1, a language analysis processing program that realizes the function of the language analysis unit 102 in FIG. 1, and data that realizes the function of the data addition unit 103 in FIG. It is a memory for storing each program including a grant processing program. The RAM 1003 is a memory that temporarily stores a program or data stored in the ROM 1002 when each program is executed.

  The external storage device 1006 is, for example, an SSD (solid state drive) storage device or a hard disk storage device, and includes the voice data 113, speech text 122, segment data 114, fundamental frequency data 115, morpheme data 123, and morpheme label data 124 shown in FIG. The learning label data 130 and the voice database not shown in FIG.

  The CPU 1001 controls the entire computer by reading each program from the ROM 1002 to the RAM 1003 and executing it.

  The input device 1004 detects an input operation by a user using a keyboard, a mouse, or the like, and notifies the CPU 1001 of the detection result.

  The output device 1005 outputs data sent under the control of the CPU 1001 to a display device or a printing device.

  The portable recording medium driving device 1007 accommodates a portable recording medium 1010 such as an optical disk, SDRAM, or compact flash, and has an auxiliary role for the external storage device 1006.

  The communication interface 1008 is a device for connecting, for example, a LAN (local area network) or WAN (wide area network) communication line.

  The system according to the present embodiment includes a speech analysis processing program that realizes the function of the speech analysis unit 101 in FIG. 1, a language analysis processing program that realizes the function of the language analysis unit 102 in FIG. 1, and the function of the data addition unit 103 in FIG. Is realized by the CPU 1001 reading out the data addition processing program for realizing the above from the ROM 1002 to the RAM 1003 and executing it. For example, the program may be recorded and distributed in the external storage device 1006 or the portable recording medium 1010, or may be acquired from the network by the network connection device 1008.

  The functions of the speech analysis unit 101 and the language analysis unit 102 in FIG. 1 realized by the CPU 1001 executing the speech analysis processing program and the language analysis processing program are as described above with reference to FIGS. As a result, the RAM 1003 in FIG. 10 stores the segment data 114 in the data configuration example in FIG. 2 (specific example is FIG. 3), the morpheme data 123 in the data configuration example in FIG. 4 (specific example in FIG. 5), and the data configuration in FIG. Example morpheme label data 124 (specific example is FIG. 7) and basic frequency data 115 of the data configuration example of FIG. 9 are generated.

  FIG. 11 is a flowchart illustrating an example of a data providing process for realizing the function of the data adding unit 103 in FIG. This data addition process is a process in which the CPU 1001 reads out and executes a data addition process program stored in the ROM 1002 while using the RAM 1003 as a work memory.

  In the data providing process, the CPU 1001 first executes an association process (step S1101).

  Next, the CPU 1001 determines whether or not an error has been detected as a result of the association processing (whether or not an error flag described later on the RAM 1003 is on) (step S1102). If no error is detected (NO in step S1102), the CPU 1001 executes accent correction processing (step S1103). As a result, the CPU 1001 outputs the learning label data 130 (both refer to FIG. 1) corresponding to the input voice data 113 and the utterance text 122, and ends the data addition process. When an error is detected (when the determination in step S1102 is YES), the CPU 1001 does not adopt the input voice data 113 and the utterance text 122 for learning, and does not output the learning label data 130. Finally, the data providing process is terminated.

  12 and 13 are flowcharts showing a detailed example of the association processing in step S1101 of FIG. In this association processing, as described above, the phoneme string (see FIGS. 2 and 3) of the segment data 114 generated on the RAM 1003 while the positional relationship between the blank segment and the morpheme break is determined. An appropriate correspondence with the phoneme string of the morpheme label data 124 (FIGS. 6 and 7) is generated.

  In the associating process, the CPU 1001 first determines whether the phoneme label on the segment data 114 side and the phoneme label on the morpheme label data 124 side match sequentially from the top of the set of segment data 114 and the top of the set of morpheme label data 124. Processing for determining whether or not (step S1205), processing for determining whether or not the phoneme label on the segment data 114 side is “#” corresponding to the blank segment (step S1204), and determination in step S1204 is YES In this case, the process of determining whether or not the phoneme label on the morpheme label data 124 side is “#” corresponding to the punctuation mark (step S1208) is repeatedly executed.

