JP6565416B2 - 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.

  In the voice search, a search technique is used that specifies, from a voice signal, a portion where a voice corresponding to a search word (query) to be searched is spoken. In this voice search technology, it is important to realize fast and accurate voice search.

  As one of the voice search techniques, Non-Patent Document 1 discloses a technique for comparing a search target voice signal with a query voice signal to be searched at high speed. In the technology disclosed in Non-Patent Document 1, the feature amount of the search target speech signal and the feature amount of the query speech signal are compared at high speed.

Y. Zhang and J.H. Glass. “An inner-product lower-bound estimate for dynamic time warping,” in Proc. ICASSP, 2011, pp. 5660-5663.

  However, in the technique disclosed in Non-Patent Document 1, when the input person of the search target speech and the query input person are the same person, the acoustic features approximate even though the utterance contents are different. There is a case where the position of different utterance contents is erroneously detected, and there is a problem that search accuracy is lowered.

  The present invention is for solving the above-described problems, and a voice search device capable of performing a search with higher accuracy even when a search target voice input person and a query input person are the same person. Another object is to provide a voice search method and program.

In order to achieve the above object, a voice search device according to the present invention provides:
For each frame included in the search target speech signal, a second probability, which is a probability that the feature amount of the search target speech signal matches the feature amount of each state of the phoneme created from the acoustic model, is determined as a phoneme of the acoustic model. Output probability storage means for storing in association with each state of
Acquired for each frame included in the query speech signal, the probability that the feature amount of the query speech signal matches the feature amount of each state of the phoneme created from the acoustic model, and each state of the phoneme of the acoustic model Estimation that the query speech signal is estimated to be generated from among the speech signals to be searched based on the associated first probability and the second probability stored in the output probability storage means An identification means for identifying the section;
It is characterized by providing.

  According to the present invention, even when the query input user's voice is partially included in the search target voice, the search can be performed with higher accuracy.

It is a figure which shows the physical structure of the speech search device which concerns on Embodiment 1 of this invention. It is a figure which shows the function structure of the speech search device which concerns on Embodiment 1 of this invention. It is a figure for demonstrating the state of a phoneme. It is a figure for demonstrating a search index. (A) is a waveform diagram of an audio signal to be searched. (B) is a figure which shows the flame | frame set in the audio | voice signal of search object. (C) is a figure which shows the likelihood acquisition area designated in the audio | voice signal of search object. (A) is a waveform diagram of a query voice signal. (B) is a figure which shows the flame | frame set in a query audio | voice signal. It is a figure for demonstrating the output probability of a query audio | voice signal. It is a figure which shows the example which makes output probability Lower-Bound. It is a flowchart which shows the flow of the voice search process which the voice search device concerning Embodiment 1 of this invention performs.

  Hereinafter, a voice search device, a voice search method, and a program according to an embodiment of the present invention will be described with reference to the drawings. In the drawings, the same or corresponding parts are denoted by the same reference numerals.

(Embodiment 1)
As shown in FIG. 1, the speech search apparatus 100 according to the first embodiment physically includes a ROM (Read Only Memory) 1, a RAM (Random Access Memory) 2, an external storage device 3, and an input device 4. And an output device 5, a CPU (Central Processing Unit) 6, and a bus 7.

  The ROM 1 stores a voice search program. The RAM 2 is used as a work area for the CPU 6.

  The external storage device 3 is composed of, for example, a hard disk, and stores an audio signal to be analyzed, a search index described later, and an acoustic model as data.

  The input device 4 is composed of a voice input device such as a microphone. The input device 4 supplies a search term (query voice signal) input by the user to the CPU 6 as voice data. The output device 5 includes, for example, a speaker and a liquid crystal display screen. The output device 5 outputs the audio data output by the CPU 6 from the speaker, and displays the position of the searched search word in the audio signal on the screen.

  The bus 7 connects the ROM 1, RAM 2, external storage device 3, input device 4, output device 5, and CPU 6. The CPU 6 reads the voice search program stored in the ROM 1 into the RAM 2 and executes the voice search program, thereby realizing the following functions.

  As shown in FIG. 2, the speech search apparatus 100 functionally includes an acoustic model storage unit 101, an output probability storage unit 102, a query output probability storage unit 103, a query speech signal acquisition unit 111, a frame sequence, Creation unit 112, query feature amount acquisition unit 113, query output probability acquisition unit 114, section specification unit 115, second output probability acquisition unit 116, replacement unit 117, likelihood acquisition unit 118, and repetition unit 119 and the specifying unit 120. The acoustic model storage unit 101, the output probability storage unit 102, and the query output probability storage unit 103 are constructed in the storage area of the external storage device 3.

