JP4314376B2 - Writing support device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To facilitate hearing and dictation by applying processes such as proper division, speed reduction if necessary, and repetition to a program voice signal by using detected pauses. <P>SOLUTION: A dictation assisting device which assists a dictation of text data from a speech while the speech is heard is equipped with a pause detection part 1 which detects pauses of an input speech signal and a division part 11 which divides the input speech signal by a time or the number of characters enabling a dictator to memorize the divided signal for a short period on the basis of the detected pauses. <P>COPYRIGHT: (C)2004,JPO&amp;NCIPI

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a writing support apparatus that supports writing of program audio or the like using detection of a pause portion.
[0002]
[Summary of Invention]
The present invention relates to a device for supporting a book cause the program audio, such as using detection of pause portion, when causing write this particular hearing program audio, appropriate to utilize the detected pose Division is performed, and the program audio signal is subjected to processing such as speed reduction or repetition as necessary to assist easy listening and writing.
[0003]
By performing this process, it is possible to reduce the tension and fatigue associated with the time and work required for writing. In particular, when this processing method is applied to the transcription of subtitle text, it is possible to shorten the production time of subtitle programs and to increase the number of transcription workers by simplifying the work, thereby reducing the cost and cost of subtitle broadcast programs. Contribute.
[0004]
[Prior art]
When a receiver uses a TV broadcast program with subtitles, it is important that the subtitles are easy to read and understand. For this reason, in the production of subtitle manuscripts in the production of subtitle programs, skilled manpower is used to produce subtitles that are easier to read and understand, taking much effort and time.
[0005]
[Problems to be solved by the invention]
However, in the subtitle broadcasting that is expanding the application program field and the number of programs in the future, this type of subtitle program production system that requires a lot of labor and time is a big bottleneck in the production of subtitle programs. The improvement is urgent.
[0006]
The most popular subtitle program production format is a program tape that superimposes the time code into video and, if necessary, a program script as a material. A summary of the speech was created, 2) a caption display image (according to the procedure for preparing caption manuscripts specified separately), and 3) the start and end time codes were entered, and a caption manuscript was created. Based on this subtitle manuscript, the operator creates digitized subtitle data, and previews and proofreads in the presence of the person in charge of subtitle production, the manuscript creator, and the digitized operator to obtain the completed subtitle.
[0007]
The most time required for this work is to listen to the program speech of 1) and write and create a subtitle manuscript, which is mostly based on human intelligence and manual work.
[0008]
In a specific example of a program speech transcribe operation, a program tape is played and listened to by voice, and the tape is reproduced and listened from the speech start point in the voice, and written by a word processor or writing. Actually, in order to confirm the writing speed and contents of the writing operator, the writing operation is performed by repeating the cueing / playing operation of the recording tape for a single speech segment.
[0009]
Therefore, transcription work is a burdensome task for human intelligence such as cumbersome tape operations such as cueing and replaying of tapes, listening confirmation of speech, and transcription. Appropriate support functions that can be reduced are necessary, and their development is desired.
[0010]
In view of the above circumstances, the present invention performs appropriate division using the detected pose, and performs processing such as slowing down or repetition as necessary on the program audio signal so that it can be easily heard and written. An object of the present invention is to provide a transcription support device that can support.
[0011]
[Means for Solving the Problems]
In order to achieve the above object, the invention of claim 1 is a writing support apparatus for supporting a work of writing text data from a voice while listening to the voice, and a pause for detecting a pause of an input voice signal. a detecting unit, based on the detected pose, and a division unit for dividing the input audio signal with the number of characters converted from the time capable of short-term memory book cause operator or the time, the dividing unit Among the poses detected by the pose detection unit, a primary division unit that divides the input audio signal by a pose having the highest reliability, and a result of division by the primary division unit enables short-term storage. Secondary division means for dividing the input voice signal by a highly reliable pause for a portion that is longer than a predetermined time or the number of characters converted from the time, and the second division means As a result of the division, the input audio signal is divided by the next highly reliable pose for a portion that is longer than a predetermined time capable of short-term memory or more than the number of characters converted from the time, And tertiary dividing means for mechanically dividing the determined time or the number of characters converted from the time .