  In order to control this repetitive processing, the CPU 1001 sets the start address of the segment data 114 on the RAM 1003 and the start address of the morpheme label data 124 on the RAM 1003 in the variables seg and morph on the RAM 1003 (step S1201). In the data configuration example of the segment data 114 in FIG. 2, the address of segment [0] is set in the variable seg, and in the data configuration example of the morpheme label data 124 in FIG. 6, the address of morph_label [0] is set in the variable morph. The Then, the CPU 1001 repeatedly executes a series of processes from step S1203 to S1215 shown below while it is determined that neither the variable seg nor the variable morph exceeds the terminal data (when the determination in step S1202 is YES). .

  First, the CPU 1001 turns off both the increment flag on the segment data 114 side and the increment flag on the morpheme label data 124 side stored in the RAM 1003 (step S1203).

  Next, the CPU 1001 determines whether or not the phoneme label (the value of the phone variable in FIG. 2) on the segment data 114 side is “#” corresponding to the blank segment (step S1204).

  If the determination in step S1204 is NO, the CPU 1001 determines whether or not the phoneme labels on the segment data 114 side and the phoneme label on the morpheme label data 124 side match (step S1205).

  The phoneme label on the segment data 114 side is not “#” indicating a blank segment (NO in step S1204), and both the phoneme label on the segment data 114 side and the phoneme label on the morpheme label data 124 side If the phoneme label does not match and the determination in step S1205 is also NO, the voice data 113 and the utterance text 122 do not correspond well in the first place. Therefore, in such a case, the speech data 113 and the utterance text 122 input with an error should not be used for learning, and the learning label data 130 should not be output. Therefore, if the determination in step S1204 is NO and the determination in step S1205 is also NO, the CPU 1001 sets the error flag stored in the RAM 1003 to ON (step S1207), and thereafter, FIG. 12 and FIG. The association process in step S1101 of FIG. 11 shown in the flowchart of FIG. As a result, the determination in step S1102 in FIG. 11 is YES, the accent correction process in step S1103 is not executed, and the data providing process ends without generating the learning label data 130.

  If the determination in step S1205 is YES, the CPU 1001 sets both the increment flag on the segment data 114 side and the increment flag on the morpheme label data 124 side on the RAM 1003 to be on (step S1206).

  Thereafter, the CPU 1001 determines whether or not the increment flag on the segment data 114 side is ON (step S1212), but the determination in step S1212 is YES due to the processing in step S1206. As a result, the CPU 1001 sets the address of the next segment data 114 in the variable seg (step S1213). Specifically, the CPU 1001 sets the value of the next variable (see FIG. 2) of the current segment data 114 to the variable seg.

  Subsequently, the CPU 1001 determines whether or not the increment flag on the morpheme data 123 side is on (step S1214), but the determination in step S1214 is YES due to the processing in step S1206. As a result, the CPU 1001 sets the address of the next morpheme label data 124 in the variable morph (step S1215). Specifically, the CPU 1001 sets the value of the next variable (see FIG. 6) of the current morpheme label data 124 to the variable morph.

  Thereafter, the CPU 1001 proceeds to the process of step S1202 and proceeds to the next repetition process.

  As described above, when the phoneme labels on the segment data 114 side and the phoneme label on the morpheme label data 124 side match, both the variable seg and the variable morph are incremented, so that the next segment data The process shifts to a process for associating the phoneme 114 and the phoneme of the next morpheme label data 124.

  If the determination in step S1204 is YES, the CPU 1001 determines whether the phoneme label (phone in FIG. 6) on the morpheme label data 124 side is “#” corresponding to the punctuation mark (step S1208).

  If the determination in step S1208 is YES, the CPU 1001 turns on both the increment flag on the segment data 114 side and the increment flag on the morpheme label data 124 side on the RAM 1003 (step S1209).

  Thereafter, the CPU 1001 proceeds to the determination process of step S1212 described above, but the determination of step S1212 is YES due to the process of step S1209. As a result, the CPU 1001 shifts to the processing of step S1213 described above, and sets the address of the next segment data 114 in the variable seg.

  Subsequently, the CPU 1001 proceeds to the determination process of step S1214 described above, but the determination of step S1214 is YES due to the process of step S1206. As a result, the CPU 1001 proceeds to the process of step S1215 described above, and sets the address of the next morpheme data 123 in the variable morph.

  Thereafter, the CPU 1001 proceeds to the process of step S1202 and proceeds to the next repetition process.