  The acoustic model storage unit 101 stores a monophone model acoustic model. The monophone model is an acoustic model generated for each phoneme, and is an acoustic model that does not depend on adjacent phonemes. The voice search device 100 learns the monophone model by a general method and stores it in the acoustic model storage unit 101 in advance.

  As the monophone model, for example, an HMM (Hidden Markov Model) that is an acoustic model used in general speech recognition can be used. The HMM is a model for probabilistically estimating phonemes constituting a speech signal from the speech signal by a statistical method. In the HMM, a standard pattern using as parameters the transition probability indicating temporal state fluctuations and the probability (output probability) matching the feature quantity input from each state is used.

  A phoneme is a unit of components constituting speech uttered by a speaker. For example, the word “xoxadedo” is composed of 11 phonemes “k, i, z, o, k, u, s, e, i, d, o”. A phoneme is further divided into a plurality of states.

  A state is a minimum time unit constituting a phoneme. A case where the number of states defined for each phoneme is “3” will be described as an example. For example, as shown in FIG. 3, the phoneme “a” of the voice “A” includes a first state “a1” including the start of utterance of the phoneme, a second state “a2” that is an intermediate state, The state is divided into three states including a third state “a3” including the end of utterance. That is, one phoneme is composed of three states. When all phonemes are composed of three states, there are (m × 3) states where m is the number of all phonemes used in the acoustic model.

  Returning to FIG. 2, the output probability storage unit 102 stores the search index shown in FIG. The search index is an output probability (second probability) in which a plurality of frames are set in the speech signal to be searched, and the feature amount of the speech for each frame matches the feature amount of each state of the phoneme of the acoustic model. This is a table storing probabilities.

  The feature amount of each phoneme state is a numerical value representing the feature of speech extracted from the speech signal for each phoneme state. This feature amount is a combination of a frequency axis system characteristic parameter obtained by converting audio data onto the frequency axis and a power system characteristic parameter obtained by calculating the square sum of the energy of the audio data and its logarithm. Can be obtained.

  For example, as is well known, the feature amount is a frequency axis feature obtained by taking a difference between a frequency axis feature parameter 12 component (12 dimensions), a power feature parameter 1 component (1 dimension), and each component of the immediately preceding time window. The sum of the 12 parameter components (12 dimensions), the power system feature parameter 1 component (1 dimension), and the frequency axis system feature parameter 12 components (12 dimensions) obtained by subtracting the difference between the previous time window components. It is configured as a 38-dimensional vector quantity having 38 components.

  The audio signal to be searched is, for example, an audio signal related to audio such as news broadcast, recorded conference audio, recorded lecture audio, movie audio, and the like.

  The frame is a time window for comparing the search target audio signal and the query audio signal. In the present embodiment, the speech detection is performed by comparing the search target speech signal with the query speech signal for each phoneme state. For example, 40 ms is used as the frame length.

A method for setting a section for each frame in the audio signal to be searched will be described with reference to FIG. FIG. 5A is a waveform diagram of an audio signal to be searched for a time length T from the beginning to the end. The vertical axis indicates the intensity of the audio signal, and the horizontal axis indicates time. FIG. 5B shows a frame set in the audio signal shown in FIG. As shown in FIG. 5B, the section of the frame length t is shifted by one shift length S, and the sections of frame numbers f ₁ to f _N are set in the audio signal to be searched. Section of the frame number f ₁ is a section of the length of time t starting at the beginning of the audio signal. Section of the frame number f ₂ is the section of the length of time t starting from 1 shift length S shifted by a position from the head of the audio signal. Hereinafter similarly, is set to the frame number f _N are shifted by the shift length S.

  The shift length S is a length that determines the accuracy of the search. The shift length S is a fixed value set to a value shorter than the frame length t. For example, when the frame length is set to t = 40 ms, the shift length is set to S = 10 ms.

In the search index, an output probability (second probability), which is a probability that the section of frame numbers f ₁ to f _N set in the speech signal to be searched matches each state of the phoneme of the acoustic model, is set as each state of the phoneme. It is a table stored in association with each other. In order to create this search index, first, the feature amount of the audio signal to be searched is acquired for each frame of frame numbers f ₁ to f _N.

