JP6569343B2 - Voice search device, voice search method and program - Google Patents

Description

  The present invention relates to a voice search device, a voice search method, and a program.

  A voice search technique called a two-pass search method for searching a voice representing a query from a recorded voice is known.

  In the two-pass search method, the voice search is performed on the assumption that the voice speed of the recorded voice substantially matches the predetermined average voice speed. For this reason, the two-pass search method has a problem that the search accuracy decreases as the speech speed of the recorded voice and the average speech speed differ.

  As a method for solving this problem, it is conceivable to estimate the speech speed of the recorded speech and perform a speech search using the estimated speech speed. As a technique for estimating the speech speed, for example, a technique described in Patent Document 1 is known.

JP 2010-26323 A

  In the technique of Patent Document 1, there is a case where the speech speed of the recorded voice is erroneously estimated due to noise or the like.

  When a voice search is performed based on the erroneously estimated speech speed of the recorded voice and an inappropriate search result is presented, the user selects an appropriate speech speed (appropriate search result) as the speech speed used for the voice search. It is necessary to manually set the speech speed that matches the speech speed of the recorded voice).

  However, in the technique of Patent Document 1, there is no way for the user to intuitively grasp the speech speed used for the voice search. For this reason, the user cannot intuitively set an appropriate speech speed as the speech speed used for the voice search.

  The present invention has been made in view of the above problems, and provides a voice search device, a voice search method, and a program that perform an accurate voice search using an appropriate speech speed intuitively designated by a user. For the purpose.

In order to achieve the above object, a voice search device according to the present invention provides:
A query acquisition unit for acquiring a query;
A speech rate designation unit for accepting a speech rate designated by the user from a predetermined range of speech rate;
A demo voice output unit that outputs a demo voice representing the acquired query at a speech speed accepted by the speech speed designation unit;
Based on the speech speed received by the speech speed designating unit, a search unit that performs a voice search from the recorded voice that is the target of the voice search for the voice representing the query;
It is characterized by providing.

  According to the present invention, it is possible to provide a voice search device, a voice search method, and a program that perform an accurate voice search using an appropriate speech speed that is intuitively designated by a user.

It is a figure showing an example of composition of a voice search device concerning an embodiment of the present invention. (A)-(c) is a figure which shows the example of the search screen which the audio | voice search apparatus which concerns on embodiment of this invention displays. (A) is a figure showing an example of a search screen. (B) is a figure showing another example of a search screen. (C) is a figure which shows another example of a search screen. It is a figure for demonstrating the relationship between the waveform of a sound recording, a likelihood acquisition area, and a frame. It is a flowchart for demonstrating the voice search process which the voice search apparatus which concerns on embodiment of this invention performs. It is a flowchart for demonstrating the speech speed designation | designated process which the speech search device which concerns on embodiment of this invention performs. It is a flowchart for demonstrating the audio | voice output process which the audio | voice search device which concerns on embodiment of this invention performs. It is a figure which shows the structural example of the speech search device which concerns on the modification of embodiment of this invention.

  Hereinafter, functions and operations of the voice search device according to the embodiment of the present invention will be described with reference to the drawings. In the drawing, the same or equivalent parts are denoted by the same reference numerals.

  The voice search device uses a two-pass search technique to search for voices representing queries from recorded voices. Details of the two-pass search method will be described later.

  As shown in FIG. 1, the voice search apparatus 100 includes a ROM (Read-Only Memory) 10, a RAM (Random Access Memory) 20, an external storage unit 30, an input unit 40, an output unit 50, a CPU ( Central Processing Unit) 60.

  The ROM 10 permanently stores various programs including a voice search program. Details of the voice search program will be described later.

  The RAM 20 temporarily stores data and programs. The RAM 20 functions as a work memory for the CPU 60.

  The external storage unit 30 stores data in a fixed manner. The external storage unit 30 includes, for example, a hard disk. Specifically, the external storage unit 30 of the present embodiment acquires and stores a recorded voice that is a target of voice search from the outside in advance. The external storage unit 30 includes a monophone model DB (Data Base) 31, a triphone model DB 32, a time length DB 33, a prosody DB 34, and a speech DB 35.

  The monophone model DB 31 stores a monophone model. The monophone model is an acoustic model generated for each phoneme (monophone). The acoustic model is obtained by modeling the waveform feature amount (for example, frequency characteristics) of each phoneme. The monophone model is an acoustic model that does not depend on adjacent phonemes, that is, a state transition between the phoneme states before and after is fixed.

  The triphone model DB 32 stores a triphone model. The triphone model is an acoustic model generated for every three phonemes (triphones), and is an acoustic model that depends on adjacent phonemes, that is, takes into account state transitions with previous and subsequent phoneme states.

  The external storage unit 30 acquires a monophone model and a triphone model generated using a known technique from the outside in advance, and stores them in the monophone model DB 31 and the triphone model DB 32, respectively.

  The time length DB 33 stores the average length of each phoneme. The average length of each phoneme is the time required for the speaker to utter each phoneme at the average speech speed. The external storage unit 30 previously acquires the average length of each phoneme acquired using a known technique from the outside and stores it in the time length DB 33.

  The prosody DB 34 stores the prosody of each phoneme (pitch transition, phoneme length, and power) in association with the phoneme string including this phoneme. The external storage unit 30 previously acquires the prosody of each phoneme acquired using any known technique from the outside and stores it in the prosody DB 34.

  The speech DB 35 stores speech segments representing each phoneme. The external storage unit 30 acquires a speech unit acquired using a known technique from the outside in advance and stores it in the speech DB 35.

  The input unit 40 receives input of various data by the user. Specifically, the input unit 40 receives an input of a query or an operation instruction by the user. The input unit 40 supplies various input data to the CPU 60. The input unit 40 includes, for example, a keyboard and a microphone.

  The output unit 50 outputs various data to the outside. The output unit 50 includes, for example, a display and a speaker, outputs data as sound through the speaker, and outputs data as an image through the display.

  The sound output (sound output) by the output unit 50 includes recorded sound and demo sound described later.

  The image output (image output) by the display of the output unit 50 includes a search screen Wn shown in FIGS. The voice search device 100 sequentially displays the search screen Wn shown in FIGS. 2A to 2C on the display of the output unit 50 (outputs an image) when executing a voice search process to be described later. FIG. 2A shows a search screen Wn displayed by the voice search device 100 when the user turns on the power of the voice search device 100. FIG. 2B shows a search screen Wn displayed by the voice search device 100 when a result of the first pass search described later is presented to the user. FIG. 2C shows a search screen Wn displayed by the voice search device 100 when a result of a second pass search described later is presented to the user. Details of the voice search processing, the first pass search, and the second pass search will be described later.

  The CPU 60 executes various programs stored in the ROM 10. Specifically, the CPU 60 executes a voice search program stored in the ROM 10 to functionally provide the speech speed designation unit 101, the voice output unit 102, the query acquisition unit 103, the search unit 104, and the presentation shown in FIG. The unit 105 is realized.

  The speech speed designation unit 101 accepts designation by the user of the speech speed used for voice search. Specifically, the speech speed designation unit 101 receives the speech speed designated by the user from the speech speed in a predetermined variable range.

