JP2011186396A - Speech recording device, speech recording method and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a speech recording device, a speech recording method and a program, capable of precisely coordinating a speech with additional information for speech retrieval. <P>SOLUTION: The device includes: a recording sentence acquiring section 11 for acquiring a recording sentence T for expressing speech to be recorded, in a sentence unit; a recording sentence dividing section 12 for dividing the recording sentence into exhalation paragraph units by language analysis processing; a recording sentence dividing section 13 for dividing speech corresponding to the recording sentence in the exhalation paragraph unit; a start/end detecting section 14 for detecting a start/end point of speech from a recording result of voice; a silence period length calculation section 15 for calculating a silence period length between exhalation paragraph units included in the recording sentence; a data creating section 16 for creating the speech data corresponding to the recording sentence, and the additional information used for retrieval of the speech, from a recording result of the speech, a detection result of the start/end point, and a calculation result of the silence period; and a data storage section 17 for storing the speech data corresponding to the recording sentence and the additional information in the sentence unit. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to an audio recording device, an audio recording method, and a program.

  Voice utilization systems using recorded voice, such as voice guidance systems and corpus-based voice synthesis systems, have become widespread. In these systems, recorded voices are stored (in a database), and desired voices are generated by searching, connecting, and processing signals in an appropriate voice section when using the voices.

  For this reason, the voice is stored in association with the additional information for voice search. In general, in order to define a speech section, additional information includes language information and acoustic information representing speech, or identification information defined based on these. The additional information is, for example, a voice text notation, a symbol representing a voice smear or accent position, voice identification information, section information of phonemes in the voice, and the like. The additional information is described in, for example, Patent Document 1 below. As a result, the utility value of the voice utilization system is evaluated from the three viewpoints of recorded voice quality, additional information quality, and database construction cost.

JP 2001-34283 A

  The additional information is used for various purposes such as the definition of a voice section, statistical analysis of acoustic parameters, construction of an acoustic model, and the like, and when used, it is necessary to ensure the quality to be satisfied according to the purpose.

  As an example for ensuring the quality of the additional information, in the phoneme labeling process used for constructing a speech corpus, the additional information is manually corrected. Also, in speech analysis of utterances with irregular “between”, inflections, speed, utterances with speech and restatement, especially free utterances without the recording engineer, speech sentence units and phrases (exhalation paragraph) In many cases, the accuracy of associating voice and additional information in units cannot be ensured. These problems cause a rise in database construction costs and a decrease in quality, which in turn prevent the spread of voice utilization systems.

  In order to solve these problems, it is possible to simply shorten the audio unit to be recorded and record it in the exhalation paragraph unit or word unit to improve the accuracy of the additional information. However, in general, it is necessary to generate prosodic information for each sentence in order to construct a speech utilization system that supports sentence-by-sentence input. Therefore, recorded voices are also required to be collected and databased in sentence units. The

  Accordingly, the present invention is intended to provide a voice recording device, a voice recording method, and a program capable of associating voice and additional information for voice search with high accuracy.

  According to an aspect of the present invention, a recorded sentence acquisition unit that acquires a recorded sentence representing speech to be recorded in sentence units, a recorded sentence division unit that divides the recorded sentence into exhalation paragraph units by language analysis processing, and a recorded sentence A voice recording unit that records the speech corresponding to each exhalation paragraph, a start / end detection unit that detects the start / end time of the speech from the audio recording results, and a silence interval length between the exhalation paragraphs included in the recorded sentence is calculated A silent section length calculation unit, a voice generation result corresponding to a recorded sentence, and a data generation unit that generates additional information used for searching for the voice from the recording result of the voice, the detection result of the start and end points, and the calculation result of the silent section There is provided an audio recording device including audio data corresponding to a recorded sentence and a data storage unit that stores additional information in units of sentences.

