JP2015155982A - Voice section detection device, speech recognition device, method thereof, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide voice section detection technology capable of recognizing the voice of a main speaker with high accuracy, not just in a quiet environment but even in a noisy environment, against the voices of a plurality of speakers mixed in a single microphone in an actual environment.

SOLUTION: A voice section detection device 10 includes: a noise suppression unit 110 for suppressing the noise included in a voice digital signal including a voice, noise, and reverberation by using a voice model and obtaining a noise-suppressed voice digital signal; a reverberation estimation unit 120 for estimating the reverberation component included in the noise-suppressed voice digital signal and obtaining a reverberation signal; a main speaker voice feature emphasis unit 140 for obtaining a noise/reverberation-suppressed voice digital signal that is a difference between the noise-suppressed voice digital signal and the reverberation signal; and a main speaker identification unit 150 for identifying a main speaker voice section that is a section in which the main speaker is talking from the noise/reverberation-suppressed voice digital signal by using the voice model.

COPYRIGHT: (C)2015,JPO&INPIT

Description

  The present invention relates to a technique for detecting a voice section from a voice digital signal, and a voice recognition technique performed for the detected voice section.

  Patent Document 1 shares information such as parameters densely between the speech signal section estimation technique and the noise removal technique, and handles the speech signal section estimation technique and the noise removal technique in an integrated manner, thereby achieving high-accuracy speech. It is known as a denoising technique that makes it possible to perform signal interval estimation and denoising. However, in Patent Document 1, the main speaker voice and the other speaker voice are not considered. The main speaker means a target person, and the main speaker voice is the voice. The other speaker means a person other than the main speaker, and the other speaker voice is the voice.

  Further, Non-Patent Document 1 focuses on the energy ratio of direct sound and reflected sound of other speaker's voice in a mobile environment, and detects the main speaker voice section from reverberation components estimated based on multistep linear prediction. Known as.

JP 2009-210647 A

Kiyoshi Kamado, Satoshi Kohashikawa, Keisuke Kinoshita, Hirokazu Masatoshi, Satoshi Takahashi, "Application of dereverberation method to main speaker speech detection in mobile speech recognition", Proceedings of the Acoustical Society of Japan, 2013 , Pp.145-146

  However, in Patent Document 1, although it is possible to distinguish between a voice and a non-sound, it is impossible to separate the voices of other speakers and the main speakers, and all voices are determined as “speech intervals”. Therefore, the main speaker speech section cannot be estimated.

  Further, in Non-Patent Document 1, when noise is included in an input signal, the accuracy of reverberation estimation decreases due to the influence of noise, and as a result, the extraction accuracy of the main speaker's speech may decrease. In addition, a configuration that performs noise suppression in the previous stage of reverberation estimation is also conceivable, but in that case, noise suppression, reverberation suppression, and speech section detection are performed independently, so noise is excessively suppressed and the accuracy of reverberation estimation is reduced. As a result, the detection accuracy of the main speaker voice may be lowered.

  The present invention is highly accurate not only in a quiet environment but also in a high noise environment with respect to mixed speech (including main speaker speech and other speaker speech) to a single microphone in a real environment. It is an object of the present invention to provide a speech segment detection technique capable of recognizing a main speaker speech.

  In order to solve the above-described problem, according to one aspect of the present invention, a speech segment detection apparatus suppresses noise included in a speech digital signal including speech, noise, and reverberation using a speech model. A noise suppression unit that obtains a suppressed speech digital signal, a reverberation estimation unit that estimates a reverberation component included in the noise suppressed speech digital signal, and a noise reverberation suppression that is a difference between the noise suppressed speech digital signal and the reverberation signal A main speaker voice feature emphasizing unit for obtaining a voice digital signal, and a main speaker identification unit for identifying a main speaker voice section, which is a section in which the main speaker is speaking, from a noise dereverberation voice digital signal using a voice model including.

In order to solve the above problem, according to another aspect of the present invention, in the speech interval detection method, the noise suppression unit is included in a speech digital signal including speech, noise, and reverberation using a speech model. A noise suppression step for obtaining a noise-suppressed speech digital signal by suppressing noise, and a reverberation estimating step for obtaining a reverberation signal by estimating a reverberation component included in the noise-suppressed speech digital signal;
The main speaker voice feature enhancement unit obtains a noise reverberation suppression voice digital signal, which is the difference between the noise suppressed voice digital signal and the reverberation signal, and the main speaker identification unit uses a voice model. And a main speaker identifying step of identifying a main speaker voice section, which is a section in which the main speaker is speaking, from the noise dereverberation suppressed voice digital signal.

