JP2018084748A - Voice section detection device, voice section detection method and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a voice section detection technology that can detect a voice section of a target speaker even under a noise environment.SOLUTION: A voice section detection device comprises: a likelihood calculation unit 110 that calculates a voice likelihood L, and a non-voice likelihood L(j is a frame number) for each frame of voice data caught by a microphone i; a target speaker voice arrival direction estimation unit 120 that estimates a target speaker voice arrival direction θfrom the voice data; a weight calculation unit 130 that calculates weight wof the microphone i from θand a direction θand a distance rrepresenting a position of the microphone i; a target speaker voice arrival direction time change calculation unit 135 that calculates target speaker voice arrival direction time change Δθfrom the θ; a weighted likelihood ratio calculation unit 140 that calculates a weighted likelihood ratio Lfrom the L, the L, the wand the Δθ; and a voice section determination unit 150 that generates a voice section determination result j from the L.SELECTED DRAWING: Figure 2

Description

  The present invention relates to a voice section detection technique, and more particularly to a technique for detecting a voice section spoken by a target speaker using voice signals collected by a plurality of microphones.

  In the voice recognition technology, only a section in which a speaker speaks (hereinafter referred to as a voice section) is cut out using a voice section detection technique, and voice recognition is performed. By extracting only the speech segment from the speech data to be recognized and removing the speech segment where the speaker is not speaking (hereinafter, the noise segment is also referred to as a non-speech segment), the speech recognition is performed with high accuracy. Voice recognition can be performed (Non-patent Document 1). As a method for detecting a speech section, for each frame of input speech data, a speech likelihood that is a likelihood that speech is included and a non-speech likelihood that is a likelihood that speech is not included are calculated. There is a method to determine.

  Voice recognition technology is applied to communication robots that are conversational robots and devices for digital signage. In this communication robot and digital signage device, it is assumed that the speaker's position will be in front of the speaker to some extent, so the sound is picked up by a specific microphone (specifically, a microphone located in front of the robot or device). Speech recognition can be performed after emphasizing the voice data. This makes it possible to improve recognition accuracy.

Masayoshi Fujimoto, “Fundamentals of Speech Interval Detection and Recent Research Trends”, IEICE Technical Report, SP2010-23, pp.7-pp.12, 2010.

  However, noise is present in an actual usage environment such as a communication robot. In an environment with noise, there is a problem that the speech section cannot be detected accurately due to the influence of the noise, resulting in a decrease in speech recognition accuracy.

  One method for solving this problem is to use a noise suppression technique. By using the noise suppression technique, the influence of noise can be reduced to some extent. However, it is difficult to reduce the influence of speech of a speaker different from the target speaker that is the target of speech recognition, which is coherent noise, and sufficient performance cannot be exhibited in crowds.

  Therefore, an object of the present invention is to provide a speech section detection technique that can detect a speech section of a target speaker even in a noisy environment.

In one embodiment of the present invention, the microphone installed in the direction in which the target speaker is supposed to speak (hereinafter referred to as the front direction) is the microphone 0, and the other M (M ≧ 1) microphones are the microphone 1. , Microphones 2,..., Microphone M, M + 1 microphones i (i = 0) with the front direction as viewed from the microphone 0 being a 0 degree direction and the microphone 0 being the center and the 0 degree direction as a reference , 1, 2,..., M) as θ _i (0 ≦ θ _i ≦ 2π, where θ ₀ = 0), the microphone 0 and the M + 1 microphones i (i = 0, 1, 2, ..., M) is the distance r _i (r _i ≥0, r ₀ = 0), and the target speaker's voice collected by the microphone i (i = 0, 1, 2, ..., M) is The included speech data x _i (i = 0, 1, 2,..., M) is divided into frames, and for each frame, speech likelihood L _{speech_i, j} and non-speech likelihood L _{noise_i, j} (i = 0, 1 , 2,…, M, j are the frame numbers A likelihood calculator for calculating the box), the audio data _{x i (i = 0, 1} , 2, ..., purpose M), which for each frame, relative to the direction of 0 degrees about the microphone 0 A target speaker voice arrival direction estimation unit for estimating a target speaker voice arrival direction θ _{est, j} (0 ≦ θ _{est, j} ≦ 2π), which is the direction of the speaker, and the target speaker voice arrival direction θ _{est, j} From the direction θ _i representing the position of the microphone i (i = 0, 1, 2,..., M) and the distance r _i , the microphone i (i = 0, 1, 2,. ) Weight w _{i, j} and the target speaker voice arrival for the past t frames (t is an integer of 1 or more) for each frame from the target speaker voice arrival direction θ _{est, j.} A target speaker voice arrival direction time change calculating unit for calculating a target speaker voice arrival direction time change Δθ _j which is a change in direction; and the speech likelihood L _{speech_i, j} , the non-voice likelihood L _{noise_i, j} ( i = 0, 1, 2,..., M), the weights w _{i, j} (i = 0, 1, 2,..., M) and the target speaker voice arrival direction time change Δθ _j are weighted for each frame. generating a weighted likelihood ratio calculation unit for calculating a likelihood ratio L _j, from the weighted likelihood ratios L _j, for each frame, the speech segment determination result j is determined whether the result is a voice section A speech segment determination unit.