  According to such a configuration, the voice corresponding to the recorded sentence is recorded in units of expiratory paragraphs, the start and end points of the voice are detected, and the silent section length between the expiratory paragraphs is calculated. Then, voice data and additional information corresponding to the recorded sentence are generated from the voice recording result, the detection result at the start and end points, and the calculation result of the silent section, and stored in sentence units. This makes it possible to accurately detect the start and end points of speech and to associate speech with additional information with high accuracy. As a result, it is possible to provide a voice utilization system that is excellent in recorded voice quality, additional information quality, and database construction cost.

  The data generation unit may generate at least one of a start / end time point of the uttered voice, a morphological analysis result of the recorded sentence, and a silent section length between the following exhalation paragraph as additional information.

  The recorded sentence dividing unit may divide the recorded sentence into exhalation paragraph units by a language analysis method capable of predicting voice boundary information and pause positions.

  The voice recording unit may record a voice uttered in accordance with a guidance indicating an exhalation paragraph unit.

  The silence interval length calculation unit may calculate a silence interval length between exhalation paragraphs based on a result of dividing the recorded sentence.

  The silence interval length calculation unit may calculate a silence interval length between exhalation paragraphs based on a recording result of sound corresponding to the recorded sentence.

  Further, according to another aspect of the present invention, a step of acquiring recorded sentences representing speech to be recorded in units of sentences, a step of dividing the recorded sentences into units of expiratory paragraphs by language analysis processing, and corresponding to the recorded sentences A step of recording speech in units of expiratory paragraphs, a step of detecting the start and end points of speech from the recording results of speech, a step of calculating a silent interval length between exhalation paragraphs included in the recorded sentence, a recording result of speech and Speech including speech data corresponding to recorded sentences, additional information used for speech search, and storing speech data and additional information corresponding to the recorded sentences in sentence units from the calculation result of the silent section A recording method is provided.

  Moreover, according to another viewpoint of this invention, the program for making a computer perform the said audio | voice recording method is provided. Here, the program may be provided using a computer-readable recording medium or may be provided via communication means.

  As described above, according to the present invention, it is possible to provide a voice recording device, a voice recording method, and a program capable of associating voice and additional information for voice search with high accuracy.

It is a block diagram which shows the main function structures of the audio | voice recording apparatus which concerns on embodiment of this invention. It is a flowchart which shows the operation | movement procedure of an audio | voice recording apparatus. It is a figure which illustrates the division | segmentation result of the collection text in a breath paragraph unit. It is a figure which illustrates the guidance notified to a speaker at the time of recording of speech sound. It is a figure which illustrates the production | generation result of audio | voice data and additional information. It is a figure which illustrates the process result of an acoustic analysis process. It is a figure which illustrates the process result of an audio | voice extraction process.

  Exemplary embodiments of the present invention will be described below in detail with reference to the accompanying drawings. In addition, in this specification and drawing, about the component which has the substantially same function structure, duplication description is abbreviate | omitted by attaching | subjecting the same code | symbol.

  FIG. 7 shows the processing result of the sound extraction process. In the voice extraction process, the voice is program-analyzed based on the recorded voice and the recorded sentence T, the voice section is estimated from the phoneme sequence acoustic information, and the voice section and the language information are associated on the time axis.

  FIG. 7 shows a case where the speech segment Sva corresponding to “June 8th” is cut out from the speech segment Sv corresponding to the continuous utterance speech “June 8th, Wednesday”. Continuous speech sounds are recorded as shown in FIG. In FIG. 7B, the voice section Sv (including voice sections Sva and Svb) is cut out, and the cut-out process has failed. On the other hand, in FIG.7 (c), audio | voice area Sva is cut out and the cut-out process was successful.

  Here, the cause of the failure is, for example, misdetection of the start and end points of speech sound due to large stationary noise or pulse-like sudden noise, fitting with an inappropriate acoustic model, speech speech due to speech / rephrase, etc. For example, the divergence of the sentence T can be cited. In particular, when the start and end points of the speech voice cannot be detected with high accuracy, the voice section cannot be appropriately cut out. However, with the current technology, it is considered difficult to automatically cut out the uttered voice with high accuracy in consideration of all these causes.