  In the present embodiment, the user designates the speech speed by designating a speech speed parameter that is a magnification with respect to the average speech speed. For example, when the user designates “2” as the speech speed parameter (magnification with respect to the average speech speed) (designates “2 times” as the magnification with respect to the average speech speed), the speech speed designation unit 101 doubles the average speech speed. Is accepted as the designated speech speed. In the present embodiment, the speech speed designating unit 101 accepts designation of speech speed parameters belonging to the range from “0.5” to “2.0” (the user passes the speech speed designating unit 101 via The speech speeds belonging to the range (predetermined variable range) from 0.5 times the average speech speed to 2.0 times the average speech speed can be specified.

  Specifically, the speech speed designation unit 101 accepts designation of the speech speed in a state where the display of the output unit 50 displays the search screen Wn shown in FIGS. As shown in FIGS. 2A to 2C, the search screen Wn includes a speech speed seek bar SB. The user designates the speech speed parameter by moving the speech speed slider SS on the speech speed seek bar SB via the input unit 40. As shown in FIGS. 2A to 2C, the upper limit value of the speech speed parameter that can be specified in the speech speed seek bar SB is “2.0”, and the lower limit value is “0.5”. The user arranges the speech speed slider SS at a position corresponding to a desired speech speed parameter, and then selects the search icon RS or the demo sound reproduction icon P2 on the search screen Wn. When detecting the selection of the search icon RS or the demo sound reproduction icon P2 by the user, the speech speed specifying unit 101 acquires a speech speed parameter corresponding to the position of the speech speed slider SS at the time of detection. The speech speed designation unit 101 accepts the speech speed obtained by multiplying the acquired speech speed parameter by the average speech speed as the speech speed designated by the user.

  The CPU 60 realizes the speech speed designation unit 101 by controlling the input unit 40.

  The voice output unit 102 outputs a demo voice having a voice speed (speak speed used for voice search) designated by the user via the voice speed designation unit 101. In this embodiment, the audio output unit 102 includes a demo audio output unit 102a and a recorded audio output unit 102b as shown in FIG. Output audio in a comparable manner. The CPU 60 realizes the audio output unit 102 by controlling the speaker of the output unit 50.

  Specifically, the demo voice output unit 102a outputs the demo voice representing the query acquired by the query acquisition unit 103 (to be described later) at the speech speed received by the speech speed specification unit 101 (via the speech speed specification unit 101). Demo voices having a speech speed designated by the user and representing a query acquired by the query acquisition unit 103 described later are output as speech). The demo voice output unit 102a outputs the demo voice in a manner that allows the user to compare the voice speed of the demo voice with the voice speed of the recorded voice that is the target of the voice search.

More specifically, the demo voice output unit 102a sequentially executes the following processes (A1) to (A4), thereby generating a demo voice representing a query having a speech speed designated by the user.
(A1) The prosody (pitch transition, phoneme length, and power) of each phoneme included in the phoneme string representing the query is acquired from the prosody DB 34.
(A2) Among the prosody obtained in the processing of (A1), the phoneme length of each phoneme is corrected to the time length obtained by dividing the average phoneme length by the speech speed parameter designated by the user via the speech speed designating unit 101. .
(A3) The speech segment of each phoneme included in the phoneme string representing the query is acquired from the speech DB 35.
(A4) A demo speech is generated by adjusting the pitch transition, phoneme length, and power of the speech segment acquired in the processing of (A3) so as to match the prosody corrected in the processing of (A2). .

The recorded voice output unit 102b outputs the recorded voice that is the target of the voice search. Specifically, the recorded voice output unit 102b outputs the recorded voice by the following method (B1) or (B2), so that the user can determine the speech speed of the demo voice and the voice speed of the recorded voice. Allows comparison.
(B1) The recorded voice output unit 102b starts outputting voice of the recorded voice from an arbitrary reproduction position designated by the user via the input unit 40 (outputs the recorded part designated by the user out of the recorded voice).
(B2) The recorded voice output unit 102b outputs the recorded voice corresponding to the result designated by the user among the results of the voice search performed by the search unit 104 described later (the search unit 104 described later among the recorded voices). The recorded part obtained by voice search is output as voice).

  Hereinafter, the details of the methods (B1) and (B2) will be described.

  In the method (B1), the user moves the recording voice slider VS included in the search screen Wn shown in FIGS. 2 (a) to 2 (c) by moving the recording voice seek bar VB via the input unit 40. Specify the audio playback position. The user selects the recording sound reproduction icon P1 in the search screen Wn after arranging the recording sound slider VS at a position on the recording sound seek bar VB corresponding to a desired reproduction position. When the recording sound output unit 102b detects the selection of the recording sound reproduction icon P1 by the user, the recording sound output unit 102b starts outputting the recording sound from a position in the recording sound corresponding to the position of the recording sound slider VS at the time of detection. The user can compare the speech speeds by listening to and comparing the demo sound and the recorded sound at the designated playback position (the recorded portion of the recorded sound designated by the user).

  In the method (B2), the user selects any one of the first pass result icons IA1 to IA5 included in FIGS. 2B and 2C, or the second pass result icons IB1 to IB5 included in FIG. Is selected via the input unit 40, and the result of the voice search executed by the search unit 104 is designated. The search unit 104 specifies a section estimated to include a query in the recorded voice by performing a voice search. The first pass result icons IA1 to IA5 and the second pass result icons IB1 to IB5 are all displayed on the search screen Wn by the presentation unit 105 described later. The first-pass result icons IA1 to IA5 and the second-pass result icons IB1 to IB5 respectively represent sections in the recorded voice specified by the search unit 104 by voice search. When the recording voice output unit 102b detects that the user has selected one of the first pass result icons IA1 to IA5 or the second pass result icons IB1 to IB5, the recorded voice (sound recording) of the section represented by the selected icon is recorded. Among the voices, a recording part obtained by voice search by a search unit 104 (to be described later) outputs the voice. Details of the first pass result icons IA1 to IA5 and the second pass result icons IB1 to IB5 will be described later. The user can compare the speaking speeds of the demo voice and the recorded voice corresponding to the result of the voice search by comparing them.

  According to the method (B2), the user can compare the demo voice representing the query with the recorded voice of the section estimated to include the voice representing the query acquired as a result of the voice search. it can. For this reason, the user can compare the speech speed more accurately as compared with the method (B1) of listening and comparing the demo voice and the recorded voice at an arbitrary reproduction position. On the other hand, according to the method (B1), unlike the method (B2), the speech speed of the demo voice and the recorded voice is not obtained when the voice search process is not executed and the search result is not yet acquired. Can be compared.

  The query acquisition unit 103 acquires a query.

  Based on the speech speed designated by the user via the speech speed designating unit 101 (speech speed accepted by the speech speed designating unit 101), the search unit 104 extracts the speech representing the query from the recorded speech that is the target of the speech search. Search by voice. Specifically, the search unit 104 includes a first search unit 104a and a second search unit 104b, and two sections (estimated sections) that are estimated to include a voice representing a query in the recorded voice are included in two passes. Identify using search techniques.