In one embodiment of the present invention, the microphone installed in the direction in which the target speaker is supposed to speak (hereinafter referred to as the front direction) is the microphone 0, and the other M (M ≧ 1) microphones are the microphone 1. , Microphones 2,..., Microphone M, M + 1 microphones i (i = 0) with the front direction as viewed from the microphone 0 being a 0 degree direction and the microphone 0 being the center and the 0 degree direction as a reference , 1, 2,..., M) as θ _i (0 ≦ θ _i ≦ 2π, where θ ₀ = 0), the microphone 0 and the M + 1 microphones i (i = 0, 1, 2, ..., M) is the distance r _i (r _i ≥0, r ₀ = 0), and the target speaker's voice collected by the microphone i (i = 0, 1, 2, ..., M) is The included speech data x _i (i = 0, 1, 2,..., M) is divided into frames, and for each frame, speech likelihood L _{speech_i, j} and non-speech likelihood L _{noise_i, j} (i = 0, 1 , 2,…, M, j are the frame numbers A likelihood calculator for calculating the box), the audio data _{x i (i = 0, 1} , 2, ..., purpose M), which for each frame, relative to the direction of 0 degrees about the microphone 0 A target speaker voice arrival direction estimation unit for estimating a target speaker voice arrival direction θ _{est, j} (0 ≦ θ _{est, j} ≦ 2π), which is the direction of the speaker, and the target speaker voice arrival direction θ _{est, j} From the direction θ _i representing the position of the microphone i (i = 0, 1, 2,..., M) and the distance r _i , the microphone i (i = 0, 1, 2,. ) Weight w _{i, j} , the speech likelihood L _{speech_i, j} , the non-speech likelihood L _{noise_i, j} (i = 0, 1, 2,..., M), and the weight A weighted likelihood ratio calculation unit that calculates a weighted likelihood ratio L _j for each frame from w _{i, j} (i = 0, 1, 2,..., M), and the weighted likelihood ratio L _j To, for each frame, a speech segment that is a determination result of whether or not it is a speech segment And a speech segment determination unit for generating a constant results j.

  According to the present invention, the direction in which the target speaker's voice arrives is estimated, the voice from the direction is emphasized, and the voice section by the target speaker is detected. It becomes possible to detect the voice section of the speaker with high accuracy.

The figure which shows an example of a sound collection system. The figure which shows an example of a structure of the audio | voice area detection apparatus 100. FIG. The figure which shows an example of operation | movement of the audio | voice area detection apparatus 100. The figure which shows an example of a structure of the likelihood calculation part 110. FIG. The figure which shows an example of operation | movement of the likelihood calculation part 110. The figure which shows an example of the target speaker audio | voice arrival direction. The figure which shows an example of a structure of the weighted likelihood ratio calculation part 140. FIG. The figure which shows an example of operation | movement of the weighted likelihood ratio calculation part 140. FIG. The figure which shows an example of a structure of the audio | voice area detection apparatus. The figure which shows an example of operation | movement of the audio | voice area detection apparatus 200. The figure which shows an example of a structure of the weighted likelihood ratio calculation part 240. FIG.

  Hereinafter, embodiments of the present invention will be described in detail. In addition, the same number is attached | subjected to the structure part which has the same function, and duplication description is abbreviate | omitted.

<Definition>
Hereinafter, terms used in each embodiment will be described.

  First, a sound collection system composed of a plurality of microphones that collect the speech of the target speaker will be described.

[Sound collection system]
The sound collection system is composed of one main microphone and M (M ≧ 1) sub-microphones. The main microphone is a microphone that is installed in a direction in which the target speaker is supposed to speak (hereinafter referred to as the front direction). For example, in the case of a microphone mounted on a communication robot, the microphone is attached to the front of the robot and is directed in the front direction. The sub microphone means a microphone other than the main microphone.

  Hereinafter, the main microphone is referred to as microphone 0, and the M sub-microphones are referred to as microphone 1, microphone 2,..., Microphone M, respectively.

  A straight line is drawn from the microphone 0 so as to be perpendicular to the ground, and a plane P passing through the microphone 0 with the straight line as a normal is considered (see FIG. 1). The direction in which the target speaker is supposed to speak on the plane P, that is, the front direction when viewed from the microphone 0 is defined as a direction of 0 degree.