[0014]
According to a second aspect of the present invention, in the transcription support apparatus according to the first aspect, the time for which the short-term memory is possible is approximately 2.5 seconds, or is converted from the period for which the short-term memory is possible. The number of characters is about 25 mora in number of mora.
[0015]
According to a third aspect of the present invention, in the transcription support apparatus according to the first or second aspect, the division unit overlaps the front and rear portions of the divided voice by a predetermined time or the number of characters converted from the time. It is characterized by including a means.
[0016]
According to a fourth aspect of the present invention, in the transcription support apparatus according to any one of the first to third aspects, the pause detection unit receives a normal speed from a level signal indicating an amplitude level or a power level of the input audio signal. A specific frequency component extracting means for extracting a specific frequency component indicating the fluctuation characteristics of speech in the signal, obtaining an envelope waveform of the extracted specific frequency component, setting a predetermined slice level for the obtained envelope waveform signal, and posing It is characterized by comprising pause section detecting means for detecting a section.
[0017]
DETAILED DESCRIPTION OF THE INVENTION
<Background of the present invention>
As described above, the transcription of program audio is a task that requires a lot of burden on human intelligence, such as complicated tape operations such as cueing and repeated playback of tapes, listening confirmation of speech, and writing. It is necessary to have an appropriate support function that can reduce abilities and labor.
[0018]
For example, regarding the “one-segment speech segment” and its “repetition”, if this is a single sentence, there is a normal pose part before and after that, so this pose is detected and a segment is automatically assigned, and the required number of times Process the program audio to repeat. As a result, the speech section is easy to hear and write.
[0019]
On the other hand, non-speech sections that do not require writing can be automatically removed, so that the writing worker can concentrate on writing the speech part, greatly contributing to the accurate and rapid creation of subtitle text. can do.
[0020]
Moreover, since this process also contributes to a reduction in necessary capacity and labor, there is a possibility that the writing worker can be expanded to a general word processor worker level.
[0021]
Note that pose detection has been proposed in Japanese Patent Application No. 2002-160255 “Speech Speech / Pause Interval Detection Device”, and the pose detected by such a method (this method will be described later as the first method) is the main issue. It can be applied to processing such as dividing program audio.
[0022]
《Program audio division processing method》
When the significance of the “one-line speech segment” is further examined, the setting of an appropriate division unit for transcription and the speed-down processing performed as necessary, the writing of the speech that becomes this unit within this division time is performed. However, it is possible to complete the work, and the same applies to the division and repetition of the required number of times.
[0023]
First, in terms of ease of listening, it is reasonable to divide by using a place where there is no speech in the program audio, or where there is a break (such as a pause for transcription) due to breathing.
[0024]
The length of division will be described later, but if this processing does not fit in the required division length, it is necessary to consider additional division processing.
[0025]
Therefore, it is divided first with the detected relatively long pose (for example, 1 to 2 seconds), the part where the division is insufficient is divided with the shorter pose (for example, 0.5 to 1 second), and the part where the division is still insufficient is the other The method (shorter pose, predetermined time, etc.).
[0026]
《Consideration about the length of division for transcription》
As for the length of the division for transcription, it is desirable to have a length that allows short-term memory, including unknown words and proper nouns, but it is also valid from simple experiments.
[0027]
Although the specific image of the length that can be memorized in a short term is unambiguous, there are examples of haiku, waka, and slogans as familiar examples. Since these are 17 characters (mora) and 31 characters (mora) in katakana (mora), they were examined with a goal of about 25 characters (mora). This is about 19 characters in a kana-kanji mixed sentence with the ratio of kanji being 1/3, and if one mora is 0.1 seconds, the time is 2.5 seconds.
[0028]
In order to divide the program audio so that it is easy to hear and write, the actual program text is used first to verify the validity of the number of mora and time as the division position and the division length raised here. And examined.