  In the two-pass search method, the search unit 104 executes a voice search by dividing it into two steps. In the first-stage process, the first search unit 104a selects a candidate for the estimation section by performing a voice search based on a monophone model (an acoustic model that does not depend on adjacent phonemes). In the second-stage process, the second search unit 104b performs a voice search based on the triphone model (an acoustic model that depends on adjacent phonemes), so that the estimated section selected in the first-stage voice search Identify the estimated interval from the candidates. Hereinafter, for the sake of explanation, the voice search performed by the first search unit 104a is referred to as a first pass search, and the voice search performed by the second search unit 104b is referred to as a second pass search.

  The first-pass search based on the monophone model is less accurate than the second-pass search based on the triphone model, which has a larger amount of information than the monophone model, but requires a small amount of calculation and results in obtaining a search result. Short time (fast search). The second-pass search is more computationally intensive than the first-pass search, and takes a long time to obtain a search result (slow search speed), but has high accuracy. The search unit 104 performs a high-accuracy second-pass search for the estimated section candidate selected by the first-pass search with a high search speed, thereby suppressing the amount of calculation and performing a high-speed and high-accuracy voice search. Execute.

Specifically, the first search unit 104a sequentially executes the following processes (C1) to (C7) in the first pass search.
(C1) The query is converted into a monophone phoneme string.
(C2) The average length of each phoneme included in the monophone phoneme string is acquired.
(C3) Based on the average length of each phoneme, the utterance time length of the voice with the average speech speed representing the query is acquired.
(C4) An utterance time length of speech at a speech speed designated by a user representing a query is acquired.
(C5) A likelihood acquisition interval is set.
(C6) A first likelihood for each likelihood acquisition section is acquired based on the monophone phoneme string.
(C7) Based on the first likelihood, an estimation interval candidate is selected from the likelihood acquisition intervals.

  Hereinafter, the details of the processes (C1) to (C7) will be described.

  In the process of (C1), the first search unit 104a converts the query into a monophone phoneme string by arranging the phonemes of the monophone model according to the query (in the same order as the order in the query of the characters corresponding to each phoneme). To do.

  In the process (C2), the first search unit 104a acquires the average length of each phoneme included in the monophone phoneme string converted in the process (C1) from the time length DB 33.

  In the process (C3), the first search unit 104a acquires the speech duration of the average speech speed voice representing the query by adding the average lengths of the phonemes acquired in the process (C2).

  In the process of (C4), the first search unit 104a uses the speech speed designated by the user via the speech speed designating unit 101 to determine the speech duration of the average speech speed representing the query acquired in the process of (C3). By dividing by the parameter, the utterance time length of the voice at the speech speed designated by the user representing the query is acquired.

  In the process of (C5), as shown in FIG. 3, the first search unit 104a performs P pieces from the 0th likelihood acquisition section to the (P-1) likelihood acquisition section on the time axis of the recorded voice. Set the likelihood acquisition interval. Each likelihood acquisition section has, as the time length L, the utterance time length of the speech at the speaking speed designated by the user representing the query, acquired in the process of (C4). The start positions of the respective likelihood acquisition sections are set apart by a predetermined shift length S. Details of the shift length S will be described later. In FIG. 3, the horizontal axis represents time, and the vertical axis represents the size of the waveform WV of the recorded voice. In FIG. 3, the waveform WV of the recorded voice has a time length T.

  In the process (C6), the first search unit 104a acquires the first likelihood for each likelihood acquisition section set in the process (C5). The first likelihood is an index indicating the likelihood that the likelihood acquisition section acquired based on the monophone model is a section including a voice representing a query.

Specifically, the first search unit 104a acquires the first likelihood of each likelihood acquisition section by sequentially executing the following processes (C6-1) to (C6-5).
(C6-1) A frame is set.
(C6-2) The feature amount of the recorded voice is acquired for each frame.
(C6-3) An output probability is calculated for each frame based on the monophone model.
(C6-4) The output probability is made lower-bound.
(C6-5) The first likelihood is calculated.

  Hereinafter, the details of the processes (C6-1) to (C6-5) will be described.

  In the process of (C6-1), as shown in FIG. 3, the first search unit 104a includes N pieces of frames from the 0th frame to the (N−1) th likelihood acquisition section on the time axis of the recorded voice. Set the frame (time window). Each frame has a predetermined frame length F as a time length. The start positions of the respective frames are set apart by a shift length S. The frame length F and the shift length S are set when the acoustic model is created. The frame length F is set longer than the shift length S so that each frame overlaps with an adjacent frame.

  Each likelihood acquisition section set in the process of (C5) includes a plurality of frames set in (C6-1). For example, the 0th likelihood acquisition section includes M frames from the 0th frame to the (M−1) th frame. Hereinafter, the first search unit 104a executes the processes (C6-2) to (C6-5), and the recorded speech in one likelihood acquisition section and the monophone phoneme sequence acquired in the process (C1) Are compared in units of frames, and the first likelihood of this likelihood acquisition section is acquired.

  In the process of (C6-2), the first search unit 104a acquires the feature amount of the recorded voice of each frame included in one likelihood acquisition section. This feature amount is obtained by, for example, calculating a frequency axis characteristic parameter obtained by converting speech data called cepstrum or mel cepstrum on the frequency axis, and a power system obtained by calculating the sum of squares of energy of the voice data or its logarithm It is obtained by combining feature parameters. For example, the feature amount is obtained by taking a difference between a frequency axis system characteristic parameter 12 component (12 dimensions), a power system characteristic parameter 1 component (1 dimension), and each component of the immediately preceding time window, that is, a Δ frequency axis system feature. The difference between the parameter 12 component (12 dimensions), the Δ power system characteristic parameter 1 component (1 dimension), and each component of the immediately preceding time window, that is, the ΔΔ frequency axis system characteristic parameter 12 component (12 Dimensional) and a 38-dimensional vector quantity having a total of 38 components.

  In the process (C6-3), the first search unit 104a outputs the feature value from each phoneme included in the monophone phoneme string based on the feature value of the recorded voice acquired in the process (C6-2). Output probability is calculated for each frame. Specifically, the first search unit 104a compares the feature amount of the recorded voice in each frame with the monophone model in a state corresponding to this frame among the phoneme states included in the monophone phoneme string. Then, the probability (output probability) that the feature value in each frame is output from each phoneme included in the monophone phoneme string is calculated. This output probability is represented by a normal mixed continuous distribution obtained by adding a plurality of Gaussian distributions with weights.

  In the process of (C6-4), the first search unit 104a replaces the output probability of each frame calculated in the process of (C6-3) with the maximum output probability among the output probabilities of n frames before and after the frame. To do. This process is called lower-bound conversion. In the present embodiment, n is set to the number of frames corresponding to 100 msec. By changing the output probability to lower-bound, the change in the output probability in the time direction is reduced. Therefore, by converting the output probability to Lower-Bound, an error between the average length of each phoneme stored in the time length DB 33 and the actual duration of each phoneme can be absorbed within the range of the preceding and following n frames. Further, by converting the output probability to Lower-Bound, the speech duration of the speech speed specified by the user representing the query and the actual speech duration of the recorded speech representing the query, acquired in the process of (C4), Can be absorbed within the range of the preceding and following n frames.