The direction of M + 1 microphones i (i = 0, 1, 2,..., M) (hereinafter referred to as the direction of the microphone i with respect to the microphone 0) with respect to the direction of 0 degree centered on the microphone 0 is θ _i (i = 0, 1, 2,..., M) (0 ≦ θ _i ≦ 2π, where θ ₀ = 0). Also, let r _i (i = 0, 1, 2,..., M) be the distance between microphone 0 and M + 1 microphones i (i = 0, 1, 2,..., M) (r _i ≧ 0 , r ₀ = 0). That is, the position of the microphone _i is represented by the direction θ _i of the microphone _i with respect to the microphone 0 and the distance r _i between the microphone 0 and the microphone i.

FIG. 1 shows an example (M = 2) of a sound collection system including a main microphone (microphone 0) and two sub-microphones (microphone 1 and microphone 2). Although FIG. 1 shows an example in which the sub microphone is on the plane P, the sub microphone may not necessarily be on the plane P. If the plane does not exist on the plane P, the direction θ _{i of the} microphone i with respect to the microphone 0 (where i is an integer between 1 and M) is defined by using the intersection of the plane P and the perpendicular drawn from the sub microphone to the plane P.

<First embodiment>
Hereinafter, the speech section detection apparatus 100 will be described with reference to FIGS.

As shown in FIG. 2, the speech section detection apparatus 100 includes a likelihood calculation unit 110, a target speaker voice arrival direction estimation unit 120, a weight calculation unit 130, a target speaker voice arrival direction time change calculation unit 135, a weighted likelihood. A ratio calculation unit 140, a voice segment determination unit 150, and a recording unit 190 are included. The recording unit 190 is a component that appropriately records information necessary for processing of the speech segment detection device 100. For example, the direction θ _i and the distance r _i representing the position of the microphone i (i = 0, 1, 2,..., M) used in the weight calculation unit 130 and the threshold th used in the speech section determination unit 150 are recorded in advance. Keep it.

The speech section detection apparatus 100 includes M + 1 speech data collected by the sound collection system, that is, speech including the speech of the target speaker collected by the microphone i (i = 0, 1, 2,..., M). Using data x _i as an input, a speech segment determination result that is a determination result of whether each frame is a speech segment is generated and output.

The operation of the speech segment detection device 100 will be described with reference to FIG. Likelihood calculation section 110 includes speech data x _i (i = 0, 1, 2,..., M) including the speech of the target speaker picked up by microphone _i (i = 0, 1, 2,..., M). _Is divided into frames, and for each frame, speech likelihood L _{speech_i, j} and non-speech likelihood L _{noise_i, j} (i = 0, 1, 2,..., M, j is an index representing a frame number) are calculated. (S110). The index j may be, for example, j = 0, 1,.

Hereinafter, the likelihood calculation unit 110 will be described with reference to FIGS. As shown in FIG. 4, the likelihood calculation unit 110 includes a feature amount extraction unit 111 and a likelihood calculation unit 113. The operation of the likelihood calculation unit 110 will be described with reference to FIG. The feature amount extraction unit 111 divides the input audio data x _i (i = 0, 1, 2,..., M) into frames, and extracts a feature amount j for each frame (S111). As the feature amount, MFCC (Mel-Frequency Cepstrum Coefficients) or power may be used. The likelihood calculating unit 113 calculates the speech likelihood L _{speech_i, j} and the non-speech likelihood L _{noise_i, j} (i = 0, 1, 2,..., M) for each frame from the feature quantity j extracted in S111. Calculate (S113). For example, there is a method using a Gaussian mixture distribution model GMM (Gaussian Mixture Model) (Reference Non-Patent Document 1). In this method, a speech likelihood and a non-speech likelihood are calculated from a speech GMM and a non-speech GMM, respectively.
(Reference Non-Patent Document 1) Masayoshi Fujimoto, Yasuo Ariki, “Speech recognition under noise by using speech signal estimation based on GMM and speech enhancement based on time domain SVD”, IEICE Transactions, D-II , Vol.J88-D-II, No.2, pp.250-265, 2005.

The target speaker voice arrival direction estimation unit 120 uses the direction of 0 degrees with respect to the frame 0 as a reference for each frame from the input voice data x _i (i = 0, 1, 2,..., M). The target speaker voice arrival direction θ _{est, j} (0 ≦ θ _{est, j} ≦ 2π), which is the direction of the target speaker, is estimated (S120). FIG. 6 shows the positional relationship between the target speaker and the microphone 0. The technique of Reference Non-Patent Document 2 can be used to estimate the direction of arrival of the target speaker voice.
(Reference Non-Patent Document 2) CH Knapp, GC Carter, “The Generalized Correlation Method for Estimation of Time Delay”, IEEE Transactions on Acoustics, Speech and Signal Processing, Vol.24, No.4, pp.320-327, 1976 .

  When this estimation technique is used, the target speaker can be focused using a plurality of microphones, and the influence of ambient noise can be reduced.