  As described above, when the phoneme label on the segment data 114 side is “#” corresponding to the blank segment, and the phoneme label on the morpheme label data 124 side is “#” corresponding to the punctuation mark and both correspond to each other When the variable seg and the variable morph are both incremented, the next segment data 114 phoneme and the next morpheme label data 124 phoneme are associated with each other.

  If the determination in step S1208 is NO, the CPU 1001 inserts the current morpheme data 123 in the morpheme data 123 obtained on the RAM 1003 in order to insert “#” morpheme label data before the current morpheme label data. Next to the morpheme data 123 including the phoneme of the label data 124, the morpheme data 123 indicating the reading point is inserted (step S1210). Specifically, the CPU 1001 first refers to the value of the morph_id variable (see FIG. 6) indicated by the variable prev of the current morpheme label data 124 on the RAM 1003 indicated by the variable morph, and thereby from the morpheme data 123 on the RAM 1003. Data that holds the same value as the value of the morph_id variable in the morph_id variable (FIG. 4) is searched. Next, the CPU 1001 generates an entry (see FIG. 4) of new morpheme data 123 at the end of the morpheme data 123 on the RAM 1003. Then, the CPU 1001 assigns a new value to the morph_id variable in the newly generated entry at the end, stores the value “,” representing the reading point in the original variable, the read variable, and the promote variable, respectively, and the group variable “Symbol” is stored in the “accent” variable, and accent information “0” is stored in the “accent” variable. Further, the CPU 1001 stores the value of the prev variable and the value of the morph_id variable of the searched morpheme data 123 in the prev variable and the next variable in the entry newly generated at the end, respectively. Finally, the CPU 1001 stores the value of the morph_id variable of the newly created entry at the end in the next variable of the morpheme data 123 indicated by the prev variable of the searched morpheme data 123, and then the searched The value of the prev variable of the morpheme data 123 is also changed to the value of the morph_id variable of the entry newly generated at the end.

  For example, in the specific example of the segment data 114 in FIG. 3, the specific example of the morpheme data 123 in FIG. 5, and the specific example of the morpheme label data 124 in FIG. 7, the values of the variables seg and morph are both “16”. In step S1204, it is determined that the phoneme label of the segment data 114 in which the value of the seg_id variable in FIG. 6 is “16” is “#”, and the determination result is YES. In step S1208, mlabel_id in FIG. It is determined that the phoneme label of the morpheme label data 124 whose value is “16” is “s”, and the determination result is NO. As a result, by executing step S1210, the CPU 1001 refers to the value “3” of the morph_id variable of the morpheme label data 124 whose mlabel_id value is “16” in FIG. Search for the morpheme “all” whose value of the morph_id variable is “3”. Next, the CPU 1001 generates an entry (see FIG. 4) of the new morpheme data 123 at the end of the morpheme data 123 of FIG. FIG. 14 is a specific example of the morpheme data 123 after the entry is added. In the entry at the end of FIG. 14, the CPU 1001 assigns a new value “13” to the morph_id variable, stores values “,” representing reading points in the original variable, read variable, and promote variable, respectively, and sets “ “Symbol” is stored, and accent information “0” is stored in the “accent” variable. Further, the CPU 1001 sets the value “2” of the prev variable and the value “3” of the morph_id variable of the entry of the morpheme “all” whose value of the morph_id variable searched on the morpheme data 123 of FIG. 5 is “3”, respectively. , The prev variable and the next variable of the entry at the end of FIG. 14 are set. Finally, the CPU 1001 indicates the value of the morph_id variable of FIG. 5 indicated by the value “2” of the prev variable of the entry of the morpheme “all” whose value of the morph_id variable searched for on the morpheme data 123 of FIG. 5 is “3”. The value “13” of the morph_id variable of the entry at the end of FIG. 14 is stored in the next variable of the entry of the morpheme “ha” whose value is “2”, and the morph_id variable searched for on the morpheme data 123 of FIG. The value of the prev variable of the entry of the morpheme “all” whose value is “3” is also changed to the value “13” of the morph_id variable of the last entry in FIG. As a result, the morpheme data 123 shown in FIG. 14 is completed. As a result, next to the morpheme “o” whose value of the morph_id variable is “2”, the value “13” of the next variable of the entry is referred to, so that the punctuation mark “,” whose value of the morph_id variable is “13”, ”Is connected, and the value“ 3 ”of the next variable of the entry is referenced, so that the entry of“ all ”morphemes of the value“ 3 ”of the morph_id variable is connected.