[0029]
<< Examination of division method using actual program text >>
FIG. 2 shows an example in which a news program is divided for writing by hand. Since the example shown in FIG. 2 is a program text, attention is first paid to punctuation marks that are usually present, and the text is divided in consideration of phrases.
[0030]
FIG. 3 shows an example of division for manually transcribe, taking a documentary program as an example. Similarly, in the example shown in FIG. 3, first, attention is paid to punctuation marks, and the sentence is also divided into consideration. Here, the terms used in both figures are defined as follows.
[0031]
Start: Corresponding program audio start time End: Corresponding program audio end time: Corresponding program audio elapsed time Kana: Text kana number Kanji: Number of text kanji etc. Mora: Text mora number (kanji etc.) Calculated as 1.86 mora)
/ S: Number of mora / number of mora per second in elapsed time.
[0032]
In the case of FIG. 2, the elapsed time, that is, the length of time was 2.45 seconds on average (minimum: 1.3 seconds, maximum: 3.4 seconds), and the number of mora was 24.4 on average (minimum: 12.3, maximum: 390/16 at 33.3). The total number of mora (total number of mora) / total elapsed time (total elapsed time) is 390 / 39.2 = 9.95, which is a value related to the speed of speech and is a little faster speech.
[0033]
In the case of FIG. 3, the elapsed time, that is, the length of time was an average of 2.26 seconds (minimum: 0.9 seconds, maximum: 3.5 seconds), and the number of mora was an average of 17.5 (minimum: 9.5, maximum: 25.2). The total number of mora (total number of mora) / total elapsed time (total elapsed time) is 7.76, which is a value related to the speed of speech, which is almost standard speed.
[0034]
For these two program examples, the text is manually divided as shown in FIG. 2 and FIG. 3, and the program audio is divided so as to correspond to the divided text, and a simple transcription experiment is performed while repeatedly reproducing the required number of times. went.
[0035]
As a result, with the exception of some long elapsed times, it was fairly easy to hear, easy to write, and fast, and it was possible to confirm the validity of the division in an average of 25 mola and 2.5 seconds. done.
[0036]
<< Speech / pause detection in actual program audio >>
However, the actual transcription target is speech in the program audio, and there are no punctuation marks or phrases, and there are various background sounds as well as the target speech, and the circumstances are completely different.
[0037]
In the processing of the program audio, preferably, the start / end times of each speech section are obtained correctly, and the speech is divided so as to have an appropriate time length based on the start / end times.
[0038]
However, in the actual program sound, not only the target speech but also various background sounds overlap, and it is generally difficult to correctly obtain the start / end time of each speech section. On the other hand, in the case of a non-speech section, a silent section is certain, and even if there is a background sound, it may be determined that reliability is high depending on the level and section length.
[0039]
Therefore, in the present invention, by analyzing a level signal indicating the amplitude level or power level of the input audio signal, a specific frequency component indicating a fluctuation characteristic of speech at a normal speed is extracted, and an envelope waveform of the extracted specific frequency component is obtained. A first method for detecting a pause section by setting a predetermined slice level for the obtained envelope waveform signal, or by comparing a plurality of LCP cepstrum vectors in acoustic data with each other from a reference frame The pose is detected by the second method (block cepstrum flux method) for detecting the switching point of the contents of the acoustic data more stably.
[0040]
<Description of Embodiment>
FIG. 1 is a block diagram showing an embodiment of a transcription support apparatus according to the present invention.
[0041]
The transcription support apparatus shown in the figure includes a pose detection unit 1 for detecting a pose of an input voice signal, and based on the detected pose, the input is performed for a time or the number of characters that allows a transcription worker to perform short-term storage. And a dividing unit 11 for dividing the audio signal.
[0042]
The pause detection unit 1 includes a specific frequency component extracting unit 3 that extracts a specific frequency component indicating a fluctuation characteristic of speech at a normal speed from a level signal indicating the amplitude level or power level of the input audio signal, and the extracted specific frequency. It comprises a pause section detecting means 5 that obtains a component envelope waveform, sets a predetermined slice level for the obtained envelope waveform signal, and detects a pause section below that level.