[1. Configuration of audio recording apparatus 10]
FIG. 1 shows a functional configuration of an audio recording apparatus 10 according to the embodiment of the present invention. As shown in FIG. 1, the voice recording device 10 includes a recorded sentence acquisition unit 11, a recorded sentence division unit 12, a voice recording unit 13, a start / end detection unit 14, a silent section length calculation unit 15, a data generation unit 16, and a data storage. The unit 17, the database 18, and the control unit 19 are included.

  The recorded sentence acquisition unit 11 acquires a recorded sentence T representing the uttered voice to be recorded in sentence units. The recorded sentence T is acquired as text data or the like and supplied to the recorded sentence division unit 12. The recorded sentence division unit 12 divides the recorded sentence T into units of expiratory paragraphs by language analysis processing. The divided recorded sentence T is supplied to the voice recording unit 13 and the silent section length calculation unit 15 as text data or the like.

  Here, the exhalation paragraph means a prosodic phrase, and the exhalation paragraph break means a boundary of the prosodic phrase. When the sentence is spoken, the time-varying pattern of the fundamental frequency is a local undulation (accent phrase) that is formed as a group of about one or two phrases, and a global change that gradually falls with time from the start of the utterance. It is expressed by At the boundary of many adjacent accent phrases, a global descending characteristic is maintained, and prosodic phrases that are the same prosodic unit are formed. On the other hand, at the boundary between some adjacent accent phrases, a prosodic phrase boundary is generated in which the fundamental frequency is reestablished without decreasing.

  The voice recording unit 13 records the uttered voice corresponding to the recorded sentence T in units of exhalation paragraphs through the microphone 13a and the like. The recorded result of the uttered voice is supplied to the start / end detection unit 14. The start / end detection unit 14 detects the start / end point of the uttered voice in units of expiratory paragraphs. The detection result at the start / end time is supplied to the data generation unit 16.

  The silent section length calculation unit 15 calculates the length of the silent section Ss between exhalation paragraphs included in the recorded sentence T. The calculation result of the length of the silent section Ss is supplied to the data generation unit 16. The silent section Ss is also called a pause or a short pause, and means “between” by conscious or unconscious breathing at the time of speech.

  The data generation unit 16 generates voice data and additional information corresponding to the recorded sentence T from the voice recording result, the detection result of the start and end points, and the calculation result of the silent section Ss. The generated voice data and additional information are supplied to the data storage unit 17 in sentence units.

  Specifically, the voice data is generated by generating a combination of a voice waveform and an acoustic waveform for each exhalation paragraph, and further combining all exhalation paragraph combinations included in the recorded sentence T. The additional information is, for example, the start and end points of the uttered voice, the morphological analysis result of the recorded sentence T, the length of the silent section Ss between the following exhalation paragraphs, and the like.

  The data storage unit 17 stores voice data and additional information in the database 18 in sentence units. The control unit 19 performs arithmetic processing and control processing necessary for operating the audio recording device 10.

  Note that at least a part of the above configuration may be realized by software (program) operating on the audio recording device 10 or may be realized by hardware. When implemented by software, the program may be stored in advance on the audio recording device 10 or supplied from the outside.

[2. Operation of the audio recording device 10]
FIG. 2 shows an operation procedure of the audio recording apparatus 10. The voice recording device 10 records uttered voices and stores voice data for each sentence corresponding to the recorded sentence T.

  As shown in FIG. 2, the recorded sentence acquisition part 11 acquires the recorded sentence T showing the speech sound which should be recorded per sentence (step S11). The recorded sentence T may be fetched in units of sentences from a recorded sentence set stored in a storage device (not shown), may be received through a network (not shown), or through an input device (not shown) such as a keyboard. It may be entered.