  In the process of (C6-5), the first search unit 104a obtains the likelihood including these frames by adding the output probabilities of each frame converted to the lower-bound in (C6-4) on the logarithmic axis. The first likelihood of the section is acquired.

  The 1st search part 104a acquires the 1st likelihood of one likelihood acquisition area by performing the process of (C6-2)-(C6-5). The first search unit 104a performs the processes of (C6-2) to (C6-5) for each likelihood acquisition section set in the process of (C5) above, so that the first search section 104a 1 likelihood is acquired.

  In the process (C7), the first search unit 104a sets a plurality of selection sections on the time axis of the recorded voice. Each selection section has a predetermined selection time length as a time length. The first search unit 104a first selects a likelihood acquisition section having a large first likelihood acquired in the process (C6) from a plurality of likelihood acquisition sections whose start positions are included in each selection section. Next, the first search unit 104a selects a predetermined number (5 in the present embodiment) of likelihood acquisition sections with the highest likelihood among the selected plurality of likelihood acquisition sections as estimation section candidates. To do.

  As shown in FIG. 3, the likelihood acquisition sections set in the process (C5) are set to overlap each other. For this reason, a section in a single recorded voice having a high likelihood may belong to a plurality of likelihood acquisition sections. In this case, when the first search unit 104a selects a predetermined number of likelihood acquisition sections with the highest likelihood from the entire recorded voice as candidates for the estimated section, the candidate for the estimated section is selected from only a part of the recorded voice. On the other hand, there is a possibility that a section having a high likelihood in another part of the recorded voice may be overlooked. Therefore, the first search unit 104a first selects one likelihood acquisition section from each selection section, and selects a predetermined number of likelihood acquisition sections with the highest likelihood among the selected likelihood acquisition sections as candidates for estimation sections. Select as. Thereby, the 1st search part 104a can select the candidate of an estimation area from the whole recorded audio | voice without bias.

  The selection time length (the time length of the selection section) is set to a time shorter than the time length L of the likelihood acquisition section. For example, a value (L / k) obtained by dividing the time length L of the likelihood acquisition section by a predetermined constant k is set as the selected time length.

  As described above, the first search unit 104a executes the processing of (C1) to (C7), and thereby the speech speed designated by the user via the speech speed designating unit 101 and the monophone model (adjacent to each other). Based on the phoneme-based acoustic model), a candidate for an estimated section (a section in the recorded voice that is estimated to include a voice representing a query) is selected from the recorded voice.

  The first search unit 104a supplies information indicating the position of the estimated section candidate selected in the process of (C7) in the recorded voice to the presentation unit 105 as information indicating the result of the first pass search.

In the second pass search, the second search unit 104b sequentially executes the following processes (D1) to (D3) for the estimation section candidates selected by the first search unit 104a.
(D1) The query is converted into a triphone phoneme string.
(D2) Based on the triphone phoneme string, a second likelihood is obtained for each estimation section candidate.
(D3) Based on the second likelihood, an estimation section is specified from the estimation section candidates.

  In the process of (D1), the second search unit 104b converts the query into a triphone phoneme string by arranging phonemes of the triphone model according to the query.

  In the process of (D2), the second search unit 104b acquires the second likelihood for each likelihood acquisition section selected as the estimation section candidate in (C7). The second likelihood is an index indicating the likelihood that the likelihood acquisition section acquired based on the triphone model is a section including a voice representing a query. The second likelihood is acquired based on the triphone model having a larger amount of information than the monophone model. For this reason, the second likelihood is an index (likelihood) with higher accuracy than the first likelihood acquired based on the monophone model.

Specifically, the second search unit 104b performs the following processes (D2-1) and (D2-2) to obtain the second likelihood of one estimation section candidate (likelihood acquisition section). get. The second search unit 104b executes the processes of (D2-1) and (D2-2) in order for each likelihood acquisition section selected in the process of (C7). Get two likelihoods.
(D2-1) The output probability is calculated for each frame based on the triphone model.
(D2-2) The second likelihood is acquired based on the output probability.

  Hereinafter, details of each processing of (D2-1) and (D2-2) will be described.

  In the process of (D2-1), the second search unit 104b outputs the output of each of the phonemes included in the triphone phoneme sequence acquired in the process of (D1). Probability is calculated for each frame. Specifically, the second search unit 104b compares the feature amount of the recorded voice in each frame acquired in the process (C6-2) with the model of each triphone included in the triphone phoneme string. Then, the probability that the feature value in each frame is output from each triphone is calculated.

  In the process of (D2-2), the second search unit 104b uses each triphone included in each frame and triphone phoneme string in the likelihood acquisition section based on the output probability calculated in the process of (D2-1). Is searched for by DP (Dynamic Programming) matching. And the 2nd likelihood of this likelihood acquisition area is acquired by adding the logarithm of the output probability acquired about each of the triphone matched with each frame in a likelihood acquisition area.

  In the process of (D3), the second search unit 104b uses a predetermined number (5 in this embodiment) of likelihood acquisition sections with the largest second likelihood acquired in the process of (D2) as estimation sections. Identify.

  As described above, the second search unit 104b executes the processes (D1) to (D3), and thus based on the triphone model (acoustic model depending on adjacent phonemes), the first search unit 104a. The estimation section is identified from the estimation section candidates selected by.

  The second search unit 104b supplies the position information in the recorded voice of the estimated section specified in the process (D3) to the presentation unit 105 as information indicating the result of the second pass search.

  Returning to FIG. 1, the presentation unit 105 presents the result of the voice search performed by the search unit 104 to the user.

  In the present embodiment, the presentation unit 105 displays icons indicating the results of the first pass search and the second pass search performed by the search unit 104 on the display included in the output unit 50, and displays these search results as a user. To present.

  Specifically, as shown in FIGS. 2B and 2C, the presentation unit 105 displays first pass result icons IA1 to IA5 in the search screen Wn. The first pass result icons IA <b> 1 to IA <b> 5 are icons respectively representing estimated section candidates selected by the first search unit 104 a through the first pass search. The presenting unit 105 supplies the first pass result icons IA1 to IA5 based on the information (information indicating the position of the candidate of the estimated section in the recorded voice) indicating the result of the first pass search supplied from the first search unit 104a. Is displayed. That is, the presentation unit 105 presents the first pass search result to the user by displaying the first pass result icons IA1 to IA5 in the search screen Wn.

  Further, as shown in FIG. 2C, the presentation unit 105 displays second pass result icons IB1 to IB5 in the search screen Wn. The second pass result icons IB1 to IB5 are icons respectively representing estimated sections specified by the second search unit 104b by the second pass search. The presentation unit 105 displays the second pass result icons IB1 to IB5 based on the information (information indicating the position in the recorded voice of the estimated section) indicating the result of the second pass search supplied from the second search unit 104b. To do. That is, the presentation unit 105 presents the second pass search result to the user by displaying the second pass result icons IB1 to IB5 in the search screen Wn.

  The CPU 60 realizes the presentation unit 105 by controlling the output unit 50.

  Hereinafter, the voice search processing executed by the voice search device 100 having the above-described physical / functional configuration will be described with reference to the flowchart shown in FIG.

  The voice search device 100 acquires a recorded voice that is a target of voice search from the outside in advance and stores it in the external storage unit 30.