The weight calculator 130 calculates the target speaker voice arrival direction θ _{est, j} estimated in S120, the direction θ _i representing the position of the microphone i (i = 0, 1, 2,..., M), and the distance r _i . For each frame, the weights w _{i, j} of the microphone i (i = 0, 1, 2,..., M) are calculated (S130). Here, the weight of the microphone i is a real number determined on the basis of the distance between the microphone 0 and the microphone i and the deviation between the direction of the microphone i relative to the microphone 0 and the arrival direction of the target speaker voice. For example, the weights w _{i, j} (i = 0, 1, 2,..., M) are calculated using Expression (1).

Here, 1/2 (1 + r _i ) is a term that depends on the distance r _i to the microphone 0, and takes a larger value as the distance from the microphone 0 is closer. Cos (θ _i −θ _{est, j} ) is a term that depends on the direction θ _i with respect to the microphone 0 and the direction of arrival of the target speaker voice θ _{est, j} , and changes from 0 to π (that is, θ _i and θ _The larger the deviation of _{est, j} ), the smaller the value. w _{0, j} represents the weight of the microphone 0 with respect to the frame j, and the previous equation (1) becomes the following equation (2).

Therefore, when the target speaker voice arrival direction θ _{est, j} is the front direction (θ _{est, j} = 0), w _{0, j} = 1.

In general, the weights w _{i, j} are the distance between the microphones 0 and r, the direction of the microphone with respect to the direction of 0 degrees with respect to the microphone 0 as the reference, and the direction of 0 degrees with respect to the microphone 0 as the reference. as the target speaker voice arrival direction θ _est is the direction of the target speaker, distance r and is a deviation in the direction θ and the target speaker voice arrival direction θ _est θ-θ _est of the function w (r, θ-θ _est ) (Where r is a real number greater than or equal to 0 and θ-θ _est is a real number). The function w (r, θ−θ _est ) is a monotonically decreasing function with respect to r. The function _{w (r, θ-θ est} ) , for theta-theta _est, maximum when θ-θ _est = 0, so as to minimize the time of θ-θ _est = π monotonically decreasing, w (r , − (θ−θ _est )) = w (r, θ−θ _est ). The weights w _{i, j} may be calculated using the function w (r, θ−θ _est ) having the above properties.

The target speaker voice arrival direction time variation calculation unit 135 calculates the target speaker voice for the past t frames (t is an integer of 1 or more) for each frame from the target speaker voice arrival direction θ _{est, j} estimated in S120. A target speaker voice arrival direction time change Δθ _j which is a change in the arrival direction is calculated (S135). For example, when one frame is 20 msec, Δθ _j may be calculated with t = 5 or so. That is, the target speaker voice arrival direction theta _est of the current _{frame, j} and its 5 frames before the target speaker voice arrival direction theta _{est, j-5} of the difference theta _est, and _j - [theta] _est, the _j-5 [Delta] [theta] _j (Δθ _j = θ _{est, j} -θ _{est, j-5} ). Also, Δθ _j can be calculated by recording the target speaker voice arrival direction in the recording unit 190 sequentially. This process assumes that the position of the target speaker does not move significantly in a very short time interval (that is, the target speaker's voice to be picked up is normal), and the time change Δθ _{j in the} direction of arrival of the target speaker's voice Is a process to be taken into consideration in the process in the weighted likelihood ratio calculation unit 140 described later.

The weighted likelihood ratio calculation unit 140 includes the speech likelihood L _{speech_i, j} calculated in S110, the non-speech likelihood L _{noise_i, j} (i = 0, 1, 2,..., M), and the weight calculated in S130. A weighted likelihood ratio L _j is calculated for each frame from w _{i, j} (i = 0, 1, 2,..., M) and the target speaker voice arrival direction time change Δθ _j calculated in S135 ( S140). Hereinafter, the weighted likelihood ratio calculation unit 140 will be described with reference to FIGS. As shown in FIG. 7, the weighted likelihood ratio calculation unit 140 includes a weighted likelihood calculation unit 141 and a weighted likelihood ratio calculation unit 143. The operation of the weighted likelihood ratio calculation unit 140 will be described with reference to FIG.

The weighted likelihood calculating unit 141 includes a speech likelihood L _{speech_i, j} , a non-speech likelihood L _{noise_i, j} (i = 0, 1, 2,..., M) and a weight w _{i, j} (i = 0, 1, 2,..., M), a weighted speech likelihood L _{speech_all, j} and a weighted non-speech likelihood L _{noise_all, j} are calculated for each frame (S141). The weighted speech likelihood L _{speech_all, j} and the weighted non-speech likelihood L _{noise_all, j} are calculated using weighted addition as shown in equations (3a) and (3b), respectively.

The weighted likelihood ratio calculation unit 143 calculates each frame from the weighted speech likelihood L _{speech_all, j} calculated in S141, the weighted non-speech likelihood L _{noise_all, j,} and the target speaker speech arrival direction time change Δθ _j. The weighted likelihood ratio L _j is calculated for (S143). The weighted likelihood ratio L _j is calculated using Equation (4).

