JP2008107641A - Voice data retrieving apparatus - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To accurately retrieve a desired part of a recorded voice data. <P>SOLUTION: Voice in a conference is stored as the voice data in a voice data storing section 17. A feature of the voice data is extracted for each predetermined frame by a CPU 11, and stored as a feature data string together with time information in an analysis data storing section 18. When retrieved, an operator inputs a desired word toward a microphone 16. The feature of the voice is extracted for each predetermined frame by the CPU 11, and stored as the feature data string in a RAM 13. Then, coincidence of the feature data string in the RAM 13 and the feature data string in the analysis data storing section 18 is detected. The time information attached to the feature data string in the analysis data storing section 18, which is detected as coincident, is extracted. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to a speech data retrieval apparatus for retrieving a desired portion from stored speech data.

  If you want to search the stored voice data for the part where the desired keyword is spoken, for example, save the meeting voice as text data, enter the text data as the search key, There is a method for searching for a text data portion that matches a search key (Patent Document 1).

In addition, a method has been proposed in which the switching state of the operation of the application software for presentation is recorded in synchronization with the conference voice, and the voice is searched using the operation state of the presentation as a key (Patent Document 2).
JP 2002-366552 A Japanese Patent No. 3637937

  However, in Patent Document 1, a voice that is clearly pronounced like a narration can be made into text with high accuracy. However, as in a meeting, it is possible to make texts that are spoken by various people in a normal conversation. Current speech recognition technology is not accurate enough to make it into text. If the text is inaccurate, there is a problem that desired data can hardly be searched.

  Further, in Patent Document 2, since there are many conferences that do not use application software for presentation, there is a problem that there are many cases where the conference software cannot be used at all. Even if application software for presentation is used, there is a problem in that accurate search cannot be performed because the audio data to be searched does not necessarily correspond to the switching timing of the presentation operation.

  In order to solve the above-described problems, the present invention provides an audio data search device capable of accurately searching for desired audio data even when application software for presentation is not used. Objective.

In order to solve the above problems, in the present invention,
Sound collection means for outputting sound data corresponding to the collected sound;
Voice data storage means for storing voice data output by the sound pickup means;
Feature data generating means for analyzing the voice data output by the sound collecting means and generating feature data indicating the characteristics;
Feature data storage means for storing the feature data generated by the feature data generation means together with time data corresponding to the generation time;
Search key input instruction means for instructing input of a search key;
Search feature data storage means for storing the feature data generated by the feature data generation means as search feature data when the search key input instruction means is instructed to input a search key;
A search means for comparing the feature data for search in the feature data storage means for search with the feature data in the feature data storage means and searching for feature data that is considered to be matched;
Reading means for reading out the voice data stored in the voice data storage means from an address corresponding to the time data of the characteristic data searched by the search means.

In another aspect of the invention,
Sound collection means for outputting sound data corresponding to the collected sound;
Voice data storage means for storing voice data output by the sound pickup means;
Feature data generating means for analyzing the voice data output by the sound collecting means and generating feature data indicating the characteristics;
Feature data storage means for storing the feature data generated by the feature data generation means together with time data corresponding to the generation time;
A character string input means for inputting a character string;
A table in which phonemes that are constituent elements of a character string are associated with voice feature data when the phonemes are pronounced;
Conversion means for converting each character of the character string input by the character string input means into feature data with reference to the table;
Search feature data storage means for storing the feature data converted by the conversion means as search feature data;
A search means for comparing the feature data for search in the feature data storage means for search with the feature data in the feature data storage means and searching for feature data that is considered to be matched;
Reading means for reading out the voice data stored in the voice data storage means from an address corresponding to the time data of the characteristic data searched by the search means.

In another preferred embodiment of the present invention,
The sound collecting means generates a plurality of microphones and sound data corresponding to the sound picked up by each microphone, and adds identification data for identifying which microphone the sound data is from. Audio data generating means for attaching to audio data;
The feature data storage means classifies and stores the feature data based on the identification data,
When the identification data is specified, the search means compares the feature data classified by the specified identification data with the search feature data in the search feature data storage means.

In another preferred embodiment of the present invention,
The sound collection means includes an array microphone whose sound collection direction is variable, a sound collection direction control means for controlling the sound collection direction of the array microphone and outputting direction data indicating the sound collection direction, and the array microphone. Voice data generating means for generating voice data corresponding to the sound that has been sounded,
The feature data storage means classifies and stores the feature data based on the direction data,
When the direction data is specified, the search means compares the feature data classified by the specified direction data with the search feature data in the search feature data storage means.