  With respect to voices mixed with a plurality of speakers in a single microphone in an actual environment, the main speaker voice can be recognized with high accuracy not only in a quiet environment but also in a high noise environment.

The functional block diagram of the audio | voice area detection apparatus which concerns on 1st embodiment. The figure which shows the example of the processing flow of the audio | voice area detection apparatus which concerns on 1st embodiment. The functional block diagram of the audio | voice area detection noise suppression part which concerns on 1st embodiment. The figure which shows the example of the processing flow of the audio | voice area detection noise suppression part which concerns on 1st embodiment. FIG. 5A is a diagram showing an image of an audio analog signal corresponding to the voice of the main speaker, and FIG. 5B is a diagram showing an image of an audio analog signal corresponding to the external sound of the main speaker. 6A is a diagram illustrating an image in which a reverberation signal corresponding to the voice of the main speaker is represented by a vector, and FIG. 6B is a diagram illustrating an image in which a reverberation signal corresponding to the external sound of the main speaker is represented by a vector. FIG. 7A is a diagram illustrating an image in which a difference corresponding to the voice of the main speaker is represented by a vector, and FIG. 7B is a diagram illustrating an image in which a difference corresponding to the external sound from the main speaker is represented by a vector. The figure for demonstrating arrangement | positioning with the audio | voice area detection apparatus which concerns on 1st embodiment, and the speech recognition apparatus which concerns on 2nd embodiment.

  Hereinafter, embodiments of the present invention will be described. In the drawings used for the following description, constituent parts having the same function and steps for performing the same process are denoted by the same reference numerals, and redundant description is omitted. Further, the processing performed for each element of a vector or matrix is applied to all elements of the vector or matrix unless otherwise specified.

<Points of first embodiment>
The main speaker speech included in the signal in which the noise component processed in Patent Document 1 is suppressed is emphasized using a reverberation estimation method. For example, the reverberation estimation method of Reference 1 can be used.
[Reference Document 1] International Publication No. WO 2007/100137 Pamphlet In Reference Document 1, in the mobile environment, the difference between the reverberation components included in the main speaker voice and the other speaker voice recorded with a single microphone is large. Can be used to enhance the voice of the main speaker with high accuracy. Further, since the signal recorded by the single microphone is not used as it is, but the signal from which the noise component is suppressed is used, the accuracy of reverberation estimation is improved as compared with the case where the original signal is processed as it is.

  Furthermore, in this embodiment, the speech section is determined again by inputting a signal in which noise and reverberation are suppressed from the original signal to the noise removal apparatus of Patent Document 1.

  With such a configuration, it is possible to statistically calculate a speech section using a speech model based on a speech feature amount of a signal in which the speech of the main speaker is accurately emphasized. As described above, it is possible to detect the main speaker voice section with higher accuracy compared to the case where the reverberation signal (logarithm signal obtained by smoothing the power thereof) is determined as a threshold value.

<Audio section detection device according to the first embodiment>
FIG. 1 is a functional block diagram of the speech section detection device 10, and FIG. 2 is a diagram showing an example of the processing flow thereof. For example, the speech section detection device 10 removes the other-speaker speech / silence / noise sections from the input speech (hereinafter also referred to as “speech analog signal”) used for speech recognition processing, thereby accurately processing the speech of the main speaker. Can be recognized.

  First, the voice section of the main speaker included in the input signal to the single microphone is extracted from the reverberation component included in the input signal. Furthermore, by converting the main speaker's speech into speech features, and using this as input, the speech likelihood is calculated together with the speech model for speech segment detection. To do. Further, by using the voice section of the main speaker, the voice of the main speaker can be recognized with high accuracy. Usually, voices and noises other than the main speaker are recognized as voices without being determined as non-speech, and the voice recognition result is generated as misrecognition. Therefore, by extracting only the voice section of the main speaker with high accuracy, it is possible to reduce adverse effects on the voice recognition system due to voice or noise that is not the recognition target.

  The speech segment detection device 10 includes a speech signal acquisition unit 100, a speech segment detection noise suppression unit 110, a reverberation estimation unit 120, a gain adjustment unit 130, a main speaker speech feature enhancement unit 140, and a main speaker speech segment extraction unit 160. . Further, the voice section detection noise suppression unit 110 includes a main speaker identification unit 150. The speech section detection device 10 receives a speech analog signal picked up by the microphone 90, corresponds to the speech section of the main speaker, and a speech digital signal in which noise components are suppressed (hereinafter also referred to as “noise-suppressed speech digital signal”). Is output. Details of each part will be described below.