The recorded sentence dividing unit 12 divides the recorded sentence T into units of expiratory paragraphs by language analysis processing (S12). For the language analysis processing, an analysis method capable of predicting voice boundary information and pause positions is used. As an analysis method, for example, a method described in the following document may be used.
"Prediction of prosodic phrase boundaries and pause positions using probabilistic context-free grammar", Fujio et al., IEICE Transactions D-II, Vol. J80-D-II, no. 1, pp. 18-25, January 1997

  In general, a boundary having a pause length exceeding a predetermined threshold among the boundaries of adjacent accent phrases can be regarded as a breath paragraph break. For this reason, the recorded sentence T is analyzed, the pose length between accent phrases is predicted, the boundary with a pose length exceeding a predetermined threshold is regarded as an exhalation paragraph break, and the recorded sentence T is divided into exhalation paragraph units. .

  FIG. 3 shows a division result of the recorded sentence T in the exhalation paragraph unit. In FIG. 3, the recorded sentence T “All the reality is twisted toward you” is divided into first to third exhalation paragraphs.

  In FIG. 3, first, the recorded sentence T is “everything”, “reality”, “all”, “my”, “toward”, “twisted” based on the reading and accent position of the recorded sentence T. To accent phrases. Next, the recorded sentence T is divided into first to third exhalation paragraphs of “every reality”, “all to yourself” and “twisted” based on the pose length of the accent phrase boundary. The For example, in FIG. 3, a pause length (Breath Paragraph Pause) of 0.35 seconds is predicted between the first and second expiratory paragraphs, and a pause length of 0.15 seconds is predicted between the second and third expiratory paragraphs. ing.

  The voice recording unit 13 records the uttered voice corresponding to the recorded sentence T through the microphone 13a or the like in units of exhalation paragraphs (S13). The voice recording unit 13 visually or audibly guides the utterance by the speaker S when recording the uttered voice.

  FIG. 4 shows the guidance G that is notified to the speaker S when the utterance voice is recorded. In FIG. 4, the recorded sentence T is divided into first to third exhalation paragraphs and is shown together with a reading of the recorded sentence T. FIG. 4 (a) shows guidance G1 that prompts utterance in the voice section corresponding to the first exhalation paragraph, and FIG. 4 (b) shows guidance that prompts utterance in the voice section corresponding to the second exhalation paragraph. G2 is shown.

  As shown in FIG. 4, the speaker S is instructed to record as one take for each exhalation paragraph and to utter a speech segment corresponding to the exhalation paragraph at a breath. As a result, it is possible to detect the exhalation paragraph break with high accuracy in the recorded uttered speech, and the subsequent acoustic analysis processing result can be obtained with high accuracy.

  The start / end detection unit 14 detects the start / end time of the uttered voice in units of expiratory paragraphs (S14). The start time point is detected as the time point when the power time series of the uttered voice exceeds the predetermined threshold from the start of the utterance, and the end time point is detected as the time point when the power time series exceeds the threshold value after the end of the utterance.

  Further, the start / end time may be detected by extracting a plurality of start / end time candidates on the power time series. In this case, an acoustic analysis process such as a phoneme labeling process is performed on the uttered speech, a sequence of acoustic feature quantities is evaluated, and a candidate that can obtain an optimal matching is adopted as the start / end point.

  The control unit 19 determines whether the utterance voice is recorded for all the exhalation paragraphs included in the recorded sentence T (S15). When the recording is completed, the process proceeds to the subsequent process. When the recording is not completed, the process returns to the process of step S13.

  The silent section length calculation unit 15 calculates the length of the silent section Ss between the exhalation paragraphs included in the recorded sentence T (S16). The length of the silent section Ss is calculated as a time length corresponding to the pose length between accent phrases calculated in the process of step S12. Further, the length of the silent section Ss may be calculated based on the actually recorded speech sound. In this case, the length of the silent section Ss is calculated by actually recording the uttered voice corresponding to the recorded sentence T and detecting the silent section Ss in the recorded uttered voice.

  The data generation unit 16 generates voice data corresponding to the recorded sentence T in sentence units from the voice recording result, the detection result at the start and end points, and the calculation result of the silent section Ss (S17). The data generation unit 16 combines the sound waveform corresponding to the length of the silent section Ss at the end of the speech waveform of the expiratory paragraph, and combines the sound waveform of the next expiratory paragraph at the end of the soundless shape. The data storage unit 17 processes all the exhalation paragraphs included in the recorded sentence T in the same manner, and generates voice data corresponding to the recorded sentence T.