[0043]
The dividing unit 11 includes a primary dividing unit 13 that divides the input audio signal according to the most reliable pose among the poses detected by the pause detecting unit 1, and the primary dividing unit 13. As a result of the division, the second division means 15 that divides the input voice signal by the next highly reliable pose for a portion that is longer than a predetermined time or a predetermined number of characters that can be short-term memorized, and this second As a result of the division by the next dividing means, the input speech signal is divided or pre-determined with the next most reliable pose for a portion that is longer than a predetermined time or a predetermined number of characters capable of short-term storage. A third dividing unit 17 that mechanically divides by the divided time or the number of characters, and an overlapping unit 19 that overlaps the front and rear portions of the divided voice by a predetermined time or the number of characters. It has been.
[0044]
Next, the operation of this embodiment will be described separately for the process of the pause detection unit 1 and the process of the division unit 11.
[0045]
<< Processing of Pause Detection Unit 1 >>
4 to 7 are explanatory diagrams showing the procedure of the pause detection process.
[0046]
In the flowchart of FIG. 4, first, a program audio signal having an amplitude waveform as shown in FIG. 5A is fetched, and the audio amplitude value is normalized (steps ST1 and ST3). In this process, only the low frequency component of the envelope is extracted from the captured program audio signal (step ST5), and the signal component is set to a predetermined level based on the amplitude value of the low frequency component. . Generally, when the level of the low frequency component is large, the level of the high frequency component is also considered to be large. This is because if there is a difference in level, the detection accuracy will be affected, so that it is necessary to achieve a certain level of standardization.
[0047]
The audio signal whose audio amplitude value has been standardized in this way is then subjected to absolute value processing, and the amplitude value waveform signal (Lch, Rch stereo audio that is turned back to the + side as shown in FIG. 5B). If there is a signal, the speech portion is expanded using the sum and difference signals) (step ST7). The waveform in FIG. 6 (A) is obtained by enlarging the time axis in the range indicated by the arrow in FIG. 5 (B) 10 times.
[0048]
Next, a digital high-pass filtering process (step ST9) and a digital band filtering process (step ST11) are executed from the amplitude value waveform signal converted into an absolute value. In the digital high-pass filtering process, a signal having a frequency of, for example, 2 Hz or more is extracted, and in the digital band filtering process, a signal having a frequency of, for example, 4 to 7 Hz is extracted. The waveform in FIG. 6B is a signal waveform having a 4 to 7 Hz component.
[0049]
It is known that the speech power corresponding to the vowel phonetic symbols tends to be higher than the other portions when the fluctuation characteristics of the speech in the time axis direction are compared with the phonetic symbol strings of the speech. The prior application showed that the fluctuation in the time axis direction of the signal waveform in the normal speed speech was a frequency of about 4 to 7 Hz (Japanese Patent Application No. 2002-160255). In this embodiment, paying attention to the fluctuation of the waveform in the time axis direction, the pose portion is detected by capturing rough periodicity.
[0050]
When a signal having a frequency of 4 to 7 Hz is extracted by digital band filtering and further subjected to absolute value processing (step ST12), a signal having a waveform similar to speech is obtained. On the other hand, when the arithmetic processing (step ST13) is executed, components other than the main speech component are extracted as the difference between these extracted signals, and further, an absolute value processing (step ST15) and a digital low-pass filtering processing (step ST15). Step ST17) is executed. In the level correction process (step ST19), the waveform of the low-frequency component of 0.5 Hz or less generated by the low-pass filtering process mainly includes components other than the speech component. The level of the speech-like waveform signal is corrected in the reverse direction, and as a result, the speech component is enhanced. The speech approximate waveform after this processing is shown in FIG.
[0051]
Thus, the level-corrected envelope waveform signal is displayed on the display for verification of its suitability (step ST21). Then, as shown in FIG. 7, the slice is sliced at a predetermined slice level (threshold) (step ST23).