  In the above-described example, in the example of the segment data 114 in FIG. 3, between the phoneme string “arayurugeNjitsuo” (notation: “every reality”) and the phoneme string “subete” (notation: “all”) Due to the occurrence of breathing in pronunciation, seg_id = 15, phoneme label phone = “o” segment data 114 and seg_id = 17, phoneme label phone = “s” segment data 114, seg_id = 16, phoneme Segment data 114 of a blank segment with label phone = “#” is generated. On the other hand, in the example of the morpheme data 123 of FIG. 5, no reading mark is detected between the morpheme “O” and the morpheme “All”. In this embodiment, in this embodiment, the CPU 1001 may insert a reading mark “,” between the morpheme “o” and the morpheme “all” on the morpheme data 123 of FIG. it can.

  After the process of step S1210 of FIG. 12 described above, the CPU 1001 turns on the increment flag on the segment data 114 side in the RAM 1003 (step S1211). On the other hand, the increment flag on the morpheme label data 124 side in the RAM 1003 remains turned off in step S1203.

  Thereafter, the CPU 1001 proceeds to the determination process of step S1212 described above, but the determination of step S1212 is YES due to the process of step S1211. As a result, the CPU 1001 shifts to the processing of step S1213 described above, and sets the address of the next segment data 114 in the variable seg.

  Subsequently, the CPU 1001 proceeds to the determination process of step S1214 described above, but the determination of step S1214 is NO by the process of step S1203. As a result, the CPU 1001 skips the processing of step S1215 described above, and keeps the value of the variable morph as the address of the current morpheme data 123.

  Thereafter, the CPU 1001 proceeds to the process of step S1202 and proceeds to the next repetition process.

  As described above, the phoneme label on the segment data 114 side is “#” corresponding to the blank segment, and the phoneme label on the morpheme label data 124 side is not “#” corresponding to the punctuation marks, but is punctuated in the morpheme data 123. Is additionally inserted, only the variable seg is incremented, and the process shifts to a process for associating the phoneme of the next segment data 114 with the phoneme of the current morpheme label data 124.

  In the specific example of the segment data 114 in FIG. 3, the specific example of the morpheme data 123 in FIGS. 5 and 14, and the specific example of the morpheme label data 124 in FIG. 7, the values of the variables seg and morph are both “16”. In addition, in step S1204, it is determined that the phoneme label of the segment data 114 in which the value of the seg_id variable of FIG. 6 is “16” is “#”, and the determination result is YES. Then, the determination in step S1208 is NO. In step S1210, punctuation marks are inserted into the morpheme data 123 as shown in FIG. Thereafter, the process proceeds from step S1211 → S1212 to YES → S1213 → S1214 from NO to S1202, so that only the value of the variable seg is the segment data 114 in which the value of the seg_id variable in FIG. 3 is “16”. It is incremented to the value “17” indicated by the next variable. As a result, the target of the next comparison process is the phoneme label “s” of the segment data 114 in which the value of the seg_id variable in FIG. 3 retrieved by the changed value “17” of the variable seg is “17”, and the variable It is found that the phoneme label “s” of the morpheme label data 124 in which the value of the mlabel_id variable in FIG.

  As a result of the above iterative processing, when it is determined that either the variable seg or the variable morph has exceeded the terminal data (NO in step S1202), the CPU 1001 executes the processing from step S1216 in FIG. To do. First, the CPU 1001 combines all the morpheme expressions stored in the original variable of each morpheme data 123 shown in FIG. 14, for example, obtained on the RAM 1003 by the processing so far into text data, and the text data The morpheme analysis process (corresponding to the morpheme analysis unit 120 in FIG. 1) is executed again, and the corresponding morpheme data 123 and morpheme label data 124 are generated on the RAM 1003 (step S1216).

  After that, the CPU 1001 sequentially determines the phoneme label on the segment data 114 side and the phoneme label on the morpheme label data 124 side from the top of the set of segment data 114 and the top of the set of morpheme label data 124 newly generated on the RAM 1003. The process of determining again whether or not (step S1220) is repeated.