<Audio signal acquisition unit 100>
Input: Audio analog signal output: Audio digital signal The audio signal acquisition unit 100 receives an analog audio signal (audio analog signal), converts it into a digital audio signal (audio digital signal) (s100), and outputs it.

<Speech Interval Detection Noise Suppression Unit 110>
Input: Voice digital signal Output: Noise suppression voice digital signal Voice section detection noise suppression unit 110 receives a voice digital signal, and uses a voice model to suppress noise contained in the voice digital signal to generate a noise-suppressed voice digital signal. Obtained (s110), and output to the main speaker voice feature enhancement unit 140, the reverberation estimation unit 120, and the main speaker voice section extraction unit 160.

  For example, the speech section detection noise suppression unit 110 performs noise suppression and speech section detection at the same time, and is realized by the noise removal device disclosed in Patent Document 1. An overview of the processing of the speech section detection noise suppression unit 110 will be described.

  FIG. 3 is a functional block diagram of the speech section detection noise suppressing unit 110, and FIG. 4 is a diagram illustrating an example of a processing flow thereof.

  The speech section detection noise suppression unit 110 includes an acoustic signal analysis unit 111, a model parameter storage unit 112, a forward estimation unit 113, a backward estimation unit 114, a parameter storage unit 115, a state probability ratio calculation unit 116, a speech A signal section estimation unit 117 and a noise removal unit 118 are included.

  The acoustic signal analysis unit 111 receives the audio digital signal, extracts the audio feature amount of the audio digital signal for each frame that is a predetermined time interval, and outputs it (s111).

  The model parameter storage unit 112 stores the probability model parameters of the probability model in which the output probabilities of the clean speech signal and the silence signal are expressed by a mixed normal distribution containing a plurality of normal distributions as the above-described speech model before use. Keep it.

  The forward estimation unit 113 receives the speech feature amount and the probability model parameters of the clean speech signal and the silence signal stored in the model parameter storage unit 112, and performs parallel nonlinear Kalman filtering from the past frame to the current frame. The noise model parameters of the current frame are sequentially estimated and output (s113).

  The backward estimation unit 114 receives the noise model parameters output from the forward estimation unit 113, and the probability model parameters of the clean speech signal and the silence signal stored in the model parameter storage unit 112, and inputs the current model from the future frame. The noise model parameters of the current frame are sequentially and backward estimated by the parallel Kalman smoother toward the frame, and each of speech (noise + clean speech) signal and non-speech (noise + silence) signal is determined based on the backward estimated noise model parameter. The probability model parameters of the probability model each representing the output probability with a mixed normal distribution are sequentially estimated, and the output probabilities of the speech signal and the non-speech signal are calculated and output (s114).

  The parameter storage unit 115 stores the calculation results obtained in the process of the forward estimation unit 113 and the backward estimation unit 114 (s115).

  The state probability ratio calculation unit 116 receives the output probabilities of each of the speech signal and the non-speech signal, calculates the speech state probability, the non-speech state probability, and the ratio of the speech state probability to the non-speech state probability. It outputs (s116).

  The speech signal section estimation unit 117 receives the state probability ratio, compares the state probability ratio with the threshold value for each frame, and determines whether each frame belongs to the speech state or the non-speech state. Is output (s117).

  The noise removal unit 118 includes an average for each normal distribution that is each probability model parameter of the speech signal and the non-speech signal, an average for each normal distribution that is each probability model parameter of the clean speech signal and the silence signal, and the speech state probability and The non-voice state probability is input. The noise removal unit 118 compares each average value of each normal distribution, which is each probability model parameter of the clean speech signal and the silence signal, with respect to the average of each normal model that is each probability model parameter of the speech signal and the non-speech signal. Generates a frequency response filter that removes the noise signal by weighted averaging using the speech state probability and the non-speech state probability, converts the frequency response filter to an impulse response filter, and convolves the impulse response filter with the speech digital signal Generates and outputs a noise-suppressed voice digital signal (s118).

  In Patent Document 1, since noise removal is performed only on frames belonging to the speech state based on the determination result of the speech signal section estimation unit 117, processing in the speech signal section estimation unit 117 is necessary. When noise removal is performed, the processing in the speech signal section estimation unit 117 may be omitted.

  The voice section detection noise suppression unit 110 includes a main speaker identification unit 150. The processing content of the main speaker identification unit 150 will be described later.

<Reverberation estimation unit 120>
Input: Noise suppression voice digital signal output: Reverberation signal The reverberation estimation unit 120 estimates a reverberation component included in the noise suppression voice digital signal (s120), and acquires a reverberation signal. Hereinafter, an outline of a method for estimating reverberation components will be described.