[0052]
Next, a minute pause section removal process is executed (step ST25). This is because, in this processing, for example, a slight vowel interval variation or a breathing-interval interval is excluded from the detection target from the sliced audio amplitude value signal. As the detection time range, for example, set to "0.2 to 2 seconds" extent, its is a process of removing the detected outside the following times. This effectively prevents pose detection that has no meaning and low reliability.
[0053]
The detected pause section is displayed on the screen (step ST27). The speech section is measured for the purpose of optimizing the pause detection accuracy and slicing (step ST23) level setting, and compared with the measured speech section as shown in FIG. 7 (step ST29). As a result, the pose interval derived from the actually measured speech interval is compared with the detected pose interval, and the comparison can be performed so that the pose detection accuracy can be checked or the slice level can be optimized. .
[0054]
<< Processing of Dividing Unit 11 >>
When a pose is detected as described above, a relatively long pose is the most reliable as shown in the flowchart of FIG. To divide the input audio signal (step ST41). This process is executed by the primary dividing means 13.
[0055]
As a result of this division, the input voice signal is divided by the next long pause (about 0.5 to 1 second) for a portion that is longer than a predetermined time or a predetermined number of characters that can be stored in the short term for the worker (step 0.5 to 1 second) ST43, ST45). This process is executed by the secondary dividing means 15.
[0056]
As a result of this division, for the portion that is longer than a predetermined time or a predetermined number of characters that can be short-term memorized by the operator, the input voice signal is output by the next long pause (for example, 0.2 to 0.5 seconds). Or mechanically divided by a predetermined time or number of characters (steps ST47 and ST49). This process is executed by the third dividing means 17.
[0057]
When the dividing unit 11 performs division by the dividing units 15 and 17 that may have a problem in reliability, the overlapping unit 19 converts the front and rear parts of the divided voice to a predetermined time or number of characters. (Step ST51). For example, by providing an overlapping portion of about 1 second, it is possible to prevent the transcription operator from being inaudible and lack of transcription caused by the low reliability of pause detection.
[0058]
《Pause detection by block cepstrum flux method》
A brief description will be given of the block cepstrum flux method as a second method for detecting a speech interval.
[0059]
In this method, a plurality of LCP cepstrum vectors in acoustic data are compared with each other from a reference frame, thereby detecting the switching point of the contents of the acoustic data more stably.
[0060]
The graph of FIG. 9 shows the result of analyzing the sound of an actual television program that is considered to be an average background sound level by both methods.
[0061]
In FIG. 9, (A) is a processing value (cflx analysis value) by the block cepstrum flux method, (B) is a 4-7 Hz component extraction value (by the first method) of the speech envelope, and (C) is a value of (B). Data obtained by slicing the waveform at an appropriate level and binarizing it.
[0062]
In the waveform of (C), the high level range corresponds to speech, and the low level range corresponds to the non-speech (pause) section. In this case, (C) matches the measured speech section fairly well. I can say that.
[0063]
Therefore, it is possible to accurately grasp the speech section in the speech from the waveform (C). In the waveform of (C), P1 is a short pose, P3 is a relatively long pose, P3 is the most reliable pose, and P2 is the next most reliable pose.
[0064]
The block cepstrum flux method (second method) of (A) was also compared in the same manner. Although the result is slightly worse than the first method, it can be applied in some cases.
[0065]
<Division experiment using pause detection of actual program audio>
Detect the place where there is no speech in the actual program audio, or breaks due to breathing, etc. by appropriate processing of the program audio (use it as a pose for writing), and use this pose etc. as a guide for the speech to be transcribed Divided into easy-to-work units, and slowed down as necessary.
[0066]
As described above, it is desirable that the speech time during which the writing operation is easy is about 2.5 seconds. Since the target program has a fairly high level background sound (music), the division method here first detects and divides a relatively long pause (for example, about 2 seconds or more).
[0067]
FIG. 10 illustrates pause portion detection from the sound of the target program (there is a fairly high level background sound (music)).