  Further, the speech search apparatus 100 previously obtains a monophone model, a triphone model, an average length of each phoneme, a prosody of each phoneme, and a phoneme unit of each phoneme obtained from any known technique from an external storage. Stored in the unit 30.

  When the user turns on the power of the voice search device 100, the voice search device 100 displays the search screen Wn of FIG. 2A on the display provided in the output unit 50.

  The user arranges the speech speed slider SS at a position corresponding to the speech speed (speech speed parameter) desired to be used for the speech search via the input unit 40 in the search screen Wn of FIG. . In FIG. 2A, the speech speed slider SS is disposed at a position corresponding to a speech speed 1.3 times the average speech speed (position corresponding to the speech speed parameter “1.3”).

  The user inputs the query text via the keyboard provided in the input unit 40. In FIG. 2A, the character string “ramen” is input as a query.

  When the user selects the search icon RS in the search screen Wn via the input unit 40, in response to this, the CPU 60 of the voice search device 100 starts the voice search process shown in the flowchart of FIG.

  When the voice search process is started, the speech speed designation unit 101 first acquires a speech speed parameter corresponding to the position of the speech speed slider SS at the time when the search icon RS is selected (acquires the speech speed) (step S101). ). Next, the query acquisition unit 103 acquires a query input by the user (step S102).

  The first search unit 104a converts the query acquired in step S102 into a monophone phoneme string (step S103). The first search unit 104a acquires the average speech length of the monophone phoneme sequence acquired in step S103 from the time length DB 33, thereby acquiring the speech duration of the average speech speed representing the query (step S104).

  The first search unit 104a divides the utterance time length acquired in step S104 by the speech speed parameter acquired in step S101, thereby acquiring the utterance time length of the speech with the speech speed specified by the user representing the query (step S105).

  The first search unit 104a sets a plurality of likelihood acquisition sections having the utterance time length acquired in step S105 as the time length L on the time axis of the recorded voice (step S106).

  Next, the first search unit 104a sets a plurality of frames having a predetermined frame length F as the time length on the time axis of the recorded voice (step S107).

  The first search unit 104a acquires the feature amount of the recorded voice for each frame set in step S107 (step S108).

  Based on the feature amount acquired in step S108 and the corresponding monophone model, the first search unit 104a uses the feature probability output from each phoneme included in the phoneme sequence acquired in step S103 as a frame. It is calculated every time (step S109).

  The first search unit 104a converts the output probability calculated for each frame in Step S109 into Lower-Bound (replaces it with the maximum output probability in n frames before and after each frame) (Step S110).

  For each likelihood acquisition section set in step S106, the first search unit 104a calculates the sum on the logarithmic axis of the output probabilities that are lower-bounded by the processing in step S110 of all frames included in each likelihood acquisition section. As a result, the first likelihood of each likelihood acquisition section is acquired (step S111).

  The first search unit 104a sets a plurality of selection sections having a predetermined selection time length as the time length on the time axis of the recorded voice (step S112).

  For each selection section set in step S112, the first search unit 104a acquires the likelihood having the largest first likelihood acquired in step S111 among a plurality of likelihood acquisition sections in which the start position is included in each selection section. A section is selected (step S113).

  The first search unit 104a selects a predetermined number (5 in this embodiment) of likelihood acquisition intervals having the largest first likelihood acquired in step S111 among the plurality of likelihood acquisition intervals selected in step S113. Then, it is selected as a candidate for the estimated section (step S114). The first search unit 104a supplies information indicating the position of the selected estimated section candidate in the recorded voice to the presentation unit 105.

  The presenting unit 105 includes the first-pass result icons IA1 to IA5 representing the estimation section candidates selected in step S114 based on the information supplied from the first search unit 104a, and the search screen Wn illustrated in FIG. Is displayed on the display of the output unit 50 (the result of the first pass search is presented to the user) (step S115).

  The second search unit 104b converts the query acquired in step S102 into a triphone phoneme string (step S116).

  The second search unit 104b calculates an output probability for each frame set in step S107, based on the triphone phoneme sequence acquired in step S116 (step S117).

  The second search unit 104b obtains the second likelihood of each estimation section candidate by performing DP matching based on the output probability calculated in step S117 for each of the estimation section candidates selected in step S114. (Step S118).

  The second search unit 104b estimates a predetermined number (5 in this embodiment) of estimation section candidates having the largest second likelihood acquired in step S118 among the estimation section candidates selected in step S114. The section is specified (step S119). The second search unit 104b supplies information indicating the position of the identified estimated section in the recorded voice to the presentation unit 105.

  The presenting unit 105 displays the search screen Wn shown in FIG. 2C including the second pass result icons IB1 to IB5 representing the estimated section identified in step S119 based on the information supplied from the second search unit 104b. (The result of the second pass search is presented to the user) (step S120).

  Next, the CPU 60 determines whether or not the speech speed change icon CS on the search screen Wn has been selected (step S121). If it is determined that the speech speed change icon CS is not selected (step S121; NO), the CPU 60 determines whether or not the end icon ES in the search screen Wn is selected (step S122). If it is determined that the speech speed change icon CS is not selected (step S121; NO) and the end icon ES is not selected (step S122; NO), the process returns to step S121. The CPU 60 repeats the processes of steps S121 to S122 until either the speech speed change icon CS or the end icon ES is selected.

  After the result of the second pass search (estimated section) is presented in step S120, the user determines whether the presented search result is accurate by listening to the recorded voice. Specifically, the user determines whether or not all sections in the recorded voice including the voice representing the query are specified as estimation candidates (whether there is no omission of search) or by listening to all the recorded voice. In addition, the user determines whether or not the recorded voices of all the sections specified as the estimated sections represent queries (whether there is no erroneous detection) and listens to the recorded voices of all the estimated sections.

  As a result of listening to the recorded voice, if the user determines that the search result is inaccurate (there is a detection failure or a false detection), the speech speed used for the voice search in the previous steps S101 to S120 is inappropriate. Is likely. In this case, the user selects the speech speed change icon CS in the search screen Wn via the input unit 40 in order to manually correct the speech speed used for the voice search. In response to this, the CPU 60 determines that the speech speed change icon CS has been selected (step S121; YES), and starts the speech speed designation process (step S123).

  Hereinafter, details of the speech speed designation processing in step S123 will be described with reference to the flowchart of FIG. 5 is displayed on the display of the output unit 50 during the speech speed designation process in the flowchart of FIG.

  When the speech speed designation process is started, first, the recorded voice output unit 102b determines whether or not the recorded voice playback icon P1 has been selected (step S201). When the user desires to designate the playback position of the recorded voice, the user designates the playback position of the recorded voice by moving the recording voice slider VS on the recorded voice seek bar VB, and selects the recorded voice playback icon P1. In response to this, the recording sound output unit 102b determines that the recording sound reproduction icon P1 is selected (step S201; YES), and outputs the sound of the recording sound from the position corresponding to the position of the recording sound slider VS at that time. (Reproduction) is started (step S203).