Here, K is a parameter for adjusting the degree to which a frame is determined to be a speech section. Usually, K = 1 is sufficient. If it is judged that it is better to miss the correct speech segment that is a frame that includes speech even if a segment other than speech is erroneously detected, the value of K is adjusted to a value greater than 1 and determined as a speech segment It should be easy to be done.

Further, as the target speaker voice arrival direction time change Δθ _j increases, the weighted likelihood ratio L _j decreases. The reason why the target speaker voice arrival direction time variation Δθ _j is reflected in the calculation of the weighted likelihood ratio L _j in this way is as follows. For example, considering the conversation between the communication robot and the target speaker, the target speaker can basically stop and speak. As a result, when the target speaker is speaking because the target speaker is not misaligned (the target speaker voice arrival direction time change Δθ _j is small), the target speaker is misaligned (the target speaker 's voice direction of arrival time change Δθ _j is large) target speaker from it can be determined that not speaking. Therefore, the target speaker voice arrival direction time variation Δθ _j is used for weighting in the form as shown in equation (4) to calculate the weighted likelihood ratio L _j .

Generally, the weighted likelihood ratio L _j is the ratio of the weighted speech likelihood L _{Speech_all} and weighted non-speech likelihood _L noise_all L speech_all / L noise_all, past t frame (t is an integer of 1 or more) The calculation is performed using a function L (L _{speech_all} / L _{noise_all} , Δθ) of the target speaker voice arrival direction time change Δθ, which is a change in the target speaker voice arrival direction. The function L (L _{speech_all} / L _{noise_all} , Δθ) is a monotonically increasing function for L _{speech_all} / L _{noise_all} . In addition, the function L (L _{speech_all} / L _{noise_all} , Δθ) monotonically decreases with respect to Δθ so as to be maximum when Δθ = 0 and minimum when Δθ = π, and L (L _{speech_all} / L _{noise_all} , −Δθ) = L (L _{speech_all} / L _{noise_all} , Δθ) is a function of period 2π. The weighted likelihood ratio L _j may be calculated using the function L (L _{speech_all} / L _{noise_all} , Δθ) having the above properties.

The speech segment determination unit 150 generates a speech segment determination result j, which is a determination result of whether or not each frame is a speech segment, from the weighted likelihood ratio L _j calculated in S140 (S150). For example, the weighted likelihood ratio L _j is compared with the threshold th, and when the weighted likelihood ratio L _j is greater than the threshold th (or greater than or equal to the threshold th), it is determined that it is a speech section, and the frame j is A speech segment determination result j indicating that it is a speech segment is generated. Otherwise, it is determined that it is a non-speech segment, and a speech segment determination result j indicating that the frame j is a non-speech segment is generated.

  According to the embodiment of the present invention, the speech likelihood / weighting weighted so as to estimate the direction in which the target speaker's voice arrives using voice data collected by a plurality of microphones and to emphasize the voice from the direction. By detecting the speech section by the target speaker using the non-speech likelihood, the speech section of the target speaker can be accurately detected even in a noisy environment.

  In addition, since voice sections can be cut out with high accuracy, the reliability of voice recognition can be improved.

<Second embodiment>
In the first embodiment, whether or not it is a speech section is determined based on a weighted likelihood ratio reflecting a change in time of arrival direction of the target speaker voice indicating the degree of movement of the target speaker. The weighted likelihood ratio may be calculated without considering the motion. Therefore, in the present embodiment, a method for determining whether or not a speech section is determined without using the target speaker voice arrival direction time change will be described.

  Hereinafter, the speech section detection apparatus 200 will be described with reference to FIGS. 9 to 10.

  As shown in FIG. 9, the speech segment detection apparatus 200 includes a likelihood calculation unit 110, a target speaker speech arrival direction estimation unit 120, a weight calculation unit 130, a weighted likelihood ratio calculation unit 240, a speech segment determination unit 150, a recording Part 190.

The speech section detection apparatus 200 includes M + 1 speech data collected by the sound collection system, that is, speech including the speech of the target speaker collected by the microphone i (i = 0, 1, 2,..., M). Using data x _i as an input, a speech segment determination result that is a determination result of whether each frame is a speech segment is generated and output.

The operation of the speech segment detection device 200 will be described with reference to FIG. Likelihood calculation section 110 includes speech data x _i (i = 0, 1, 2,..., M) including the speech of the target speaker picked up by microphone _i (i = 0, 1, 2,..., M). _Is divided into frames, and for each frame, speech likelihood L _{speech_i, j} and non-speech likelihood L _{noise_i, j} (i = 0, 1, 2,..., M, j is an index representing a frame number) are calculated. (S110).