[0068]
The original audio waveform of the target program is shown at the bottom. Using the first method described above, the frequency range component of 4 to 7 Hz is extracted from the envelope of this waveform, and the amplitude value is expressed as “processed audio waveform. ". The “processed speech waveform” was compared in magnitude with an appropriate threshold level to obtain a “detection pose”. In order to evaluate the suitability of this “detected pose”, data of actual speech was also displayed for reference. It can be said that at least a relatively long pause (for example, about 2 seconds or more) is correctly detected.
[0069]
Using this pose detection result, the first division when the pause was 1 to 2 seconds and the second division of 0.5 to 1 second were performed. Even in this division, the portion that is about 2.5 seconds or longer was divided to be about 2.5 seconds by one of the following methods.
[0070]
(1) A shorter pause (for example, 0.2 to 0.5 seconds or more) is detected and further divided using this.
(2) Divide this part equally so that it takes about 3 seconds.
[0071]
In addition, at least a portion to which the processing of (1) or (2) was applied was provided with an overlapping portion of about 1 second so as to avoid omission and inaudibility. Further, in order to indicate an overlapping portion, an overlapping display sound CH is inserted at the beginning of the portion so that it can be seen that the portion is overlapping.
[0072]
<Example of division / duplication processing>
An actual processing example is shown in the following text. The underlined portion indicates duplicate processing. Further, PP represents a detection pause of 1 to 2 seconds or more, SP: a detection pause of 0.5 to 1 second or more, and CH: an overlapping display sound (underlined portions are overlapping portions).
[0073]
Example of processing (1) 1 second overlap 18 divisions
PP Nature, there are always creatures
There is a wonder and excitement that the CH creatures weave. SP
Now, go to the middle of nature and travel to the earth. SP
Seeking a new encounter, aside of
CH is the beginning of a special earth trip. PP
The US Costa Rica medium is the stage of today
CH Rice is a tropical forest that spreads over Costa Rica . SP
In this forest, many creatures
CH creatures are living with ingenuity. SP
Above all, living with unique wisdom
The white-tailed capuchin monkey lives with CH. PP
The wisdom of capuchin capuchin monkeys varies. SP
And care Tata nuts in stone, medium
Tighten CH and take out the seeds inside . SP
Furthermore, live in peace with different kinds of monkeys
There is even the wisdom of bargaining that lives in CH peace . PP
Throat Giro capuchin is, look in the life of the forest, SP wisdom
I looked at the wisdom of CH. PP
[0074]
(2) Processing example Equal division between PPs, 1 second overlap 16 divisions
PP Nature, there are always living things
There is a wonder and excitement that the creatures weave. here we go
Now , in the midst of nature, a living place
Creatures land ball travelogue. Seeking to meet new out,
In search of a meeting, it is the beginning of a special earth trip. PP
Today's stage is spread to Central America Costa Rica
A tropical forest that stretches over mosquitoes . In this forest
In this forest, many living creatures live with ingenuity.
I'm alive . Above all, even the wisdom of the German special
The lives it also the wisdom of the special is the throat Giro capuchin monkeys. PP
The wisdom of capuchin capuchin monkeys varies. wood
And slammed the fruit of the tree in stone, take out the seeds in
Take out the seeds . Furthermore, different types of monkeys and flat
Different live monkeys and peace, but there's up to the wisdom of bargaining. PP
Throat Giro capuchin is, look in the life of the forest
I watched a lot of wisdom to show in. PP
[0075]
《Transcription experiment using divided program audio》
The programs targeted for the experiment are 78 seconds in total, and the speech part is 48 seconds (62%).
[0076]
An experiment was performed using the program audio by the above dividing method. As shown in FIG. 11, the experiment was performed by using the divided audio processed as described above in MIDIPLAYER (shareware software) and using the operation function of this software. Specifically, each divided sound is repeatedly reproduced, and when writing of the hatched portion is completed, the operation proceeds to the next divided sound by a key operation. FIG. 11 shows a processing example of (1) above.
[0077]
As a result of experimenting with the software of FIG.