Then, by comparing the acquired feature quantity with the feature quantity of each phoneme state of the acoustic model, the feature quantity of the speech signal included in the frames of frame numbers f ₁ to f _N is obtained by An output probability (second probability) that is a probability that matches the feature amount is acquired for each frame, and is stored in the output probability storage unit 102 as a search index associated with each state of the phoneme. FIG. 4 shows an example of a search index in which the number of phonemes is m and the number of phoneme states is three. The first column in FIG. 4 shows the frame numbers of the frames created by shifting by the shift length S. The probability that the feature quantity for each frame matches the feature quantity in each state of the phoneme is represented by f (x, y, z). x (x = 1 to N) represents a frame number, y (y = 1 to m) represents a phoneme number, and z (z = 1 to 3) represents a state number. f (1, 1, 1) represents the probability that the feature quantity of the speech signal included in the frame of frame number f ₁ matches the feature quantity of state 1 of phoneme 1 included in the acoustic model. Feature quantity of the audio signal included in the frame of the frame number f _X is, it represents the probability of matching the feature quantity of state z phoneme number y that is included in the acoustic model f (x, y, z).

  Returning to FIG. 2, the user inputs a search word (query) for searching for a portion where the target voice is uttered from the search target voice signal to the voice search device 100, and the query voice. The signal acquisition unit 111 acquires a query voice signal input by the user via the input device 4 as voice data.

The frame sequence creation unit 112 creates a frame sequence obtained by dividing the query speech signal acquired by the query speech signal acquisition unit 111 into sections for each frame length. The frame sequence of the query audio signal will be described with reference to FIG. FIG. 6A is a waveform diagram of a query voice signal having a time length L from the beginning to the end. The time length L is the time length (speech time length) during which the query voice signal is uttered. The vertical axis indicates the intensity of the query voice signal, and the horizontal axis indicates time. FIG. 6B shows a frame set in the query voice signal shown in FIG. As shown in FIG. 6B, the frame sequence creation unit 112 shifts the section of the frame length t by one shift length S, and sets the sections of the frame numbers g ₁ to g _{k in} the query speech signal. The frame length t is the same as the frame length t (for example, 40 ms) used when creating the search index. The shift length S is also set to the same shift length S as when the search index was created. Section of the frame number g ₁ is a section of the length of time t starting at the beginning of the query speech signal. Section of frame number g ₂ is a section of the length of time t starting from 1 shift length S shifted position from the beginning of the query speech signal. Hereinafter Similarly, shifted by shift length S is set to the frame number g _k.

Returning to FIG. 2, the query feature quantity acquisition unit 113 acquires the feature quantity of the query speech signal for each frame (g _{1 to} g _k ) constituting the frame sequence created by the frame sequence creation unit 112. The feature quantity acquisition method is the same as that used when creating a search index.

Based on the feature quantity acquired by the query feature quantity acquisition unit 113, the query output probability acquisition unit 114 calculates a probability (first probability) that the feature quantity matches the feature quantity of each state of the phoneme included in the acoustic model. It is acquired for each frame (g _{1 to} g _k ) and stored in the query output probability storage unit 103 in association with each phoneme state. FIG. 7 shows a case where the number of phonemes is m and the number of phoneme states is three. The number of phoneme types “m” and the number of states “3” are the same as those at the time of creating the search index. The first column in FIG. 7 shows the frame numbers of the frames constituting the frame sequence created by the frame sequence creation unit 112. Then, the feature quantity of the frame _(g 1 ~g _k) constituting the frame column represents the probability that matches the characteristic amount of each state of the phoneme in g (a, y, z) . a (a = 1 to k) indicates a frame number of the query voice signal, y (y = 1 to m) indicates a phoneme number, and z (z = 1 to 3) indicates a state number.

  The frame number k of the query voice signal is a natural number obtained by rounding down the decimal point of the value obtained by k = L / s using the utterance time length L and the shift length s of the query voice signal.

Returning to FIG. 2, the section specifying unit 115 specifies a plurality of sections of the utterance time length L of the query voice signal as the likelihood acquisition section from the voice signal. The likelihood acquisition section is a section for acquiring the likelihood that the query voice signal is generated from the section. The likelihood is an index indicating the degree of similarity between the search target voice and the query voice signal. This will be described with reference to FIG. The section specifying unit 115 first specifies a section of the utterance time length L of the query voice signal starting from the _first frame f1 of the search target voice signal as the first likelihood acquisition section. In the present embodiment, since the number of frames frames constituting the query speech signal is set to the k, it specifies a section of the k-th frame f _k from the first frame f ₁ as the first likelihood acquisition sections.

Then, the section specifying unit 115 specifies the first (k + 1) frame _{f k + 1} of the segment as the second likelihood acquisition sections from the second frame _{f 2} of the audio signal. Similarly, the process is designated up to the P-th likelihood acquisition section. Note that the number P of likelihood acquisition sections that can be specified in the speech signal to be searched is the time length T of the speech signal, the time length of the likelihood acquisition section (utterance time length of the query speech signal) L, and the shift length S. , P = (TL−S) / S is a natural number obtained by rounding down the value after the decimal point.