  Specifically, the CPU 1001 sets the start address of the segment data 114 on the RAM 1003 and the start address of the morpheme label data 124 newly generated on the RAM 1003 in the variables seg and morph on the RAM 1003 (step S1217). .

  Next, in order to generate the learning label data 130 on the RAM 1003, the CPU 1001 sets the address of the leading empty area of the learning label data 130 on the RAM 1003 in the variable tr on the RAM 1003 (step S1218).

  While the CPU 1001 determines that neither the variable seg nor the variable morph exceeds the terminal data (when the determination in step S1219 is YES), the CPU 1001 executes a series of processing from step S1220 to step S1224 below. To do.

  First, the CPU 1001 determines whether or not the phoneme labels on the segment data 114 side and the phoneme label on the morpheme label data 124 side match (step S1220).

  By the processing of the flowchart of FIG. 12 described above, the morpheme data 123 of FIG. 14 is changed to the blank timing segment data 114 generated by the occurrence of a breath of breath at the timing corresponding to the morpheme break in the utterance of the audio data 113. Correspondingly, additional reading marks are inserted. Therefore, the phoneme string of the morpheme label data 124 generated again by executing the morpheme analysis process in step S1216 of FIG. 13 again on the text data obtained by recombining the morpheme data 123 is as follows. That is, it is well associated with the phoneme string of the segment data 114.

  On the other hand, when a blank timing occurs due to stagnation in the speech of the audio data 113, the blank timing may not be a morpheme break, but may be a halfway phoneme position. In this case, the segment data 114 at the blank timing is detected by the determination process of step S1204 of FIG. 12, the determination result is YES, and the determination result of subsequent step S1208 is NO. The morpheme data 123 is forcibly inserted at the morpheme break position with respect to the blank timing generated in the middle phoneme. In this state, when the morpheme analysis process in step S1216 of FIG. 13 is executed again, the phoneme string of the morpheme label data 124 generated again does not correspond well with the phoneme string of the segment data 114. Since the speech data 113 that originally includes staleness is not suitable for generating phoneme segment data, in such a case, the speech data 113 and the utterance text 122 that are input with an error are not used for learning. The learning label data 130 should not be output.

  Therefore, when the phoneme label of the phoneme label on the segment data 114 side and the phoneme label on the morpheme label data 124 side do not match, the determination in step S1220 is NO, the CPU 1001 is stored in the RAM 1003. Is set to ON (step S1224), and then the association process in step S1101 of FIG. 11 shown in the flowcharts of FIGS. 12 and 13 is terminated. As a result, the determination in step S1102 in FIG. 11 is YES, the accent correction process in step S1103 is not executed, and the data providing process ends without generating the learning label data 130.

  If the phoneme label on the segment data 114 side matches the phoneme label on the morpheme label data 124 side, and the determination in step S1220 is YES, the CPU 1001 stores the RAM 1003 in the RAM 1003 indicated by the variable tr on the RAM 1003. The contents of the segment data 114 indicated by the variable seg and the contents of the morpheme label data 124 indicated by the variable morph are assigned to the new area of the learning label data 130 (see FIG. 8). Specifically, the CPU 1001 copies the contents of the phone, start, and end variables of the segment data 114 (see FIG. 2) on the RAM 1003 indicated by the variable seg to the new area, and on the RAM 1003 indicated by the variable morph. The contents of each of the variables “accent” and “mola” of the morpheme label data 124 (see FIG. 5) are copied. Further, in the new area, the CPU 1001 sets a value “1” indicating a vowel to a vowel variable (see FIG. 8) when the phoneme label of the phone variable indicates a vowel phoneme, and indicates a consonant phoneme. If so, set the vowel variable to “0” to indicate the consonant Further, the CPU 1001 sets a new morpheme label ID value in the tlabel_id variable in the new area. Further, the CPU 1001 sets the morpheme label ID value (null value in the first case) of the previous learning label data 130 to the prev variable and the next variable to the next in the new area. The value of the morpheme label ID of the learning label data 130 (a null value in the case of the end) is set. The contents of the pitch variable are set in the accent correction process described later.

  After step S1221, the CPU 1001 sets the addresses of the next segment data 114 and the next morpheme label data 124 in the variables seg and morph on the RAM 1003, respectively (step S1222). Specifically, the CPU 1001 sets the value of the next variable (see FIG. 2) of the current segment data 114 and the value of the next variable (see FIG. 6) of the current morpheme label data 124 to the variables seg and morph, respectively. To do.