Transcription time by method (1): approx. 135 seconds Number of repetitions of each divided voice: 2.8 times Transcription time by method (2): approx. 152 seconds The number of repetitions of each divided voice is about 3.2 times. It was found that a division method that uses the detected pose as much as possible is desirable.
[0078]
【The invention's effect】
As described above, according to the present invention, when the detected pause is used, the program audio signal is subjected to processing such as appropriate division of the program audio, speed reduction or repetition if necessary, and it is easy to hear and write. It will be possible to support.
[0079]
In addition, since the front and rear parts of the divided specific speech are overlapped by a predetermined time or the number of characters by the overlapping means, it is possible to effectively prevent the writing operator from being inaudible or missing the writing. Can do.
[Brief description of the drawings]
FIG. 1 is a block diagram showing an example of a transcription support apparatus according to the present invention.
FIG. 2 is an explanatory diagram showing an example of a division experiment for writing a news program.
FIG. 3 is an explanatory diagram showing an example of division verified by writing a documentary number.
FIG. 4 is a flowchart showing a procedure for pause detection.
FIG. 5 is an explanatory diagram showing a program audio signal waveform and an amplitude level signal or a power level signal corresponding to the waveform.
6 is an explanatory diagram showing the amplitude level signal or the power level signal shown in FIG. 5 with the time axis enlarged, and the extracted component values in the characteristic frequency range. FIG.
7 is an explanatory diagram showing an envelope waveform and an actually measured pause (speech) portion of an amplitude level signal or a power level signal in a specific frequency range shown in FIG. 6;
FIG. 8 is an explanatory diagram showing a processing procedure for dividing program audio by a pause;
FIG. 9 is an explanatory diagram of speech approximation data obtained by specially processing an audio signal.
FIG. 10 is an explanatory diagram of pause portion detection.
FIG. 11 is an explanatory diagram illustrating a screen example of a computer used when a divided program sound is played back using sound playback software.
[Explanation of symbols]
1 Pause Detection Unit 3 Specific Frequency Component Extraction Unit 5 Pause Section Detection Unit 11 Division Unit 13 Primary Division Unit 15 Secondary Division Unit 17 Third Division Unit 19 Overlapping Unit

Claims (4)

A writing support device that supports a task of writing text data from the voice while listening to the voice,
A pause detector for detecting the pause of the input audio signal;
A division unit that divides the input voice signal by a time that allows a transcription worker to perform a short-term memory or the number of characters converted from the time based on the detected pose ;
The dividing unit is
Primary dividing means for dividing the input audio signal by the most reliable pose among the poses detected by the pose detector;
As a result of the division by the primary dividing means, the input voice signal is divided in the next highly reliable pause for a portion that is longer than a predetermined time capable of short-term storage or more than the number of characters converted from the time. Secondary dividing means;
As a result of the division by the second division means, for the portion that is longer than a predetermined time capable of short-term storage or more than the number of characters converted from the time, the input voice signal is sent to the next highly reliable pose. Tertiary dividing means for dividing, or mechanically dividing by a predetermined time or the number of characters converted from the time;
A transcription support device characterized by comprising: The transcription support apparatus according to claim 1 ,
The time for which the short-term memory is possible is about 2.5 seconds, or the number of characters converted from the time for which the short-term memory is possible is about 25 mora in number of mora. Support device. In the transcription support apparatus according to claim 1 or 2 ,
The dividing unit is
Including overlapping means for overlapping the front and rear portions of the divided speech by a predetermined time or the number of characters converted from the time,
A transcription support device characterized by that. The transcription support apparatus according to any one of claims 1 to 3 ,
The pause detection unit
Specific frequency component extraction means for extracting a specific frequency component indicating a fluctuation characteristic of speech at a normal speed from a level signal indicating an amplitude level or a power level of the input audio signal;
Pause section detection means for obtaining an envelope waveform of the extracted specific frequency component, setting a predetermined slice level for the obtained envelope waveform signal, and detecting a pause section;
A transcription support device characterized by comprising:

Images