Returning to FIG. 2, the second output probability acquisition unit 116 acquires a probability (third probability) that each frame constituting the query speech signal matches each frame constituting the speech signal to be searched. Specifically, the probability (first probability) that the feature amount of each frame of the query speech signal matches the feature amount of each state of the phoneme, and the probability (second second) stored in the search index of the search target speech signal. (Probability), the feature amount of each frame (g _{1 to} g _k ) of the query speech signal matches the feature amount of each frame (f _{1 to} f _N ) of the speech signal to be searched ( _{first order} ). 3).

This will be specifically described with reference to FIGS. 4 and 7. When the section specifying unit 115 specifies the first likelihood acquisition section starting from the first frame f ₁ of the audio signal, the second output probability acquiring unit 116 includes the _first frame g ₁ of the query audio signal and the _first frame f _{1 of the} audio signal. for by multiplying the output probability of each state of the phoneme to obtain a probability that the first frame g ₁ of the query speech signal matches the first frame f ₁ of the search target speech signal.

Specifically, the second output probability acquisition unit 116 has a probability P (1, 1, 1) that the first frame g ₁ of the query speech signal and the first frame f ₁ of the speech signal are both in the state 1 of the phoneme 1. ) Is obtained from equation (1). The first frame _{f 1} of the first frame _{g 1} and the audio signal of the query audio signal, together determine the probability P (1,2,1) is a state 1 phoneme 2 from equation (2). In the same manner, the second output probability obtaining unit 116, a first frame f ₁ of the first frame g ₁ and the audio signal of the query speech signal, the probability P (1, both the state 3 phoneme m, m, 3 ) Is obtained from equation (3).
P (1,1,1) = f (1,1,1) × g (1,1,1) (1)
P (1,2,1) = f (1,2,1) × g (1,2,1) (2)
P (1, m, 3) = f (1, m, 3) × g (1, m, 3) Expression (3)

In this manner, the second output probability acquisition unit 116 acquires (m × 3) probabilities (third probabilities) for the first frame g ₁ of the query speech signal. Then, by multiplying the (m × 3) probabilities, the output probability P (1, 1), which is the probability that the first frame g ₁ of the query voice signal matches the first frame f ₁ of the search target voice signal. ) Is obtained by equation (4).

Then, the second output probability obtaining unit 116, by multiplying the output probability of each state of the phoneme corresponding to the second frame f ₂ of the second frame g ₂ and the audio signal of the query speech signals, query the voice signal the second frame g ₂ obtains a probability that match the second frame f ₂ of the search target speech signal. Specifically, the second output probability acquisition unit 116 acquires (m × 3) output probabilities for the second frame g ₂ of the query speech signal. Then, by multiplying the (m × 3) output probabilities, the output probability P (1, 1, which is the probability that the second frame g ₂ of the query speech signal matches the second frame f ₂ of the search target speech signal. 2) is obtained by equation (5).

In the same manner, the second output probability obtaining unit 116 outputs the probability P up to the k frame g _k queries audio signal (1, k) is obtained by equation (6).

The acquisition of the output probability is over the case where the query speech signal starting from the first frame f ₁ of the search target of the speech signal, the section specifying unit 115 specifies the second likelihood acquisition sections starting from the second frame f ₂ of the audio signal To do. The second output probability obtaining unit 116 performs the same calculation and the combined first frame g ₁ of the query speech signal to the second frame f ₂ of the search target speech signal.

Similarly, the second output probability acquisition unit 116 obtains the output probability up to the P-th likelihood acquisition section. The second output probability obtaining unit 116, the j-th frame g of the query speech signal when the combined first frame g ₁ of the query speech signal to the s frame f _s of the search target voice signal (the s likelihood acquisition sections) The output probability of _j is obtained by equation (8).

  Returning to FIG. 2, the replacement unit 117 replaces each of the output probabilities acquired by the second output probability acquisition unit 116 with the maximum output probability in several frames before and after the frame. This replacement process is called Lower-Bound conversion.

  Specifically, referring to FIG. 8, lower-bound conversion will be described. In FIG. 8, a solid line indicates the output probability acquired for each frame. The vertical axis shows the higher output probability as it goes down, and the horizontal axis shows time. The replacement unit 117 replaces the output probability of each frame with the maximum output probability among the frame, N1 frames before the frame, and N2 frames after the frame. N1 and N2 are natural numbers including 0, but either N1 or N2 is not 0.