The target speaker voice arrival direction estimation unit 120 uses the direction of 0 degrees with respect to the frame 0 as a reference for each frame from the input voice data x _i (i = 0, 1, 2,..., M). The target speaker voice arrival direction θ _{est, j} (0 ≦ θ _{est, j} ≦ 2π), which is the direction of the target speaker, is estimated (S120).

The weight calculator 130 calculates the target speaker voice arrival direction θ _{est, j} estimated in S120, the direction θ _i representing the position of the microphone i (i = 0, 1, 2,..., M), and the distance r _i . For each frame, the weights w _{i, j} of the microphone i (i = 0, 1, 2,..., M) are calculated (S130).

The weighted likelihood ratio calculation unit 240 uses the speech likelihood L _{speech_i, j} calculated in S110, the non-speech likelihood L _{noise_i, j} (i = 0, 1, 2,..., M), and the weight calculated in S130. A weighted likelihood ratio L _j is calculated for each frame from w _{i, j} (i = 0, 1, 2,..., M) (S240). Hereinafter, the weighted likelihood ratio calculation unit 240 will be described with reference to FIG. As shown in FIG. 11, the weighted likelihood ratio calculation unit 240 includes a weighted likelihood calculation unit 141 and a weighted likelihood ratio calculation unit 243.

The weighted likelihood calculating unit 141 includes a speech likelihood L _{speech_i, j} , a non-speech likelihood L _{noise_i, j} (i = 0, 1, 2,..., M) and a weight w _{i, j} (i = 0, 1, 2,..., M), a weighted speech likelihood L _{speech_all, j} and a weighted non-speech likelihood L _{noise_all, j} are calculated for each frame (S141).

The weighted likelihood ratio calculation unit 243 calculates a weighted likelihood ratio L _j for each frame from the weighted speech likelihood L _{speech_all, j} and the weighted non-speech likelihood L _{noise_all, j} calculated in S141. (S243). The weighted likelihood ratio L _j is calculated using Equation (5).

  It should be noted that a form that does not use K, that is, K = 1 may be used.

The speech segment determination unit 150 generates a speech segment determination result j that is a determination result of whether or not each frame is a speech segment for each frame from the weighted likelihood ratio L _j calculated in S240 (S150).

  According to the embodiment of the present invention, the speech likelihood / weighting weighted so as to estimate the direction in which the target speaker's voice arrives using voice data collected by a plurality of microphones and to emphasize the voice from the direction. By detecting the speech section by the target speaker using the non-speech likelihood, the speech section of the target speaker can be accurately detected even in a noisy environment.

  In addition, since voice sections can be cut out with high accuracy, the reliability of voice recognition can be improved.

<Modification>
The present invention is not limited to the above-described embodiment, and it goes without saying that modifications can be made as appropriate without departing from the spirit of the present invention. The various processes described in the above embodiment may be executed not only in time series according to the order of description, but also in parallel or individually as required by the processing capability of the apparatus that executes the processes or as necessary.

<Supplementary note>
The apparatus of the present invention includes, for example, a single hardware entity as an input unit to which a keyboard or the like can be connected, an output unit to which a liquid crystal display or the like can be connected, and a communication device (for example, a communication cable) capable of communicating outside the hardware entity. Can be connected to a communication unit, a CPU (Central Processing Unit, may include a cache memory or a register), a RAM or ROM that is a memory, an external storage device that is a hard disk, and an input unit, an output unit, or a communication unit thereof , A CPU, a RAM, a ROM, and a bus connected so that data can be exchanged between the external storage devices. If necessary, the hardware entity may be provided with a device (drive) that can read and write a recording medium such as a CD-ROM. A physical entity having such hardware resources includes a general-purpose computer.

  The external storage device of the hardware entity stores a program necessary for realizing the above functions and data necessary for processing the program (not limited to the external storage device, for example, reading a program) It may be stored in a ROM that is a dedicated storage device). Data obtained by the processing of these programs is appropriately stored in a RAM or an external storage device.

  In the hardware entity, each program stored in an external storage device (or ROM or the like) and data necessary for processing each program are read into a memory as necessary, and are interpreted and executed by a CPU as appropriate. . As a result, the CPU realizes a predetermined function (respective component requirements expressed as the above-described unit, unit, etc.).

  The present invention is not limited to the above-described embodiment, and can be appropriately changed without departing from the spirit of the present invention. In addition, the processing described in the above embodiment may be executed not only in time series according to the order of description but also in parallel or individually as required by the processing capability of the apparatus that executes the processing. .

  As described above, when the processing functions in the hardware entity (the apparatus of the present invention) described in the above embodiments are realized by a computer, the processing contents of the functions that the hardware entity should have are described by a program. Then, by executing this program on a computer, the processing functions in the hardware entity are realized on the computer.