  If it is determined that the recorded sound reproduction icon P1 is not selected (step S201; NO), the recorded sound output unit 102b selects either the first pass result icons IA1 to IA5 or the second pass result icons IB1 to IB5. It is determined whether or not (step S202). When the user desires to output the recorded sound corresponding to the search result, the user selects any icon. In response to this, the audio recording output unit 102b determines that any one of the first pass result icons IA1 to IA5 or the second pass result icons IB1 to IB5 has been selected (step S202; YES), and the selected icon The recorded voice of the section corresponding to the search result represented by is output as a voice (step S203).

  If it is determined that no icon is selected (step S202; NO), the process returns to step S201. The recorded sound output unit 102b performs the processing of steps S201 and S202 until the playback position of the recorded sound is specified or the search result is specified (until determined as YES in either step S201 or step S202). repeat.

  After the reproduction of the recorded voice (step S203) is finished, the recorded voice output unit 102b determines whether or not a second voice output of the recorded voice is instructed (step S204). When the user determines that it is necessary to listen to the recorded voice again, the user selects the recording voice reproduction icon P1, the first pass result icons IA1 to IA5, or the second pass result icons IB1 to IB5, thereby recording. Instructs audio output again. In response to this, the recorded voice output unit 102b determines that another voice output has been instructed (step S204; YES), and the process returns to step S201.

  If it is determined that the second sound output of the recorded sound is not instructed (step S204; NO), the demo sound output unit 102a determines whether or not the demo sound reproduction icon P2 is selected (step S205). If it is determined that the demo sound reproduction icon P2 is not selected (step S205; NO), the process returns to step S204. The sound output unit 102 performs steps S204 and S205 until a sound output of the recorded sound is instructed again or a sound output of the demo sound is instructed (YES in step S204 or step S205). Repeat the process.

  The user who has heard the recorded voice in the processes of steps S201 to S204 changes the voice speed used for the voice search to match the voice speed of the recorded voice by moving the voice speed slider SS. When the change is completed (when the speech speed slider SS is placed at a desired position on the speech speed seek bar SB), the user compares the speech speed after the change with the speech speed of the recorded voice, and the demo voice playback icon P2 By selecting, the audio output of the demo audio is instructed. In response to this, the demo audio output unit 102a determines that the audio output of the demo audio has been instructed (step S205; YES), and outputs the audio of the demo by executing the audio output process (step S206).

  Details of the voice output instruction in step S206 will be described below with reference to the flowchart of FIG.

  When the voice output process is started, the demo voice output unit 102a firstly sets the speech speed parameter corresponding to the position of the speech speed slider SS at the time when YES is determined in step S205 (when the demo voice playback icon P2 is selected). Is acquired (step S301).

  The demo voice output unit 102a acquires the prosody of each phoneme included in the monophone phoneme sequence acquired in step S103 of the flowchart of FIG. 4 from the prosody DB 34 (step S302).

  The demo speech output unit 102a corrects the prosody obtained in step S302 so that the phoneme lengths included in these prosody become values obtained by dividing these phoneme lengths by the speech speed parameter obtained in step S301. (Step S303).

  Next, the demo speech output unit 102a acquires the speech segment of each phoneme included in the monophone phoneme sequence acquired in step S103 from the speech DB 35 (step S304).

  The demo voice output unit 102a generates a demo voice by adjusting the prosody of the speech unit acquired in step S304 to match the prosody corrected in step S303 (step S305).

  The demo sound output unit 102a outputs a sound of the demo sound composed of the speech unit adjusted in step S305 (step S306), and ends the sound output process.

  Returning to FIG. 5, after outputting the demo sound in step S206, the demo sound output unit 102a determines whether or not the demo sound reproduction icon P2 is selected (step S207). As a result of hearing the demo voice in step S206, the user may determine that it is necessary to change the speech speed used for the voice search. In this case, the user readjusts the speech speed by moving the speech speed slider SS. After the readjustment of the speech speed is completed (after the speech speed slider SS is arranged at a desired position on the talk speed seek bar SB), the user compares the speech speed after the readjustment with the speech speed of the recorded voice. By selecting the demo audio reproduction icon P2, the audio output of the demo audio is instructed. In response to this, the demo sound output unit 102a determines that the second sound output of the demo sound has been instructed (step S207; YES), and executes the sound output process of step S206 again.

  In addition, when it is determined that it is necessary to listen to the demo sound again, the user may instruct to output the demo sound again by selecting the demo sound reproduction icon P2 without changing the speech speed. In this case, the demo sound output unit 102a determines that the second sound output of the demo sound has been instructed (step S207; YES), and executes the sound output process of step S206 again. In this case, the demo audio output unit 102a omits the processes of steps S301 to S305 in the flowchart of FIG. 6 and executes only the process of step S306 (demonstration audio output). This is because the user does not change the speaking speed, and the demo voice acquired in the immediately previous voice output process may be output as it is.

  If it is determined that the demo audio playback icon P2 is not selected (step S207; NO), the recorded audio output unit 102b displays the recorded audio playback icon P1, the first pass result icons IA1 to IA5, or the second pass results IB1 to IB5. It is determined whether any of them has been selected (step S208). When the user listens to the demo sound and decides that it is necessary to listen to the recorded sound again and compare it with the demo sound, the user can select one of these icons to instruct the sound output of the recorded sound. To do. In response to this, the recorded sound output unit 102b determines that the second sound output of the recorded sound has been instructed (step S208; YES), the process returns to step S201, and the recorded sound is output again.

  When it is determined that the re-speech output of the recorded voice is not instructed (step S208; NO), the speech speed designation unit 101 determines whether or not the search icon RS in the search screen Wn has been selected (step S209). If it is determined that the search icon RS is not selected (step S209; NO), the process returns to step S207.

  As a result of comparing the recorded voice and the demo voice, the user determines that there is no need to further adjust the speech speed used for the voice search, and by selecting the search icon RS, The voice search device 100 is instructed to perform voice search at a high speed. In response to this, the speech speed designation unit 101 determines that the search icon RS has been selected (step S209; YES), and sets the speech speed (speech speed parameter) corresponding to the position of the speech speed slider SS at that time to the user. Is acquired as the designated speech speed (step S210), and the speech speed designation processing is terminated.

  Returning to FIG. 4, after the speech speed designation processing in step S123 is completed, the processing returns to step S105. As a result of listening to the recorded voice, if the user determines that the search result is accurate (no omissions or false detections), no further search is necessary. In this case, the user instructs the end of the voice search process by selecting the end icon ES via the input unit 40. In response to this, the CPU 60 determines that the end icon ES has been selected (step S122; YES), and ends the voice search process.

  As described above, the voice search device 100 outputs the recorded voice and the speech speed demonstration voice used for the voice search specified by the user. The user can intuitively specify an appropriate speech speed (that is, a speech speed that approximately matches the speech speed of the recorded voice) as the speech speed used for the speech search by comparing the two voices that have been output. The voice search device 100 acquires an accurate search result by executing a voice search using an appropriate speech speed intuitively designated by the user. That is, the voice search device 100 can execute an accurate voice search using an appropriate speech speed that is intuitively designated by the user.

  As mentioned above, although embodiment of this invention was described, these embodiment is an example and the application range of this invention is not restricted to this. That is, the embodiments of the present invention can be applied in various ways, and all the embodiments are included in the scope of the present invention.