If the combined first frame _{g 1} of the query speech signal to the first frame _{f 1} of the audio signal is described as N1 = N2 = 2. Since there is no frame before the output probability P (1, 1) of the first frame g ₁ of the query voice signal, the replacement unit 117 has its own first frame g ₁ P (1, 1) and the subsequent the second frame _{g 2} of P and (1,2) replacing the largest output probability in the third frame _{g 3} of P (1, 3). Replacement unit 117, a second frame _{g 2} of the output probability P of the query speech signal (1, 2), the previous first frame _{g 1} of the (P1,1) and its second frame _{g 2} of P ( 1, 2) and replaces the subsequent third frame _{g 3} of P and (1,3) and the largest output probability in P (l, 4) of the fourth frame _{g 4.} Replacement unit 117, the third frame _{g 3} of the output probability P a (1,3), the first frame _{g 1} of the previous P (1, 1) and the second frame _{g 2} P queries audio signal (1 , 2), a P of the third frame _{g 3} itself (1,3), in the subsequent fourth frame _{g 4} of P and (1,4) of the fifth frame _{g 5} P (1, 5) To replace the maximum output probability. Thus, the replacement unit 117 performs replacement processing up to the kth frame. As a result of the replacement, the output probability indicated by the solid line in FIG. 8 is converted into an output probability having a smaller change in value in the time direction, like the output probability after Lower-Bound shown by the broken line.

  Returning to FIG. 2, the likelihood acquisition unit 118 indicates that the likelihood acquisition section specified by the section specification unit 115 based on the output probability after replacement by the replacement unit 117 is a section in which the query voice signal is being emitted. The likelihood indicating the likelihood of the is acquired. Specifically, the likelihood acquisition unit 118 adds the value obtained by taking the logarithm of the output probability after replacement over all frames from the beginning to the end of the likelihood acquisition interval, in this example, over k frames. Then, the likelihood of this likelihood acquisition section is acquired. That is, the likelihood acquired by the likelihood acquisition unit 118 becomes higher as the likelihood acquisition section includes more frames with higher output probabilities.

The repeater 119 changes the designated section in the speech signal of the likelihood acquisition section specified by the section specifying section 115 to change the section specifying section 115, the second output probability acquiring section 116, the replacing section 117, and the likelihood acquiring section 118. Control each unit to repeat the process. In the first process, so it called for the likelihood of the first likelihood acquisition sections starting from the first frame f ₁ to be searched for audio signals, a second time, starting from the second frame f ₂ of the search target speech signal The likelihood of the second likelihood acquisition section is obtained. Thereafter, the frame is shifted frame by frame, and the likelihood up to the P-th likelihood acquisition section is obtained.

  Based on the P likelihoods acquired by the likelihood acquisition unit 118, the specification unit 120 specifies an estimation interval in which it is estimated that the query speech signal is generated from the search target speech signals. For example, the specification unit 120 sorts the likelihood acquisition sections in descending order of the likelihood acquired by the likelihood acquisition unit 118, and specifies a predetermined number of sections as the estimation sections in descending order of likelihood. Alternatively, a section whose likelihood is equal to or greater than a predetermined threshold is specified as an estimated section. The position information of the section specified by the specifying unit 120 is displayed outside as a search result via a screen provided in the output device 5.

  The speech search processing executed by the speech search apparatus 100 having the above-described physical configuration and functional configuration will be described with reference to the flowchart shown in FIG.

  The user stores the acoustic model in the acoustic model storage unit 101 in advance. Further, a search index indicating the second probability shown in FIG. 4 is created from the audio signal to be searched, and stored in the output probability storage unit 102. Further, the user inputs a search word (query) from the input device 4 as an audio signal. Here, it is assumed that the query to be searched is “xoxadedo”.

  The CPU 6 reads out the voice search program from the ROM 1 and executes the voice search program, whereby the flowchart shown in FIG. 9 starts. When the user inputs a query voice signal to be searched from the input device 4, the query voice signal acquisition unit 111 acquires the query voice signal as voice data. Then, the frame sequence creation unit 112 creates a frame sequence obtained by dividing the acquired query audio signal for each frame (step S10). The query feature quantity acquisition unit 113 acquires a feature quantity for each frame constituting the frame sequence of the query audio signal (step S11).

  Next, the query output probability acquisition unit 114 acquires the probability (first probability) that the feature amount acquired for each frame of the query speech signal matches the feature amount in each state of the phoneme of the acoustic model (step S12). 7, the acquired output probability is stored in the query output probability storage unit 103 in association with each phoneme state.

  When acquisition of the output probability (first probability) of the query voice signal is completed, the section specifying unit 115 acquires a probability (third probability) that the query voice signal matches the search target voice signal (third probability). Are set, and the likelihood acquisition unit 118 acquires the likelihood that the query voice signal is generated from each likelihood acquisition section.