  Since comparison is performed using feature data extracted from voice data, it is not necessary to convert the voice data into text data or the like, and an accurate search can be performed. Also, no presentation software is required.

(First embodiment)
(A) Configuration FIG. 1 is a block diagram showing a hardware configuration of a conference system according to the first embodiment of the present invention. A CPU (Central Processing Unit) 11 shown in FIG. 1 reads a computer program stored in a ROM (Read Only Memory) 12, loads it into a RAM (Random Access Memory) 13, and executes it to execute hardware. Control each part. The RAM 13 is also used as a work area for the CPU 11. The operation unit 14 includes various keys and outputs a signal corresponding to the pressed key to the CPU 11.

  The microphone 16 collects ambient sounds and outputs them as audio signals. The input IF (Interface) 15 samples an audio signal (analog signal) output from the microphone 16 at a predetermined sampling frequency and converts it into audio data Sad. Here, FIG. 2 is an example of the audio data Sad. As shown in the figure, the data string represents the amplitude at each sampling timing along the time axis.

  Next, the audio data storage unit 17 illustrated in FIG. 1 sequentially stores the audio data Sad output from the input IF 15 under the control of the CPU 11. In this case, the audio data Sad at each sampling timing is sequentially stored in one address of the audio data storage unit 17.

  Further, the CPU 11 analyzes the voice data Sad output from the input IF 15 to generate analysis data, and sequentially stores the generated analysis data in the analysis data storage unit 18.

  Here, a method for generating analysis data will be described. In this embodiment, as shown in FIG. 3, the frequency spectrum is generated by performing fast Fourier transform (FFT) on the audio data Sad at every frame of a predetermined time interval (10 msec in this embodiment). Examples of frequency spectra in the frames fr1 to fr3 shown in FIG. 3 are shown in (a) to (c) of FIG. As shown in this figure, for each frame, the frequency and amplitude of a sine wave included in that frame are extracted. The CPU 11 performs normalization processing described below on the frequency and amplitude of the sine wave included in each frame extracted in this way.

  First, the lowest one of the sine wave frequencies for each frame is set as the pitch, and the average value of the amplitude of the sine wave in each frame is set as the average sound pressure level of each frame. Then, the frequency of each sine wave in each frame is divided by the pitch, and the amplitude of the sine wave in each frame is divided by the average sound pressure level. As a result of such processing, for each frame, a normalized frequency and amplitude data string is generated from the lower frequency side to the higher frequency side. Here, a data string of (f1, A1), (f2, A2), (f3, A3)... Is generated from the lower frequency side to the higher side. Note that the numbers indicate the order from the lowest frequency in each frame, and even if the numbers are the same in each frame, they do not indicate the same frequency and the same amplitude. In the following description, this data string is referred to as a feature data string.

  FIG. 5 is a diagram showing the storage contents of the analysis data storage unit 18. As shown in the figure, one record includes a frame number (fr1, fr2, fr3...), Time data, and a feature data string. The time data is the start time of each frame. Note that the time data in this case only needs to correspond to the generation time of the feature data string, and may be the start time and end time of the frame, or the write time to the analysis data storage unit 18. Alternatively, the sound collection time of the first audio data of each frame may be used. In short, it is only necessary to be able to specify the time, and any time corresponding to the generation time of the feature data may be used.

  The display unit 20 shown in FIG. 1 includes a display, and displays predetermined characters and diagrams under the control of the CPU 11. The reproduction unit 21 converts the audio data Sad into an audio signal under the control of the CPU 11. The speaker 22 outputs the converted audio signal as audio.

(B) Operation Next, the operation of this embodiment will be described. In the following, a case will be described as an example where the audio of the conference is stored and a desired portion is searched from among them.

  First, the microphone 16 is placed on a conference table or the like, and each speech of the conference participant is recorded. That is, the microphone 16 collects the speech of each participant and outputs it as an audio signal. As a result, audio data Sad as shown in FIG. 2 is output from the input IF 15, and the amplitude at each sampling timing is sequentially recorded in the audio data storage unit 17.

  At the same time, the CPU 11 analyzes the voice data Sad and sequentially stores the analysis results in the analysis data storage unit 18. As a result, feature data strings as shown in FIG. 5 are sequentially stored. In this way, each utterance in the conference is stored as voice data Sad in the voice data storage unit 17, and its characteristics are analyzed and stored in the analysis data storage unit 18 as a feature data string.