  The program describing the processing contents can be recorded on a computer-readable recording medium. As the computer-readable recording medium, for example, any recording medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, and a semiconductor memory may be used. Specifically, for example, as a magnetic recording device, a hard disk device, a flexible disk, a magnetic tape or the like, and as an optical disk, a DVD (Digital Versatile Disc), a DVD-RAM (Random Access Memory), a CD-ROM (Compact Disc Read Only). Memory), CD-R (Recordable) / RW (ReWritable), etc., magneto-optical recording medium, MO (Magneto-Optical disc), etc., semiconductor memory, EEP-ROM (Electronically Erasable and Programmable-Read Only Memory), etc. Can be used.

  The program is distributed by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Furthermore, the program may be distributed by storing the program in a storage device of the server computer and transferring the program from the server computer to another computer via a network.

  A computer that executes such a program first stores, for example, a program recorded on a portable recording medium or a program transferred from a server computer in its own storage device. When executing the process, the computer reads a program stored in its own recording medium and executes a process according to the read program. As another execution form of the program, the computer may directly read the program from a portable recording medium and execute processing according to the program, and the program is transferred from the server computer to the computer. Each time, the processing according to the received program may be executed sequentially. Also, the program is not transferred from the server computer to the computer, and the above-described processing is executed by a so-called ASP (Application Service Provider) type service that realizes the processing function only by the execution instruction and result acquisition. It is good. Note that the program in this embodiment includes information that is used for processing by an electronic computer and that conforms to the program (data that is not a direct command to the computer but has a property that defines the processing of the computer).

  In this embodiment, a hardware entity is configured by executing a predetermined program on a computer. However, at least a part of these processing contents may be realized by hardware.

Claims (5)