  The original audio signal s (z) is obtained by applying a short auto-regressive (hereinafter also referred to as “AR”) process to the white signal u (z) as shown in Expression (1). The Z transformation of the AR process is v (z) = 1 / (1-b (z)), and 1-b (z) is a polynomial.

The original audio signal s (z) is transmitted through the space, and the signal x (z) observed by the microphone is expressed as follows from the equation (1).

  Here, h (z) represents an indoor transfer function from the sound source to the microphone. The audio signal has a strong short-term correlation according to v (z). Therefore, by performing pre-whitening processing by linear prediction that removes short-term correlation according to Equation (3), v (z) can be regarded as a substantially white signal, and v (z) ≈1 holds.

  Here, b (p) is a linear prediction coefficient for effectively suppressing v (z), and is obtained by Expression (4).

  Here, r (i) represents an autocorrelation coefficient when the signal x (z) observed by the microphone is shifted by i samples. This linear prediction is performed with a filter length of 30 ms, and it is expected that the short-term correlation between the early reflection component and the speech included within 30 ms is removed.

  When D is a step size (delay) and L is a filter length, the reverberation signal d (n) can be formulated as follows.

  Here, a (l) (Roman letter L) represents a linear prediction coefficient, and x to (n) represent pre-whitening-processed observation sounds obtained by Expression (3). a (z) obtained by z-converting a (l) is obtained by Expression (6).

Where h _min (z) and h _max (z) are the minimum phase component of h (z) (the component corresponding to the zero in the unit circle on the Z plane) and the maximum phase component (unit on the Z plane). Represents the component corresponding to the zero point outside the circle). Also, min [h _max (z)] represents a function that minimizes h _max (z).

In general, D is set to a value corresponding to 10 to 200 ms, and L is set to a value corresponding to 100 ms to 500 ms.
This technique is detailed in Reference Document 1, for example.

  Using the above-described method and other existing reverberation estimation techniques, the reverberation estimation unit 120 estimates the reverberation component included in the noise-suppressed speech digital signal x (n) and acquires the reverberation signal d (n).

<Gain adjustment unit 130>
Input: Reverberation signal output: Gain-adjusted reverberation signal The gain adjustment unit 130 receives the reverberation signal, multiplies the reverberation signal by the gain G (s130), and obtains and outputs the gain-adjusted reverberation signal. As the gain G, a value smaller than 1 and larger than 0 is used. For example, a value of 0.8 to 1.0 is used. Thereby, in the main speaker voice feature emphasizing unit 140 described later, it is possible to reduce distortion that occurs when obtaining the difference between the noise-suppressed voice digital signal and the reverberation signal.

<Main speaker voice feature enhancement unit 140>
Input: noise-suppressed voice digital signal, gain-adjusted reverberation signal output: noise-reverberation-suppressed voice digital signal The main speaker voice feature enhancement unit 140 receives the noise-suppressed voice digital signal and the gain-adjusted reverberation signal, and The signal difference is calculated (s140) and output as a noise dereverberation speech digital signal. Note that the noise dereverberation speech digital signal may be referred to as a speech digital signal in which the main speaker speech is emphasized.

  FIG. 5A is a diagram illustrating an image of a speech analog signal corresponding to the voice of the main speaker, and FIG. 5B is a diagram illustrating an image of a speech analog signal corresponding to the speech of another speaker. As shown in FIG. 5A, the voice analog signal corresponding to the voice of the main speaker has a large direct sound D and a small reflected sound R (reverberation component). On the other hand, as shown in FIG. 5B, the voice analog signal corresponding to the voice of another speaker has a small direct sound D and a large reflected sound R (reverberation component). 6A is a diagram illustrating an image in which a reverberation signal corresponding to the voice of the main speaker is represented by a vector, and FIG. 6B is a diagram illustrating an image in which a reverberation signal corresponding to the speech of another speaker is represented by a vector. FIG. 7A is a diagram showing an image in which a difference corresponding to the voice of the main speaker is represented by a vector, and FIG. 7B is a diagram showing an image in which a difference corresponding to the voice of another speaker is represented by a vector. A small arrow in FIG. 7 represents a reverberation component R ′ that could not be removed.

  The main speaker speech feature emphasizing unit 140 performs reverberation suppression by calculating a difference between the gain-adjusted reverberation signal and the noise-suppressed speech digital signal. As a result of the subtraction process, the feature of the main speaker voice is emphasized (see FIG. 7), and the feature amount of the main speaker voice in the voice digital signal can be extracted with high accuracy.