  Next, when a request for listening to a desired portion of the recorded audio data is generated, the operator presses a predetermined button of the operation unit 14 and utters a word as a keyword for search toward the microphone 16. . For example, when the keyword "Hello", by pressing a predetermined button in the operation unit 14 say "Hello", the words of the audio data Sad is generated, along with stored in RAM 13, when the recording of the meeting It is analyzed by the same process. The analysis result is stored in a predetermined area of the RAM 13 as search feature data. FIG. 6 shows the stored contents. In this way, feature data string "Hello" is detected for each frame FR1, FR2 ....

Subsequently, the CPU 11 sequentially collates the keyword feature data string stored in the RAM 13 with the conference voice feature data string stored in the analysis data storage unit 18. Here, the coincidence of feature data strings between frames will be described. For example, for the first frame, the feature data of the frames FR1 and fr1 are sequentially compared from the lower frequency to determine whether or not they match, and a predetermined allowable range is set for the match determination. .
For example, even if the values of f1 of the frame FR1 and f1 of the frame fr1 do not completely coincide with each other, they are regarded as coincident if they are within an allowable error (for example, 10%). Similarly, when the relative error of the amplitude A1 is within 10%, for example, the amplitudes are considered to match. When both the frequency component and the amplitude component are regarded as matching, the sine wave components are regarded as matching. In this way, when comparison is made in the order of (fr1, A1), (fr2, A2), (fr3, A3), etc., 90% of all samples (for example, 50 to 100) are regarded as matching Is determined that the frames FR1 and fr1, which are the first frames, match. This determination is performed for each frame.
In this case, since the frequency and amplitude are normalized as described above, even if the keyword pronounced by the operator differs from the pronunciation of the conference by the speaker in the pitch (pitch) and sound pressure level, the feature data If they match, it is determined that the words match. Therefore, it does not become a different search target depending on the individuality of pronunciation of the operator or the speaker. The allowable range described above can be set as appropriate according to the implementation status. The setting may be performed by key operation of the operation unit 14 or may be stored in advance in the ROM 12 or RAM 13 as a default value.

  Here, the content of the matching search process will be further described. CPU11 is recognized, the successive frames correspond to the pronunciation of the single word (hereinafter, referred to as a frame group), analyzes the characteristic data sequence in the analytical data storage unit 18 as the keyword "Hello" one word stored in the RAM13 Then, a matching frame group is extracted. In other words, the slide into comparing the features data columns of the conference pronunciation's an operator from the beginning of the frame in the order of "Hello".

  In this case, although the length of the pronunciation may be different between the operator and the conference speaker, the CPU 11 performs DP (Dynamic Programming) for two feature data strings corresponding to the pronunciation of the operator and the conference speaker. Comparing sequentially according to the matching algorithm. By performing the DP matching process, the frames in which the characteristics of the operator voice and the meeting speaker voice are matched are associated. As a result, be different from the length of the pronunciation, the search can be performed if the pronunciation of the same "Hello". That is, be different from the number of frames corresponding to pronounce "Hello" is extracted from the feature data string of the operator was blown "Hello" and analysis conference speaker stored in the data storage unit 18, both match search is possible if the pronunciation of the same "Hello".

  In this case, there is the frame that matches the "Hello" is more detected from within the analytical data storage unit 18. In the present embodiment, CPU 11 is consistent for a frame group "Hello" is be detected, with reference to the degree of matching of each frame in both frame group, and calculates the degree of matching between frame group.

For example, if the frames detected as matching both have 100 frames, the feature data strings are considered to match in 97 frames, and the other three frames do not match. When it is regarded, the calculation is performed such that the degree of coincidence of the feature data string of the conference speaker is 97%. Or it is good also considering the average of the coincidence degree of each frame as a coincidence degree of a frame group. Further, the degree of coincidence between the frame groups when the number of frames is different may be determined by proportional distribution. For example, when performing a match determination between a frame group of 30 frames and a frame group of 90 frames, the degree of match is such that the number of matching frames in the former frame is tripled and divided by 90. % Can be obtained.
On the other hand, the permissible ratio of frames that are considered to be inconsistent among the frames included in the frame group is set in advance. However, if there is only one mismatched frame, it may be set not to be recognized as a match. You may set it to accept the discrepancy.