  The data generation unit 16 also generates additional information of the voice data in units of sentences (S17). For each expiratory paragraph included in the recorded sentence T, the data generating unit 16 represents, for example, the start and end times of the uttered speech, the morphological analysis result of the recorded sentence T, and the length of the silent section Ss between the following expiratory paragraphs. Generate data.

  FIG. 5 shows the generation result of the audio data and additional information. FIG. 5 shows audio data and additional information of the recorded sentence T shown in FIG.

  The voice data includes voice waveforms of voice sections Sv1 to Sv3 corresponding to the first to third exhalation paragraphs, and silent sections Ss1 and Ss2 that make boundaries between the voice sections Sv1 to Sv3. Specifically, a silence interval Ss1 of 0.35 seconds is coupled to the end of the first speech segment Sv1, and a speech segment Sv2 is coupled to the end of the silence segment Ss1. Further, a silent section Ss2 of 0.15 seconds is coupled to the end of the speech section Sv2, and a speech section Sv3 is coupled to the end of the silent section Ss2.

  Further, the additional information includes the morphological analysis results of the recorded sentence T at the start and end points of the speech sections Sv1 to Sv3. Specifically, the start and end times of the voice section Sv1 are 0.00 seconds and 0.90 seconds, the start and end times of the voice section Sv2 are 1.25 seconds and 2.25 seconds, and the start and end times of the voice section Sv3 are 2. 40 seconds and 3.30 seconds. Further, the morphological analysis results of the speech sections Sv1 to Sv3 are “a / ra / yu / ru / ge / n / ji / tsu / o”, “su / be / te / ji / bu / n / no / ho”. / O / e "and" ne / ji / ma / ge / ta / no / da ".

  The data storage unit 17 stores voice data and additional information in the database 18 in sentence units (S18). The control unit 19 determines whether the utterance voice is recorded for all the recorded sentences T to be recorded (S19). Then, when the recording is completed, the process is terminated. When the recording is not completed, the process returns to the process of step S11, and the next recorded sentence T is acquired.

  FIG. 6 shows a processing result of an acoustic analysis process such as a phoneme labeling process. FIG. 6 shows a processing result of the phoneme labeling process for the speech voice “June 8th”. FIG. 6A shows a processing result based on the prior art, and FIG. 6B shows a processing result by the audio recording device 10.

  In FIG. 6A, the speech section Sv corresponding to the reading “Rokugatsuyo” of the speech voice “June 8th” is detected as the speech section corresponding to “Rokugatsuyo”. In other words, since the start / end time of the uttered voice is erroneously detected, an appropriate labeling process corresponding to the utterance of the uttered voice is not performed.

  On the other hand, in FIG. 6B, the voice section Sv is detected as a voice section corresponding to “Rokugatsu Yoka”. In other words, since the start and end points of the uttered speech are appropriately detected, an appropriate labeling process corresponding to the utterance of the uttered speech is performed.

  As described above, since the voice recording apparatus 10 can accurately detect the start and end points of the speech voice by recording the speech voice for each exhalation paragraph, the speech voice and the additional information for voice search are associated with high accuracy. be able to.

  In addition, since the voice recording device 10 can keep deterioration in sound quality due to noise or incomplete speech contained in the exhalation paragraph and erroneous detection at the start and end points in the exhalation paragraph, the use efficiency of the entire recorded speech is improved. Can be made.

  As described above, according to the audio recording device 10 according to the embodiment of the present invention, the audio corresponding to the recorded sentence T is recorded in units of exhalation paragraphs, the start and end points of the audio are detected, and silence between exhalation paragraphs is detected. The length of the section Ss is calculated. Then, voice data and additional information corresponding to the recorded sentence T are generated from the voice recording result, the detection result of the start and end points, and the calculation result of the silent section Ss and stored in sentence units. This makes it possible to accurately detect the start and end points of speech and to associate speech with additional information with high accuracy. As a result, it is possible to provide a voice utilization system that is excellent in recorded voice quality, additional information quality, and database construction cost.