  Further, the CPU 1001 sets the address of the next empty area of the learning label data 130 on the RAM 1003 in the variable tr on the RAM 1003 (step S1223).

  Thereafter, the CPU 1001 proceeds to the process of step S1219 in FIG. 13 and proceeds to the next repetition process.

  As a result of the above iterative processing, if it is determined that either the variable seg or the variable morph has exceeded the terminal data (NO in step S1219), the CPU 1001 is shown in the flowcharts of FIGS. The association process in step S1101 in FIG. In this case, since the learning label data 130 is generated without an error, the determination in step S1102 in FIG. 11 is NO, and the accent correction process in step S1103 in FIG. 11 is executed.

  FIG. 15 is a flowchart showing a detailed example of the accent correction process in step S1103 of FIG. In the accent correction process, the learning label data 130 generated in the RAM 1003 by the association process in step S1101 (FIGS. 12 and 13) in FIG. 11 and the functions of the fundamental frequency analysis unit 112 in the speech analysis unit 101 in FIG. Using the fundamental frequency data 115 generated in the RAM 1003 by the fundamental frequency analysis processing corresponding to, a new accent position is generated for each accent block, and the accent information of the learning label data 130 corresponding to the segment belonging to the accent block Is corrected based on the newly calculated accent position described above.

  First, the CPU 1001 calculates, for each learning label data 130 on the RAM 1003 (see FIG. 8), an average value of fundamental frequencies within a segment section determined by the value of the start variable and the value of the end variable of the learning label data 130. Calculate (step S1501). Specifically, the CPU 1001 extracts each basic frequency data 115 in which the value of the time variable falls within the interval from the basic frequency data 115 (see FIG. 9) on the RAM 1003, and the pitch variable of the basic frequency data 115. Are extracted, and an average value of the fundamental frequencies is calculated. Then, the CPU 1001 holds the calculated average value in the pitch variable (see FIG. 8) of the learning label data 130 as the fundamental frequency of the segment interval corresponding to the learning label data 130.

  Next, the CPU 1001 sets the start address of the learning label data 130 generated on the RAM 1003 in the variable tr on the RAM 1003 in step S1502, and then the next address of the learning label data 130 in the variable tr in step S1507. Are sequentially set, until a determination is made in step S1503 that the address of the variable tr exceeds the end of the learning label data 130, a series of processes from steps S1504 to S1506 are repeatedly executed.

  In the above iterative process, the CPU 1001 first determines whether or not the segment corresponding to the learning label data 130 indicated by the variable tr is an accent block break (step S1504). Specifically, the CPU 1001 determines that the value of the mora number indicated by the mola variable in the learning label data 130 indicated by the variable tr (see FIG. 8) is one immediately before indicated by the prev variable in the learning label data 130. When the value of the mola number indicated by the mola variable in the learning label data 130 is smaller than the value of the mora number, it is determined that the accent block to which the previous segment of the learning label data 130 belongs has ended, and the determination in step S1504 Becomes YES. Alternatively, when the phone variable of the learning label data 130 indicated by the variable tr holds a phoneme label “#” indicating a blank segment, the CPU 1001 immediately before that indicated by the prev variable in the learning label data 130 It is determined that the accent block to which the segment of the learning label data 130 belongs has ended, and the determination in step S1504 is YES.

  If the determination in step S1504 is YES, the CPU 1001 determines the segment determined by the value of the start variable and the value of the end variable (see FIG. 8) in the range of the accent block to which the previous segment of the learning label data 130 belongs. Among the learning label data 130 whose section belongs to the accent block, the phoneme position (for example, the value of the start variable) of the learning label data 130 having the highest fundamental frequency held by the pitch variable (see FIG. 8) is used as the accent block. Is determined as a new accent position (mora number). When the value of the vowel variable (see FIG. 8) of the learning label data 130 is “0”, that is, when the phoneme of the segment is a consonant, the phoneme position of the learning label data 130 of the vowel phonemes before and after the segment. Is a new accent position (step S1505).

  After step S1505, the CPU 1001 determines that the segment variable determined by the start variable value and the end variable value (see FIG. 8) in the range of the accent block includes the accent variable ( The difference between the mora number held in the mola variable (see FIG. 8) and the new accent position calculated in step S1505 (a negative value before the accent position and a positive value after the value) (Step S1506).