  As described above, the frame group corresponding from the analysis data storage unit within 18 to "Hi", the degrees of the frame group is detected. Here, a display example on the display unit 20 when the frame group is detected matching "Hello" in Fig. FIG. 7 shows a display example when coincidence is detected at three locations in the analysis data storage unit 18. As shown in the figure, the number indicating the detection order, the time, and the degree of coincidence are displayed. The time in this case is the time (see FIG. 5) of the first frame of the frame group in the analysis data storage unit 18 determined to match.

  A cursor Csr is displayed on the display unit 20, and the cursor Csr moves under the control of the CPU 11 as a predetermined key of the operation unit 14 is pressed. When a predetermined key (such as the Enter key) is pressed, the CPU 11 calls the time specified by the cursor Csr and recognizes it as the start time, and reads out audio data corresponding to this time from the audio data storage unit 17. Since the audio data in the audio data storage unit 17 is sequentially stored according to the sampling timing, the difference of one address corresponds to the sampling period, so that the address corresponding to the read start time can be easily obtained. The audio data read out in this way is supplied to the reproduction unit 21 where a reproduction signal is generated and generated from the speaker 22.

  As described above, the sound that matches the was blown into the operator "Hello", ie the voice of the reproduction is made from the point where the conference speaker is pronounced "Hello". When the sound reproduced in this way is not the desired one, the operator can select and listen to another candidate from the displayed list, thereby easily listening to the sound of the desired part. Search and listen to. Thus, in this embodiment, desired voice data can be detected by directly comparing the characteristics of recorded voice and search voice without using any character string input or voice recognition.

(Second Embodiment)
Next, a second embodiment of the present invention will be described. In the following description, parts common to those in the first embodiment are denoted by common reference numerals and description thereof is omitted.
(A) Configuration The present embodiment is different from the first embodiment described above in that a text phoneme feature conversion unit 19 is provided. The text phoneme feature conversion unit 19 has a function of converting text data input from the keyboard of the operation unit 14 into a feature data string.

  For example, from the keyboard provided in the operation unit 14, a character string "Hello" if entered, it converts the input character string in hiragana strings representing actual pronunciation by morphological analysis. Here, the morphological analysis is a process for recognizing a word from a character string. In other words, unlike English documents, Japanese sentences are not “separated”, so there is no space between words, and it is difficult to cut out and recognize words. Therefore, in the morpheme analysis, the morpheme analysis is performed based on a morpheme dictionary database (not shown) stored in advance, and the part of speech is determined by dividing into word units. In the present embodiment, the sound is converted into a sound corresponding to a sound that is actually generated. For example, to explain the word “Konchi ni”, the hiragana string representing this pronunciation is “Konchiwa”. That is the word "hello" is extracted from the morpheme dictionary database, further its actual pronunciation by referring to the internal pronunciation dictionary database (not shown) is recognized as "Hello", corresponding to the recognition result " Ask for "kana".

  When the phoneme is obtained in this way, the text phoneme feature conversion unit 19 refers to the text phoneme feature conversion table (see FIG. 9) stored in the phoneme feature conversion unit 19 and searches for the phoneme corresponding to the actual pronunciation “Konchiwa”. Generate feature data. In the text phoneme feature conversion table, as shown in FIG. 9, a frame group corresponding to each phoneme “A”, “I”, “U”... Is set, and a feature data string is stored in each frame in each frame group. Has been written. This feature data string is normalized data in the same manner as the feature data string stored in the analysis data storage unit 18 of the first embodiment. Note that the frame group corresponding to each phoneme in FIG. 9 is shown only for five frames in order to simplify the description, but actually includes a larger number of frames.

  In the case of English text, morpheme analysis is not required, but phonemes are extracted by referring to the dictionary database from the spelling of the input character string, and the feature data string corresponding to the extracted phoneme is shown in the text in FIG. Obtained by referring to the phoneme feature conversion table. In this case, in the text phoneme feature conversion table, it is necessary to set in advance a feature data string corresponding to English phonemes.

(B) Operation Next, the operation of this embodiment will be described. From the keyboard of the operator the operation unit 14, for example, if you enter the keyword "Hello", the text phoneme feature transformation unit 19 converts the input string to Hiragana string "Hello" to represent the actual pronunciation by morphological analysis A frame group having a feature data string corresponding to this is obtained with reference to the text phoneme feature conversion table shown in FIG. CPU11 writes the frame group corresponding to the "Hello" text phoneme feature transformation unit 19 was determined to RAM13.
Next, as in the first embodiment described above, the CPU 11 obtains a frame group that matches the frame group written in the RAM 13 from the frame group in the analysis data storage unit 18, and displays the searched candidates on the display unit 20. indicate. If the operator selects a desired candidate from the display content of the display unit 20, the corresponding sound is emitted from the speaker 22. This operation is the same as in the first embodiment.