Therefore, the section specifying unit 115, first, designating the first likelihood acquisition sections starting from the first frame f ₁ search index (Step S13). Then, the second output probability acquisition unit 116 obtains the probability (third probability) that the first frame g ₁ of the query speech signal matches the first frame f ₁ of the speech signal to be searched using Equation (4). Similarly, the second output probability obtaining unit 116, the output probability up to the k frame _{g k} queries audio signal (third probability) is obtained by equation (6) (step S14).

  When the second output probability acquisition unit 116 acquires the output probability, the replacement unit 117 calculates the output probability acquired for each frame of the frame, N1 frames before the frame, and N2 frames after the frame. By replacing it with the maximum output probability in the total (1 + N1 + N2) frames, Lower-Bound processing is executed (step S15).

  The likelihood obtaining unit 118 obtains the likelihood of the likelihood obtaining section designated by the section designating unit 115 by adding the output probabilities after Lower-Bound to each frame by taking a logarithm (step S16). When the likelihood acquisition unit 118 acquires the likelihood, the repetition unit 119 determines whether or not the likelihood acquisition of all the sections in the search target audio signal has been completed (step S17).

  When the likelihood acquisition of all the sections has not been completed (step S17; No), the repetition unit 119 designates the next likelihood acquisition section in which the position of the search index is advanced by one frame (step S18). And the process of step S14-S16 mentioned above with respect to the likelihood acquisition area newly designated by the area designation | designated part 115 is repeated.

  When the section specifying unit 115 specifies the s-th likelihood acquisition section, the second output probability acquisition unit 116 obtains an output probability for each of the k frames included in the s-likelihood acquisition section by Expression (8). (Step S14). Then, a lower-bound process is performed on the obtained output probability for each frame (step S15). The likelihood obtaining unit 118 obtains the likelihood of the likelihood obtaining section designated by the section designating unit 115 by adding the output probabilities after Lower-Bound to each frame by taking a logarithm (step S16).

  As described above, the repetition unit 119 controls the section specifying unit 115, the second output probability acquisition unit 116, the replacement unit 117, and the likelihood acquisition unit 118 so as to sequentially acquire the likelihoods up to the Pth likelihood acquisition section. To do.

  Finally, when the likelihood acquisition of all the sections is completed (step S17; YES), the speech search apparatus 100 proceeds to a process of specifying a section corresponding to the query speech signal based on the acquired likelihood (step S19). To do. For example, the specifying unit 120 sorts P likelihoods acquired for each likelihood acquisition section in descending order of likelihood, and specifies a section having the likelihood equal to or higher than a threshold as a search result section. When the specifying unit 120 specifies a section corresponding to the query voice signal, the specifying unit 120 outputs a specifying result via the output device 5 (step S20). Thus, the voice search process executed by the voice search device 100 ends.

  As described above, the speech search apparatus 100 according to the first embodiment compares and compares the feature quantities of the search target speech signal and the query speech signal with the probability of matching the feature quantity of the acoustic model. Thereby, the influence by the acoustic characteristic peculiar to each person can be reduced. Therefore, even when the query input user's voice is partly included in the search target voice, it is possible to reduce the erroneous search due to the acoustic features unique to each person and increase the search accuracy.

  In the above description of the embodiment, the case where a voice search is performed using a monophone model as an acoustic model has been described. However, the application of the present invention is not limited to this. For example, the present invention can be applied to a case where a voice search is performed using a biphone model or a triphone model.

  Moreover, although the above embodiment demonstrated the case where this invention was applied to the primary search using a monophone model, the secondary search using a triphone model is performed after step S20 of Embodiment 1, Further, the search accuracy may be improved.

  In the description of the first embodiment, the section specifying unit 115 shifts the start position of the likelihood acquisition section by one shift length, and obtains the third probability each time the shift is performed, and performs the Lower-Bound process (replacement process). The process of performing and obtaining likelihood was described P times. However, after the section specifying unit 115 first specifies the P-th likelihood acquisition section from the first likelihood acquisition section, the third probability acquisition, the lower-bound process, and the likelihood acquisition process are performed. May be.

  In this embodiment, the process of taking the logarithm of the output probability of each frame and adding it has been described. However, since it is a process of multiplying the output probability of each frame, the output probability is directly multiplied without taking the logarithm. Or an approximate expression may be used instead of the logarithm.