  In the above-described embodiment, the voice search device 100 previously acquires and stores the recorded voice that is the target of the voice search from the outside. However, this is only an example, and the voice search device 100 can acquire the recorded voice by an arbitrary method. For example, the voice search device 100 may generate a recorded voice by recording voice through a microphone included in the input unit 40.

  Moreover, in the above-described embodiment, the user inputs a text through a keyboard provided in the input unit 40. However, this is only an example, and the user can enter a query in any way. For example, the user can input a query by voice through a microphone included in the input unit 40.

  In the above-described embodiment, the second search unit 104b performs the second pass search based on the triphone model. However, this is only an example, and the second search unit 104b can perform a second-pass search using an arbitrary acoustic model that depends on adjacent phonemes. For example, the second search unit 104b can perform a second pass search based on the biphone model. The biphone model is an acoustic model generated for every two phonemes (biphone), and is an acoustic model that depends on adjacent phonemes, that is, takes into account the state transition between the previous and subsequent phoneme states.

  In the above-described embodiment, the first search unit 104a selects a predetermined number of likelihood acquisition sections having the largest first likelihood among the likelihood acquisition sections selected from the selection sections as candidates for the estimation section. did. However, this is only an example, and the first search unit 104a can select an estimation section candidate by an arbitrary method based on the first likelihood. For example, the first search unit 104a can select all likelihood acquisition sections whose first likelihood is greater than or equal to a predetermined threshold as candidates for estimation sections.

  In the above-described embodiment, the second search unit 104b identifies a predetermined number of estimation section candidates having the largest second likelihood as estimation sections among the estimation section candidates. However, this is only an example, and the second search unit 104b can specify the estimation interval by an arbitrary method based on the second likelihood. For example, the second search unit 104b can identify all estimation section candidates whose second likelihood is equal to or greater than a predetermined threshold as estimation sections.

  In the above-described embodiment, the voice search device 100 accepts a change in the speech speed after the result of the second pass search is presented to the user (in the flowchart of FIG. The process of S121 was executed). However, this is merely an example, and the voice search device 100 can accept a change in speech speed at an arbitrary timing.

  For example, the voice search device 100 may perform the speech speed designation process shown in the flowchart of FIG. 5 before executing the voice search process shown in the flowchart of FIG. In this case, as shown in FIG. 2A, the voice search device 100 accepts a change in speech speed in a state where search results that can be selected by the user are not presented in the search screen Wn. The user designates the speech speed by comparing the recorded voice at an arbitrary reproduction position designated by the recorded voice slider VS with the demo voice.

  According to this aspect, an appropriate speech speed can be set before starting the voice search. For this reason, it is possible to prevent an inaccurate voice search from being performed, and to perform an accurate voice search efficiently.

  In this aspect, the user may listen to the character string represented by the recorded voice, input the character string via the input unit 40, and cause the voice output unit 102 to output the demo voice representing the character string. According to this aspect, it is possible to hear and compare the recorded voice and the demo voice that represents the same character string as the recorded voice. For this reason, the user can specify the speech speed more intuitively and more accurately than when comparing the recorded voice and the demo voice representing a character string different from the recorded voice.

  Further, the voice search device 100 may accept the change in the speech speed immediately after step S115 in the flowchart of FIG. 4 (immediately after the result of the first pass search is presented to the user). In this case, as shown in FIG. 2B, the voice search device 100 accepts a change in speech speed in a state where the result of the first pass search is presented to be selectable on the search screen Wn. The user listens to the recorded voice corresponding to the result of the first pass search and the demo voice representing the query. That is, according to this aspect, the user listens and compares the demo voice that represents the query and the recorded voice in the section that is likely to include the voice that represents the query. ), You can specify the speaking speed more intuitively and more accurately.

  In addition, according to this aspect, compared to the above-described embodiment in which the recorded voice corresponding to the result of the second pass search and the demo voice representing the query are compared, an appropriate speech speed can be obtained at an early stage of the voice search. Can be specified. For this reason, it is possible to prevent the second-pass search from being performed on the candidate of the inappropriate estimated section selected in the first-pass search based on the inappropriate speech speed, and to perform an accurate and accurate voice search. In addition, according to this aspect, the search time for the first pass is shorter than the search for the second pass, so that the time required for the cycle of changing the speech speed, comparing the speech speed, and re-searching as compared with the above-described embodiment. Is short and the speech speed can be adjusted quickly. On the other hand, in the above-described embodiment in which the comparison is performed using the recorded voice corresponding to the result of the second pass search that is more accurate than the first pass search, the recording that is more likely to represent the query than this aspect. Since comparison can be made using voice, an appropriate speech speed can be set more efficiently.

  In addition, the voice search device 100 can accept a change in speech speed in parallel with an arbitrary process in the flowchart of FIG. According to this aspect, since the speech speed can be changed at an arbitrary timing during the voice search, it is possible to execute the voice search more in line with the user's intention.

  In the above-described embodiment, the speech search apparatus 100 calculates the output probability based on the monophone phoneme string as part of the speech search process. However, this is only an example, and a voice search can be executed without calculating an output probability based on a monophone phoneme.

  Hereinafter, the function and operation of the speech search apparatus 100 ′ that performs speech search without calculating the output probability based on monophone phonemes will be described.

  The voice search device 100 ′ has a configuration that is generally common to the voice search device 100 according to the above-described embodiment, but a part of the configuration is different. Specifically, as shown in FIG. 7, the voice search device 100 ′ does not include the monophone model DB 31 unlike the voice search device 100. Instead, the voice search device 100 ′ includes an output probability DB 36 in the external storage unit 30, unlike the voice search device 100.

  The output probability DB 36 includes, for each frame set on the time axis of the recorded sound, each phoneme of the monophone model, and an output probability that the feature amount of the recorded sound is output from each phoneme calculated in advance by an external device. Are stored in association with each other. In other words, the voice search device 100 ′ previously acquires the output probability of each frame calculated by the external device based on the monophone model from the external device and stores it in the output probability DB 36. The output probability based on the monophone model may be calculated prior to the voice search process by the voice search device 100 ′ itself, not the external device, and stored in the output probability DB 36.

  The voice search device 100 ′ having the physical and functional configuration shown in FIG. 7 executes processing that is substantially the same as the voice search processing shown in the flowchart of FIG. 4. However, the processing in step S109 is different.

  Specifically, in step S109, the first search unit 104a of the voice search device 100 ′ acquires the output probability stored in the output probability DB 36 in association with each phoneme included in the monophone phoneme sequence acquired in step S103. . That is, the first search unit 104a calculates the output probability of the phonemes included in the monophone phoneme sequence from the output probabilities of all phonemes of the monophone model stored for each frame in the output probability DB 36 for all frames of the recorded speech. get.

  As described above, the voice search device 100 ′ stores the output probability of the entire recorded voice as a likelihood index in advance, and performs a first pass search based on the likelihood index during the voice search. That is, the voice search device 100 ′ can perform a voice search without calculating the output probability. The voice search device 100 ′ can suppress the calculation load and can perform voice search at a higher speed than the voice search device 100 that calculates the output probability based on the monophone model.