Microphone is set to microphone 0 in the direction in which the target speaker is supposed to speak (hereinafter referred to as the front direction), M microphones (M ≧ 1) are set to microphone 1, microphone 2, ..., microphone M
The front direction when viewed from the microphone 0 is a direction of 0 degrees,
The direction of M + 1 microphones i (i = 0, 1, 2,..., M) with reference to the direction of 0 degrees with respect to the microphone 0 is defined as θ _i (0 ≦ θ _i ≦ 2π, where θ ₀ = 0), and the distance between the microphone 0 and M + 1 microphones i (i = 0, 1, 2,..., M) is r _i (r _i ≧ 0, r ₀ = 0),
Voice data x _i (i = 0, 1, 2,..., M) including the voice of the target speaker picked up by the microphone i (i = 0, 1, 2,..., M) is divided into frames, For each frame, a likelihood calculator for calculating speech likelihood L _{speech_i, j} , non-speech likelihood L _{noise_i, j} (i = 0, 1, 2,..., M, j is an index representing a frame number);
From the speech data x _i (i = 0, 1, 2,..., M), the arrival of target speaker speech that is the direction of the target speaker with respect to the direction of 0 degrees around the microphone 0 for each frame. A target speaker voice arrival direction estimation unit that estimates a direction θ _{est, j} (0 ≦ θ _{est, j} ≦ 2π);
From the target speaker voice arrival direction θ _{est, j} , the direction θ _i representing the position of the microphone i (i = 0, 1, 2,..., M) and the distance r _i , the microphone i ( a weight calculation unit for calculating weights w _{i, j} of i = 0, 1, 2,.
From the target speaker voice arrival direction θ _{est, j} , for each frame, the target speaker voice arrival direction time change Δθ that is a change in the target speaker voice arrival direction for the past t frames (t is an integer of 1 or more). _The target speaker voice arrival direction time change calculation unit for calculating _j ,
The speech likelihood L _{speech_i, j} , the non-speech likelihood L _{noise_i, j} (i = 0, 1, 2,..., M) and the weight w _{i, j} (i = 0, 1, 2 _,. M) and a weighted likelihood ratio calculation unit for calculating a weighted likelihood ratio L _j for each frame from the target speaker voice arrival direction time change Δθ _j ,
A speech segment detection device comprising: a speech segment determination unit that generates a speech segment determination result j that is a determination result of whether or not each frame is a speech segment for each frame from the weighted likelihood ratio L _j . Microphone is set to microphone 0 in the direction in which the target speaker is supposed to speak (hereinafter referred to as the front direction), M microphones (M ≧ 1) are set to microphone 1, microphone 2, ..., microphone M
The front direction when viewed from the microphone 0 is a direction of 0 degrees,
The direction of M + 1 microphones i (i = 0, 1, 2,..., M) with reference to the direction of 0 degrees with respect to the microphone 0 is defined as θ _i (0 ≦ θ _i ≦ 2π, where θ ₀ = 0), and the distance between the microphone 0 and M + 1 microphones i (i = 0, 1, 2,..., M) is r _i (r _i ≧ 0, r ₀ = 0),
Voice data x _i (i = 0, 1, 2,..., M) including the voice of the target speaker picked up by the microphone i (i = 0, 1, 2,..., M) is divided into frames, For each frame, a likelihood calculator for calculating speech likelihood L _{speech_i, j} , non-speech likelihood L _{noise_i, j} (i = 0, 1, 2,..., M, j is an index representing a frame number);
From the speech data x _i (i = 0, 1, 2,..., M), the arrival of target speaker speech that is the direction of the target speaker with respect to the direction of 0 degrees around the microphone 0 for each frame. A target speaker voice arrival direction estimation unit that estimates a direction θ _{est, j} (0 ≦ θ _{est, j} ≦ 2π);
From the target speaker voice arrival direction θ _{est, j} , the direction θ _i representing the position of the microphone i (i = 0, 1, 2,..., M) and the distance r _i , the microphone i ( a weight calculation unit for calculating weights w _{i, j} of i = 0, 1, 2,.
The speech likelihood L _{speech_i, j} , the non-speech likelihood L _{noise_i, j} (i = 0, 1, 2,..., M) and the weight w _{i, j} (i = 0, 1, 2 _,. M), for each frame, a weighted likelihood ratio calculation unit for calculating a weighted likelihood ratio L _j ,
A speech segment detection device comprising: a speech segment determination unit that generates a speech segment determination result j that is a determination result of whether or not each frame is a speech segment for each frame from the weighted likelihood ratio L _j . The speech section detection device according to claim 1 or 2,
The weights w _{i, j} (i = 0, 1, 2,..., M) are defined as the distance between the microphone 0 and the microphone r, and the direction of the microphone relative to the direction of 0 degrees with respect to the microphone 0 as θ. , As the target speaker voice arrival direction θ _est , which is the direction of the target speaker with respect to the direction of 0 degree centered on the microphone 0, the distance r, the direction θ, and the target speaker voice arrival direction θ _est function _{w (r, θ-θ est} ) between the displacement a is theta-theta _est of are those that are calculated using,
The function w (r, θ-θ _est ) is a monotonically decreasing function for r,
The function _{w (r, θ-θ est} ) , for theta-theta _est, maximum when θ-θ _est = 0, so as to minimize the time of θ-θ _est = π monotonically decreasing, w (r, A speech section detection device, which is a function having a period of 2π where − (θ−θ _est )) = w (r, θ−θ _est ). Microphone is set to microphone 0 in the direction in which the target speaker is supposed to speak (hereinafter referred to as the front direction), M microphones (M ≧ 1) are set to microphone 1, microphone 2, ..., microphone M
The front direction when viewed from the microphone 0 is a direction of 0 degrees,
The direction of M + 1 microphones i (i = 0, 1, 2,..., M) with reference to the direction of 0 degrees with respect to the microphone 0 is defined as θ _i (0 ≦ θ _i ≦ 2π, where θ ₀ = 0), and the distance between the microphone 0 and M + 1 microphones i (i = 0, 1, 2,..., M) is r _i (r _i ≧ 0, r ₀ = 0),
Voice data x _i (i = 0, 1, 2,..., M) including the voice of the target speaker picked up by the microphone _i (i = 0, 1, 2,..., M) _Is divided into frames, and for each frame, speech likelihood L _{speech_i, j} and non-speech likelihood L _{noise_i, j} (i = 0, 1, 2,..., M, j is an index representing a frame number) are calculated. A likelihood calculation step;
The speech section detection device is configured to determine the direction of the target speaker from the speech data x _i (i = 0, 1, 2,..., M) with respect to the direction of 0 degrees around the microphone 0 for each frame. A target speaker voice arrival direction estimation step for estimating a target speaker voice arrival direction θ _{est, j} (0 ≦ θ _{est, j} ≦ 2π),
The speech section detection device includes each of the target speaker speech arrival direction θ _{est, j} , the direction θ _i representing the position of the microphone i (i = 0, 1, 2,..., M), and the distance r _i . A weight calculating step for calculating a weight w _{i, j} of the microphone i (i = 0, 1, 2,..., M) for the frame;
The target speech in which the speech section detection device is a change in the target speaker voice arrival direction for the past t frames (t is an integer of 1 or more) for each frame from the target speaker voice arrival direction θ _{est, j.} A target speaker voice arrival direction time change calculating step for calculating a speaker voice arrival direction time change Δθ _j ,
The speech section detection device includes the speech likelihood L _{speech_i, j} , the non-speech likelihood L _{noise_i, j} (i = 0, 1, 2,..., M) and the weight w _{i, j} (i = 0 , 1, 2,..., M) and the target speaker voice arrival direction time change Δθ _j , a weighted likelihood ratio calculating step for calculating a weighted likelihood ratio L _j for each frame;
A speech segment determination step in which the speech segment detection device generates, from the weighted likelihood ratio L _j , a speech segment determination result j that is a determination result of whether or not each frame is a speech segment. Detection method.   A program for causing a computer to function as the speech segment detection device according to any one of claims 1 to 3.

Images