  In addition to being able to provide a voice search device having a configuration for realizing the functions according to the present invention in advance, by applying a program, an existing personal computer, an information terminal device, or the like can be used as the voice search device according to the present invention. It can also function. That is, by applying the program for realizing each functional configuration by the voice search device 100 exemplified in the above embodiment so that a CPU or the like for controlling an existing personal computer, an information terminal device, or the like can be executed. It can be made to function as the voice search device 100 according to the above. The voice search method according to the present invention can be implemented using a voice search device.

  Moreover, the application method of such a program is arbitrary. For example, the program can be stored and applied to a computer-readable recording medium (CD-ROM (Compact Disc Read-Only Memory), DVD (Digital Versatile Disc), MO (Magneto Optical disc), etc.), the Internet, etc. It is also possible to apply the program by storing it in a storage on the network and downloading it.

  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)
For each frame included in the search target speech signal, a second probability, which is a probability that the feature amount of the search target speech signal matches the feature amount of each state of the phoneme created from the acoustic model, is determined as a phoneme of the acoustic model. Output probability storage means for storing in association with each state of
Acquired for each frame included in the query speech signal, the probability that the feature amount of the query speech signal matches the feature amount of each state of the phoneme created from the acoustic model, and each state of the phoneme of the acoustic model Estimation that the query speech signal is estimated to be generated from among the speech signals to be searched based on the associated first probability and the second probability stored in the output probability storage means An identification means for identifying the section;
A voice search device comprising:

(Appendix 2)
Query feature value acquisition means for acquiring the feature value of the query sound signal for each frame that is a section for comparing the search target sound signal and the query sound signal;
Query output probability acquisition means for acquiring the first probability for each frame in association with each state of the phoneme of the acoustic model based on the feature quantity of the query speech signal acquired by the query feature value acquisition means;
The speech search device according to appendix 1, further comprising:

(Appendix 3)
Section specifying means for specifying a plurality of likelihood acquisition sections that are sections having a speech duration of a query voice signal in the search target voice signal;
Based on the first probability and the second probability, the likelihood indicating that the likelihood acquisition section specified by the section specifying means is a section in which the query voice signal is generated is acquired. Likelihood acquisition means to perform,
Further comprising
The section designating unit designates a plurality of likelihood acquisition sections by changing a head position of the likelihood acquisition section in the search target audio signal,
The likelihood acquisition means acquires likelihood for each of the plurality of likelihood acquisition sections,
The specifying means generates the query voice signal from the search target voice signal based on the likelihood acquired by the likelihood acquisition means for each of the likelihood acquisition sections specified by the section specifying means. Identify the estimated interval that is estimated to be,
The speech search device according to Supplementary Note 1 or 2, characterized in that:

(Appendix 4)
For each of the plurality of likelihood acquisition sections, second output probability acquisition that acquires a third probability obtained by multiplying the first probability and the second probability for each frame included in the likelihood acquisition section Further providing means,
The likelihood acquisition means adds the value obtained by taking the logarithm of the third probability acquired for each frame by the second output probability acquisition means to acquire the likelihood of the likelihood acquisition section.
The speech search device according to Supplementary Note 3, wherein

(Appendix 5)
Each of the third probabilities acquired by the second output probability acquisition means for each frame is determined as the maximum output probability among the frame, N1 frames before the frame, and N2 frames after the frame. Further comprising a replacement means for replacing
N1 and N2 are natural numbers including 0, and either N1 or N2 is not 0,
The likelihood acquisition means acquires the likelihood of the likelihood acquisition section specified by the section specification means based on the output probability after replacement by the replacement means.
The voice search device according to supplementary note 4, wherein:

(Appendix 6)
For each frame included in the search target speech signal, a second probability, which is a probability that the feature amount of the search target speech signal matches the feature amount of each state of the phoneme created from the acoustic model, is determined as a phoneme of the acoustic model. Output probability storage step of storing in association with each of the states,
Acquired for each frame included in the query speech signal, the probability that the feature amount of the query speech signal matches the feature amount of each state of the phoneme created from the acoustic model, and each state of the phoneme of the acoustic model Estimation that the query speech signal is estimated to be generated from the search target speech signal based on the associated first probability and the second probability stored in the output probability storage step A specific process for identifying a section;
Voice search method including

(Appendix 7)
Computer
For each frame included in the search target speech signal, a second probability, which is a probability that the feature amount of the search target speech signal matches the feature amount of each state of the phoneme created from the acoustic model, is determined as a phoneme of the acoustic model. Output probability storage means for storing in association with each state of
Acquired for each frame included in the query speech signal, the probability that the feature amount of the query speech signal matches the feature amount of each state of the phoneme created from the acoustic model, and each state of the phoneme of the acoustic model Estimation that the query speech signal is estimated to be generated from among the speech signals to be searched based on the associated first probability and the second probability stored in the output probability storage means Identification means for identifying the section,
Program to function as.