  The voice search device according to the present invention can be realized by any electronic device such as a smartphone, a computer, a digital camera, or a PDA (Personal Digital Assistance).

  Specifically, a program for operating an electronic device such as a smartphone, a computer, a digital camera, or a PDA as a voice search device according to the present invention is recorded on a recording medium (for example, a memory card or a CD) that can be read by these electronic devices. The voice search device according to the present invention can be realized by storing, distributing, and installing in a ROM (Compact Disc Read-Only Memory), DVD-ROM (Digital Versatile Disc Read-Only Memory), and the like.

  Alternatively, the above program is stored in a storage device (for example, a disk device) of a server device on a communication network such as the Internet, and a smartphone, a computer, a digital camera, a PDA, or the like downloads the program. The voice search device according to the invention may be realized.

  Further, when the function of the voice search device according to the present invention is realized by the cooperation or sharing of an operating system (OS) and an application program, only the application program portion is stored in a recording medium or a storage device. May be.

  Further, the application program may be superimposed on a carrier wave and distributed via a communication network. For example, an application program may be posted on a bulletin board (BBS: Bulletin Board System) on a communication network, and the application program may be distributed via the network. Then, the voice search device according to the present invention may be realized by installing and starting this application program on a computer and executing it in the same manner as other application programs under the control of the OS.

  The preferred embodiments of the present invention have been described above. However, the present invention is not limited to the specific embodiments, and the present invention includes the invention described in the claims and the equivalent scope thereof. included. Hereinafter, the invention described in the scope of claims of the present application will be appended.

(Appendix 1)
A query acquisition unit for acquiring a query;
A speech rate designation unit for accepting a speech rate designated by the user from a predetermined range of speech rate;
A demo voice output unit that outputs a demo voice representing the acquired query at a speech speed accepted by the speech speed designation unit;
Based on the speech speed received by the speech speed designating unit, a search unit that performs a voice search from the recorded voice that is the target of the voice search for the voice representing the query;
A voice search device comprising:

(Appendix 2)
A recording voice output unit that outputs a voice of a recording part specified by a user or a recording part obtained by voice search by the search unit of the recorded voice,
The voice search device according to appendix 1, wherein the demo voice output by the demo voice output unit and the recorded voice output by the recorded voice output unit can be compared for a user.

(Appendix 3)
The search unit
Based on the speech speed designated by the user via the speech speed designating unit and the acoustic model that does not depend on adjacent phonemes, the section in the recorded speech estimated to include speech representing the query A first search unit for selecting a candidate for an estimated interval from the recorded voice;
The voice search device according to appendix 2, wherein the recorded voice output unit outputs the recorded voice in the estimated section candidate selected by the first search unit.

(Appendix 4)
The search unit
Based on an acoustic model that depends on adjacent phonemes, a second search unit that identifies the estimated section from among the candidates for the estimated section selected by the first search unit,
4. The voice search apparatus according to appendix 3, wherein the recorded voice output unit outputs the recorded voice in the estimated section specified by the second search unit.

(Appendix 5)
A query acquisition step for acquiring a query;
A speech speed designation step for accepting a speech speed designated by the user from a predetermined range of speech speed;
A demo voice output step for outputting a demo voice representing the acquired query at a voice speed accepted in the voice speed designation step;
Based on the speech speed accepted in the speech speed designation step, a search step for performing a voice search for the voice representing the query from the recorded voice that is the target of the voice search;
A voice search method comprising:

(Appendix 6)
Computer
A query acquisition unit for acquiring a query,
A speech speed designation unit that accepts a speech speed designated by the user from a predetermined range of speech speed;
A demo voice output unit that outputs a demo voice representing the acquired query at a speech speed accepted by the speech speed designation unit,
A search unit that searches the voice representing the query from the recorded voice that is the target of the voice search, based on the speech speed received by the speech speed specifying unit,
A program characterized by functioning as

DESCRIPTION OF SYMBOLS 10 ... ROM, 20 ... RAM, 30 ... External storage part, 31 ... Monophone model DB, 32 ... Triphone model DB, 33 ... Time length DB, 34 ... Prosody DB, 35 ... Speech DB, 36 ... Output probability DB, DESCRIPTION OF SYMBOLS 40 ... Input part, 50 ... Output part, 60 ... CPU, 100, 100 '... Voice search device, 101 ... Speech speed designation part, 102 ... Voice output part, 102a ... Demo voice output part, 102b ... Recording voice output part, DESCRIPTION OF SYMBOLS 103 ... Query acquisition part, 104 ... Search part, 104a ... 1st search part, 104b ... 2nd search part, 105 ... Presentation part, Wn ... Search screen, SB ... Speech speed seek bar, SS ... Speech speed slider, RS ... Search Icon, P1 ... Recording sound playback icon, P2 ... Demo sound playback icon, VS ... Recording sound slider, VB ... Recording sound seek bar, S ... Shift length, WV ... Recording sound waveform, L ... Time length of time acquisition sections, T ... the time length of the recorded sound of the waveform, F ... frame length, CS ... speech speed change icon, ES ... end icon

Claims (6)

A query acquisition unit for acquiring a query;
A speech rate designation unit for accepting a speech rate designated by the user from a predetermined range of speech rate;
A demo voice output unit that outputs a demo voice representing the acquired query at a speech speed accepted by the speech speed designation unit;
Based on the speech speed received by the speech speed designating unit, a search unit that performs a voice search from the recorded voice that is the target of the voice search for the voice representing the query;
A voice search device comprising: A recording voice output unit that outputs a voice of a recording part specified by a user or a recording part obtained by voice search by the search unit of the recorded voice,
The voice search device according to claim 1, wherein the demo voice output by the demo voice output unit and the recorded voice output by the recorded voice output unit can be compared for a user. The search unit
Based on the speech speed designated by the user via the speech speed designating unit and the acoustic model that does not depend on adjacent phonemes, the interval in the recorded speech that is estimated to include speech representing the query A first search unit for selecting a candidate for an estimated interval from the recorded voice;
The voice search apparatus according to claim 2, wherein the recorded voice output unit outputs the recorded voice in the estimated section candidate selected by the first search unit. The search unit
Based on an acoustic model that depends on adjacent phonemes, a second search unit that identifies the estimated section from among the candidates for the estimated section selected by the first search unit,
The voice search device according to claim 3, wherein the recorded voice output unit outputs the recorded voice in the estimated section specified by the second search unit. A query acquisition step for acquiring a query;
A speech speed designation step for accepting a speech speed designated by the user from a predetermined range of speech speed;
A demo voice output step for outputting a demo voice representing the acquired query at a voice speed accepted in the voice speed designation step;
Based on the speech speed accepted in the speech speed designation step, a search step for performing a voice search for the voice representing the query from the recorded voice that is the target of the voice search;
A voice search method comprising: Computer
A query acquisition unit for acquiring a query,
A speech speed designation unit that accepts a speech speed designated by the user from a predetermined range of speech speed;
A demo voice output unit that outputs a demo voice representing the acquired query at a speech speed accepted by the speech speed designation unit,
A search unit that searches the voice representing the query from the recorded voice that is the target of the voice search, based on the speech speed received by the speech speed specifying unit,
A program characterized by functioning as