  The preferred embodiments of the present invention have been described in detail above with reference to the accompanying drawings, but the present invention is not limited to such examples. It is obvious that a person having ordinary knowledge in the technical field to which the present invention pertains can come up with various changes or modifications within the scope of the technical idea described in the claims. Of course, it is understood that these also belong to the technical scope of the present invention.

  For example, in the above description, the case where the recorded sentence is in Japanese has been described, but the recorded sentence may be in a language other than Japanese.

DESCRIPTION OF SYMBOLS 10 Voice recording device 11 Recorded sentence acquisition part 12 Recorded sentence division part 13 Voice recording part 14 Start / end detection part 15 Silent section length calculation part 16 Data generation part 17 Data storage part 18 Database 19 Control part T Recorded sentence S Speaker Sv Voice Section Ss Silent section

Claims (8)

A recorded sentence acquisition unit that acquires recorded sentences representing the sound to be recorded in sentence units,
A recorded sentence dividing unit that divides the recorded sentence into exhalation paragraph units by language analysis processing;
An audio recording unit that records audio corresponding to the recorded sentence in units of the exhalation paragraph;
A start / end detection unit for detecting a start / end time of the sound from the recording result of the sound;
A silent interval length calculation unit for calculating a silent interval length between the exhalation paragraphs included in the recorded sentence;
From the recording result of the voice, the detection result of the start and end points, and the calculation result of the silent section, the voice data corresponding to the recorded sentence, and a data generation unit that generates additional information used for searching the voice,
An audio recording apparatus comprising: the audio data corresponding to the recorded sentence and the data storage unit that stores the additional information in sentence units.   2. The data generation unit according to claim 1, wherein the data generation unit generates at least one of a start and end point of the voice, a morphological analysis result of the recorded sentence, and a silent section length between a subsequent exhalation paragraph as the additional information. Audio recording device.   The voice recording apparatus according to claim 1, wherein the recorded sentence dividing unit divides the recorded sentence into units of the exhalation paragraph by a language analysis method capable of predicting boundary information and pause positions of the voice.   The audio recording apparatus according to claim 1, wherein the audio recording unit records the audio uttered in accordance with guidance indicating the exhalation paragraph unit.   The sound recording device according to claim 1, wherein the silent section length calculation unit calculates a silent section length between the exhalation paragraphs based on a division result of the recorded sentence.   The voice recording device according to claim 1, wherein the silent section length calculation unit calculates a silent section length between the exhalation paragraphs based on a recording result of the voice corresponding to the recorded sentence. . A step of acquiring sentence-by-sentence sentences representing voice to be recorded;
Dividing the recorded sentences into exhalation paragraph units by language analysis processing;
Recording audio corresponding to the recorded sentence in units of the exhalation paragraph;
Detecting the start and end time of the sound from the recording result of the sound;
Calculating a silent section length between the exhalation paragraphs included in the recorded sentence;
Generating voice data corresponding to the recorded sentence and additional information used for searching the voice from the voice recording result, the detection result of the start and end points, and the calculation result of the silent section;
Storing voice data corresponding to the recorded sentence and the additional information additional information in sentence units. A step of acquiring sentence-by-sentence sentences representing voice to be recorded;
Dividing the recorded sentences into exhalation paragraph units by language analysis processing;
Recording audio corresponding to the recorded sentence in units of the exhalation paragraph;
Detecting the start and end time of the sound from the recording result of the sound;
Calculating a silent section length between the exhalation paragraphs included in the recorded sentence;
Generating voice data corresponding to the recorded sentence and additional information used for searching the voice from the voice recording result, the detection result of the start and end points, and the calculation result of the silent section;
A program for causing a computer to execute an audio recording method including the step of storing audio data corresponding to the recorded sentence and the additional information additional information in sentence units.

Images