DESCRIPTION OF SYMBOLS 1 ... ROM, 2 ... RAM, 3 ... External storage device, 4 ... Input device, 5 ... Output device, 6 ... CPU, 7 ... Bus, 100 ... Voice search device, 101 ... Acoustic model memory | storage part, 102 ... Output probability memory | storage 103: Query output probability storage unit, 111 ... Query voice signal acquisition unit, 112 ... Frame sequence creation unit, 113 ... Query feature amount acquisition unit, 114 ... Query output probability acquisition unit, 115 ... Section designation unit, 116 ... No. 2-output probability acquisition unit, 117 ... replacement unit, 118 ... likelihood acquisition unit, 119 ... repetition unit, 120 ... identification unit

Claims (7)

For each frame included in the search target speech signal, a second probability, which is a probability that the feature amount of the search target speech signal matches the feature amount of each state of the phoneme created from the acoustic model, is determined as a phoneme of the acoustic model. Output probability storage means for storing in association with each state of
Acquired for each frame included in the query speech signal, the probability that the feature amount of the query speech signal matches the feature amount of each state of the phoneme created from the acoustic model, and each state of the phoneme of the acoustic model Estimation that the query speech signal is estimated to be generated from among the speech signals to be searched based on the associated first probability and the second probability stored in the output probability storage means An identification means for identifying the section;
A voice search device comprising: Query feature value acquisition means for acquiring the feature value of the query sound signal for each frame that is a section for comparing the search target sound signal and the query sound signal;
Query output probability acquisition means for acquiring the first probability for each frame in association with each state of the phoneme of the acoustic model based on the feature quantity of the query speech signal acquired by the query feature value acquisition means;
The voice search device according to claim 1, further comprising: Section specifying means for specifying a plurality of likelihood acquisition sections that are sections having a speech duration of a query voice signal in the search target voice signal;
Based on the first probability and the second probability, the likelihood indicating that the likelihood acquisition section specified by the section specifying means is a section in which the query voice signal is generated is acquired. Likelihood acquisition means to perform,
Further comprising
The section designating unit designates a plurality of likelihood acquisition sections by changing a head position of the likelihood acquisition section in the search target audio signal,
The likelihood acquisition means acquires likelihood for each of the plurality of likelihood acquisition sections,
The specifying means generates the query voice signal from the search target voice signal based on the likelihood acquired by the likelihood acquisition means for each of the likelihood acquisition sections specified by the section specifying means. Identify the estimated interval that is estimated to be,
The speech search apparatus according to claim 1 or 2, wherein For each of the plurality of likelihood acquisition sections, second output probability acquisition that acquires a third probability obtained by multiplying the first probability and the second probability for each frame included in the likelihood acquisition section Further providing means,
The likelihood acquisition means adds the value obtained by taking the logarithm of the third probability acquired for each frame by the second output probability acquisition means to acquire the likelihood of the likelihood acquisition section.
The voice search apparatus according to claim 3. Each of the third probabilities acquired by the second output probability acquisition means for each frame is determined as the maximum output probability among the frame, N1 frames before the frame, and N2 frames after the frame. Further comprising a replacement means for replacing
N1 and N2 are natural numbers including 0, and either N1 or N2 is not 0,
The likelihood acquisition means acquires the likelihood of the likelihood acquisition section specified by the section specification means based on the output probability after replacement by the replacement means.
The voice search device according to claim 4. For each frame included in the search target speech signal, a second probability, which is a probability that the feature amount of the search target speech signal matches the feature amount of each state of the phoneme created from the acoustic model, is determined as a phoneme of the acoustic model. Output probability storage step of storing in association with each of the states,
Acquired for each frame included in the query speech signal, the probability that the feature amount of the query speech signal matches the feature amount of each state of the phoneme created from the acoustic model, and each state of the phoneme of the acoustic model Estimation that the query speech signal is estimated to be generated from the search target speech signal based on the associated first probability and the second probability stored in the output probability storage step A specific process for identifying a section;
Voice search method including Computer
For each frame included in the search target speech signal, a second probability, which is a probability that the feature amount of the search target speech signal matches the feature amount of each state of the phoneme created from the acoustic model, is determined as a phoneme of the acoustic model. Output probability storage means for storing in association with each state of
Acquired for each frame included in the query speech signal, the probability that the feature amount of the query speech signal matches the feature amount of each state of the phoneme created from the acoustic model, and each state of the phoneme of the acoustic model Estimation that the query speech signal is estimated to be generated from among the speech signals to be searched based on the associated first probability and the second probability stored in the output probability storage means Identification means for identifying the section,
Program to function as.

Images