  Thereafter, the CPU 1001 proceeds to the process of step S1507.

  When the determination in step S1504 is NO, the CPU 1001 skips the processes in steps S1505 and S1506 and proceeds to the process in step S1507.

  In step S1507, the CPU 1001 sets the next address of the learning label data 130 in the variable tr. Specifically, the CPU 1001 stores the value of the next variable of the learning label data 130 (see FIG. 8) on the RAM 1003 indicated by the variable tr before the change into the variable tr. Thereafter, the CPU 1001 returns to the process of step S1503.

  After the above iterative processing, if it is determined that the address of the variable tr has exceeded the end of the learning label data 130 and the determination in step S1503 is YES, the CPU 1001 determines in step S1103 of FIG. 11 shown in the flowchart of FIG. The learning label data 130 generated in the RAM 1003 is output to the external storage device 1006 or the like as final learning label data 130 for creating a speech database of phoneme segment data.

  As described above, in the present embodiment, correct accent information can be obtained by considering the blank timing of audio data, and correct correction of accent information based on the fundamental frequency obtained from audio data can be performed. It becomes possible.

  In the embodiment described above, the accent information of each segment in the accent block is corrected according to the position of the segment having the highest fundamental frequency in the accent block. In addition, the accent information of each segment in the accent block may be corrected according to the position of the segment having the maximum signal strength in the accent block.

  The above-described embodiment is an embodiment of the accent information creating apparatus 100 for creating the learning label data 130 used for creating the speech database of phoneme piece data. In addition to the speech analysis unit 101, the language analysis unit 102, and the data addition unit 103 shown in FIG. 1, a database registration unit is provided. Registering, in the speech database, a process for cutting out a portion to be processed as phoneme piece data, the phoneme piece data, and accent information that the data adding unit 103 can extract from the learning label data 130 together with a phoneme label representing the phoneme of the segment. The present invention may be implemented as a voice database creation device that executes processing.

Regarding the above embodiment, the following additional notes are disclosed.
(Appendix 1)
An acquisition process for acquiring first position data indicating at least one of an accent position and a break position from input speech data, and morpheme data including a plurality of morphemes generated from text data corresponding to the speech data A comparison process comparing the second position data indicating at least one of the position of the accent and the separation position between the plurality of morphemes and the first position data acquired from the audio data; If the first and second position data do not match, the voice information includes a processing unit that executes a process of assigning the first position data to the morpheme data instead of the second position data. Creation device.
(Appendix 2)
The voice information creation device further includes:
The speech information creation device according to attachment 1, further comprising a morpheme analysis unit that generates morpheme data including the plurality of morphemes by executing a morpheme analysis process on the text data.
(Appendix 3)
The audio information creation device according to appendix 1 or 2, wherein the processing unit executes, as the acquisition process, a process of acquiring a position of a silent section of the audio data as a break position of the first position data.
(Appendix 4)
In the comparison process, when the delimiter position indicated by the first position data coincides with the delimiter position indicated by the second position data, the processing unit further determines that the second position data is relative to the morpheme data. 4. The audio information creation device according to any one of appendices 1 to 3, which adds reading point information to the indicated position.
(Appendix 5)
The processing unit includes, as the acquisition process, a process of determining a fundamental frequency of the speech data within a section of the speech data corresponding to each of the plurality of morphemes, and a highest fundamental frequency within the section of the speech data. The audio information creation device according to any one of appendices 1 to 4, which executes a process of acquiring a high position as an accent position of the first position data.
(Appendix 6)
The processing unit includes, as the acquisition process, a process of determining a signal strength of the voice data within a section of the voice data corresponding to each of the plurality of morphemes, and a highest signal strength within the section of the voice data. The audio information creation device according to any one of appendices 1 to 4, which executes a process of acquiring a high position as an accent position of the first position data.
(Appendix 7)
A speech information creation method used in a speech information creation apparatus including a processing unit, wherein the processing unit is
Obtaining first position data indicating at least one of an accent position and a break position from the input voice data;
Second position data indicating at least one of a position of an accent given to morpheme data including a plurality of morphemes generated from text data corresponding to the sound data and a delimiter position between the plurality of morphemes; and the sound data And the first position data obtained from
When the first and second position data do not coincide with each other, the voice information creating method of adding the first position data to the morpheme data instead of the second position data.
(Appendix 8)
In a computer used as a voice information creation device,
Obtaining first position data indicating at least one of an accent position and a break position from input voice data;
Second position data indicating at least one of a position of an accent given to morpheme data including a plurality of morphemes generated from text data corresponding to the sound data and a delimiter position between the plurality of morphemes; and the sound data Comparing the first position data obtained from
When the first and second position data do not match, the first position data is given to the morpheme data instead of the second position data;
A program that executes
(Appendix 9)
The audio information creation device according to any one of appendices 1 to 6;
A speech database that extracts the phoneme data from the speech data, the phoneme data, the phoneme label that represents the phoneme data, and the accent information given to the morpheme corresponding to the phoneme data by the speech information creation device A registration processing unit for executing, a registration processing unit for executing
Voice database creation device with