  As described above, according to the second embodiment, even if a character string is typed from the keyboard, a frame group having a feature data string corresponding to the character string is specified, and the match search is a comparison between the feature data strings. It is not necessary to convert conference voice or the like into a character string by voice recognition, and a search can be performed by comparing voice features.

(Modification)
In addition, this invention is not limited to embodiment mentioned above, It can implement in a various aspect. An example is shown below.

(Modification 1)
When features common to a plurality of frames are continuous, a rule may be provided that considers the same based on the number of consecutive frames. For example, a rule may be provided such that, from the fifth frame to the 30th frame, phonemes are considered to be the same when frames having characteristic data sequences that are regarded as matching are consecutive.

(Modification 2)
As shown in FIG. 10, a plurality of microphones 16 such as microphones A, B, and C may be provided. Further, in this case, input path information (identification data) is added to each microphone input terminal, and the input path information is added to the analysis data as shown in FIG. 11, so that a speaker can be distinguished and voice data can be classified. Therefore, by performing a search using the input route information and keywords, the search range can be narrowed and the search efficiency can be improved. In the conference, each speaker rarely speaks at the same time, and the sound at a certain time is picked up by one of the microphones A, B, and C directed to one speaker as shown in FIG. It is because it can be estimated that it was done.

(Modification 3)
FIG. 10 shows an example in which three microphones are installed, but instead of this, a microphone array system 30 having a plurality of microphones Mic may be used as shown in FIG. Since the microphone array system 30 can spatially generate the voice input direction, the microphone array system 30 is configured to supply the input IF 15 with direction information indicating the input direction and a voice signal. The input IF 15 samples the audio signal at a predetermined sampling frequency and converts it into audio data Sad, and outputs direction information. The audio data storage unit 17 sequentially stores the audio data Sad output from the input IF 15 under the control of the CPU 11 and stores a direction information by providing a predetermined header. Since this direction information identifies the speaker, the voice data is classified as in the second modification, and the search efficiency can be improved.

(Modification 4)
As a mode of storing the voice data in the voice data storage unit 17, any mode may be used as long as time and amplitude are related to each other. For example, the physical address of the storage area of the audio data storage unit 17 may correspond directly to the time, or a header for storing the time may be inserted for each predetermined memory block. The memory block length may be fixed, or may be variable length memory block data in which the value of the memory block length is included in the header. When time data is assigned to each memory block, the search time is also discrete in units of memory blocks, but the time to be searched becomes ambiguous by appropriately setting the size of the memory block. Such a problem does not occur.

Further, the audio data may be stored in a continuous storage area, and a table for storing the correspondence between the time data and the physical address of the storage area may be stored in another storage area.
Also in the above-described case and in the case of the first and second embodiments, the audio data can be compressed and stored.
Furthermore, when recording conference audio or the like, it is useless to store audio data during silent periods, so it is desirable not to store audio data having no amplitude value greater than a predetermined intensity. In this case, time data at the time of resuming recording may be stored (time stamp), or time data may be included in the header of fixed-length or variable-length memory block data as described above.

(Modification 5)
The algorithm of the analysis data generation method is not limited to the fast Fourier transform (FFT). Any algorithm may be used as long as it is capable of generating a unique frequency and amplitude spectrum for each frame as shown in FIG. For example, another discrete Fourier transform or wavelet transform algorithm may be used.
Each frame may have a period that overlaps the preceding and following periods. Thereby, the analysis accuracy can be improved.

(Modification 6)
The analysis of the sound data Sad may be performed simultaneously with the sound data Sad being stored in the sound data storage unit 17 or may be performed separately. For example, the stored voice data Sad may be read and analyzed.

(Modification 7)
Each data may be stored directly in the audio data storage unit 17 or the analysis data storage unit 18, but may be buffered in a predetermined buffer memory or a storage area of the RAM 13. When buffering is performed, the data stored in the temporary storage area can be searched quickly, which is suitable for the case where it is desired to reproduce the previous message.