DESCRIPTION OF SYMBOLS 100 Accent information production apparatus 101 Speech analysis part 102 Language analysis part 103 Data provision part 110 Speech recognition part 111 Acoustic model 112 Fundamental frequency analysis part 120 Morphological analysis part 121 Morphological dictionary 122 Utterance text 1001 CPU
1002 ROM (Read Only Memory)
1003 RAM (Random Access Memory)
1004 Input device 1005 Output device 1006 External storage device 1007 Portable recording medium driving device 1008 Communication interface 1009 Bus 1010 Portable recording medium

Claims (9)

  An acquisition process for acquiring first position data indicating at least one of an accent position and a break position from input speech data, and morpheme data including a plurality of morphemes generated from text data corresponding to the speech data A comparison process comparing the second position data indicating at least one of the position of the accent and the separation position between the plurality of morphemes and the first position data acquired from the audio data; If the first and second position data do not match, the voice information includes a processing unit that executes a process of assigning the first position data to the morpheme data instead of the second position data. Creation device. The voice information creation device further includes:
The speech information creation device according to claim 1, further comprising: a morpheme analysis unit that generates morpheme data including the plurality of morphemes by executing a morpheme analysis process on the text data.   The audio information creation device according to claim 1, wherein the processing unit executes a process of acquiring a position of a silent section of the audio data as a delimiter position of the first position data as the acquisition process.   In the comparison process, when the delimiter position indicated by the first position data coincides with the delimiter position indicated by the second position data, the processing unit further determines that the second position data is relative to the morpheme data. The voice information creation device according to claim 1, wherein reading point information is assigned to the indicated position.   The processing unit includes, as the acquisition process, a process of determining a fundamental frequency of the speech data within a section of the speech data corresponding to each of the plurality of morphemes, and a highest fundamental frequency within the section of the speech data. The voice information creation device according to claim 1, wherein a process for acquiring a high position as an accent position of the first position data is executed.   The processing unit includes, as the acquisition process, a process of determining a signal strength of the voice data within a section of the voice data corresponding to each of the plurality of morphemes, and a highest signal strength within the section of the voice data. The voice information creation device according to claim 1, wherein a process for acquiring a high position as an accent position of the first position data is executed. A speech information creation method used in a speech information creation apparatus including a processing unit, wherein the processing unit is
Obtaining first position data indicating at least one of an accent position and a break position from the input voice data;
Second position data indicating at least one of a position of an accent given to morpheme data including a plurality of morphemes generated from text data corresponding to the sound data and a delimiter position between the plurality of morphemes; and the sound data And the first position data obtained from
When the first and second position data do not coincide with each other, the voice information creating method of adding the first position data to the morpheme data instead of the second position data. In a computer used as a voice information creation device,
Obtaining first position data indicating at least one of an accent position and a break position from input voice data;
Second position data indicating at least one of a position of an accent given to morpheme data including a plurality of morphemes generated from text data corresponding to the sound data and a delimiter position between the plurality of morphemes; and the sound data Comparing the first position data obtained from
When the first and second position data do not match, the first position data is given to the morpheme data instead of the second position data;
A program that executes The voice information creation device according to any one of claims 1 to 6,
A speech database that extracts the phoneme data from the speech data, the phoneme data, the phoneme label that represents the phoneme data, and the accent information given to the morpheme corresponding to the phoneme data by the speech information creation device A registration processing unit for executing, a registration processing unit for executing
Voice database creation device with