(Modification 8)
In each embodiment described above, a feature code (feature data) may be assigned to the feature data string. That is, there are cases in which similar feature data strings can be classified as one common set. A feature code is assigned to the set classified in this way. Then, this feature code is added to each record in the table as shown by the broken lines in FIGS. 5, 6, 9, and 11. With such a configuration, it is possible to detect the match between the operator's voice and the frames in the analysis data storage unit 18 by comparing the feature codes, so that the processing speed of the match detection can be greatly improved. . In this case, a certain allowable value may be provided for matching of feature codes. In other words, feature codes having similar feature data strings may be regarded as a match after giving a complete match or a certain match (80%, 90% match).
Moreover, if it carries out as mentioned above, it will suffice to memorize | store only the characteristic code in the analysis data storage part 18 instead of a characteristic data sequence, and a memory area can be made small.

  As a method for deriving a feature code from a feature data string, for example, a considerable number of Japanese alphabets are sampled in advance, and this is analyzed by the method used in the above-described embodiment, and similar to the analysis result. For example, there is a method in which characteristic codes are assigned together with a characteristic code.

It is a figure which shows the hardware constitutions of 1 embodiment of this invention. It is a wave form diagram which shows an example of audio | voice data. It is a wave form diagram for showing the frame used for analysis of voice data. It is a figure which shows the spectrum of each flame | frame. It is a figure which shows an example of analysis data. It is a figure which shows an example of the characteristic data for search. It is a figure which shows the example of a display of a search result. It is a figure which shows the hardware constitutions of the 2nd Embodiment of this invention. It is a figure which shows a text phoneme characteristic conversion table. It is a block diagram which shows input IF provided with the several microphone. It is a figure which shows the example of the analysis data containing input path information (tag). It is a timing chart which shows the timing of the sound inputted into a plurality of microphones. It is a block diagram which shows input IF provided with the microphone array system.

Explanation of symbols

DESCRIPTION OF SYMBOLS 11 ... CPU, 12 ... ROM, 13 ... RAM, 14 ... Operation part, 15 ... Input IF, 16 ... Microphone, 17 ... Audio | voice data storage part, 18 ... Analysis data storage part

Claims (4)

Sound collection means for outputting sound data corresponding to the collected sound;
Voice data storage means for storing voice data output by the sound pickup means;
Feature data generating means for analyzing the voice data output by the sound collecting means and generating feature data indicating the characteristics;
Feature data storage means for storing the feature data generated by the feature data generation means together with time data corresponding to the generation time;
Search key input instruction means for instructing input of a search key;
Search feature data storage means for storing the feature data generated by the feature data generation means as search feature data when the search key input instruction means is instructed to input a search key;
A search means for comparing the feature data for search in the feature data storage means for search with the feature data in the feature data storage means and searching for feature data that is considered to be matched;
A voice data search apparatus comprising: a reading unit that reads out voice data stored in the voice data storage unit from an address corresponding to time data of feature data searched by the search unit. Sound collection means for outputting sound data corresponding to the collected sound;
Voice data storage means for storing voice data output by the sound pickup means;
Feature data generating means for analyzing the voice data output by the sound collecting means and generating feature data indicating the characteristics;
Feature data storage means for storing the feature data generated by the feature data generation means together with time data corresponding to the generation time;
A character string input means for inputting a character string;
A table in which phonemes that are constituent elements of a character string are associated with voice feature data when the phonemes are pronounced;
Conversion means for converting each character of the character string input by the character string input means into feature data with reference to the table;
Search feature data storage means for storing the feature data converted by the conversion means as search feature data;
A search means for comparing the feature data for search in the feature data storage means for search with the feature data in the feature data storage means and searching for feature data that is considered to be matched;
A voice data search apparatus comprising: a reading unit that reads out voice data stored in the voice data storage unit from an address corresponding to time data of feature data searched by the search unit. The sound collecting means generates a plurality of microphones and sound data corresponding to the sound picked up by each microphone, and adds identification data for identifying which microphone the sound data is from. Audio data generating means for attaching to audio data;
The feature data storage means classifies and stores the feature data based on the identification data,
The search means, when the identification data is specified, compares the feature data classified by the specified identification data with the search feature data in the search feature data storage means. Item 3. The speech data retrieval apparatus according to item 1 or 2. The sound collection means includes an array microphone whose sound collection direction is variable, a sound collection direction control means for controlling the sound collection direction of the array microphone and outputting direction data indicating the sound collection direction, and the array microphone. Voice data generating means for generating voice data corresponding to the sound that has been sounded,
The feature data storage means classifies and stores the feature data based on the direction data,
The search means, when the direction data is specified, compares the feature data classified by the specified direction data with the search feature data in the search feature data storage means. Item 3. The speech data retrieval apparatus according to item 1 or 2.

Images