  The feature of the main speaker speech here means that the difference in the feature amount on the mel spectrum between the signal after dereverberation and the signal before dereverberation is small. On the other hand, the feature amount of the other-speaker voice changes greatly due to dereverberation, or if the direct wave is unclear due to the characteristics of the dereverberation method, the reverberation cannot be estimated and the sound is silenced. Therefore, the difference between the main speaker and the other speaker is emphasized in the feature amount of the signal after dereverberation and the mel spectrum expression. Since the mel spectrum is a known technique, the description thereof is omitted here. In other words, in the case of other speaker's speech, the difference in the feature amount on the mel spectrum between the signal after dereverberation and the signal before dereverberation increases (because it changes greatly due to dereverberation), or , The difference becomes 0 (because the reverberation cannot be estimated and the reverberation is silenced, it becomes 0, and the signals before and after the reverberation suppression become exactly the same signal). Since the case where the difference is small (main speaker) is clearly different from the case where the difference is large or 0 (other speaker), the difference between the main speaker and the other speaker is emphasized.

The processing of the gain adjustment unit 130 and the main speaker voice feature enhancement unit 140 can be combined and realized by a known technique called a spectral subtraction method (see Reference 2).
[Reference 2] BOLL, SF, "Suppression of Acoustic Noise in Speech Using Spectral Subtraction", IEEE Trans. Acoust., Speech, Signal Processing, 1979, vol. ASSP-27, pp. 113-120

<Main speaker identification unit 150>
Input: noise dereverberation suppressed digital signal output: section information The main speaker identification unit 150 receives the noise dereverberation suppressed digital signal, and identifies the main speaker voice segment from the noise dereverberation suppressed digital signal using a speech model. (S150), the identification result is output as section information. For example, the main speaker identification unit 150 includes an acoustic signal analysis unit 111, a model parameter storage unit 112, a forward estimation unit 113, a backward estimation unit 114, a parameter storage unit 115, a state probability ratio calculation unit 116, A speech signal section estimation unit 117. S111 to s117 are performed using a noise dereverberation suppressed voice digital signal instead of the voice digital signal (see FIG. 4), and the state probability ratio is input to the voice signal interval estimation unit 117 in the main speaker identification unit 150. Then, the ratio of the state probability and the threshold value are compared for each frame, and a determination result indicating whether each frame belongs to the voice state or the non-voice state is output as section information (s117).

  In addition, since the main speaker voice is emphasized in the noise reverberation suppression voice digital signal, it is determined that only the frame corresponding to the main speaker voice belongs to the voice state, and the main speaker outside sound including the other speaker voice is included. It is easy to determine that the frame corresponding to is in the non-voice state. The main speaker outside sound means a sound other than the main speaker voice, and includes other speaker voice, noise, silence and the like.

  As a result, a noise dereverberation-suppressed voice digital signal, which is a signal in which the features of the main speaker voice are emphasized, is input to the noise removal device of Patent Document 1 (more specifically, the acoustic signal analysis unit 111). It is possible to re-calculate the speech interval statistically using the speech model based on the speech feature that the speaker speech is emphasized with high accuracy, and the main speaker speech is more accurate than simple threshold calculation. Section detection is possible.

  Since a delay occurs during the reverberation calculation in the reverberation estimation unit 120, a process for rewinding the time of the output speech section is performed in consideration of the delay.

  In this way, by using a model similar to the speech model at the time of noise suppression processing, the speech signal section estimation technique and the noise suppression technique are handled in an integrated manner, and highly accurate speech section estimation and noise suppression are performed.

<Main speaker voice section extraction unit 160>
Input: Noise-suppressed voice digital signal, section information output: Noise-suppressed voice digital signal corresponding to the voice of the main speaker The main-speaker voice section extraction unit 160 receives the noise-suppressed voice digital signal and the section information, and receives the section information. Then, a portion corresponding to the voice of the main speaker is extracted from the noise-suppressed voice digital signal (s160) and output as an output value of the voice section detection device 10.

  For example, when the start time and the end time are used as the section information, 1 is added to the sample between the start time and the end time, and in order to secure a margin between the start time and the end time, the main speaker voice section 1 of N samples (sample length corresponding to 0.1 to 0.4ms) before start time to switch to speaker outside sound section, M samples (corresponding to 0.1 to 0.4ms) after end time to switch from 1 to 0 Perform margin processing to add 1 of (sample length). The noise-suppressed speech digital data of the main speaker speech section (that is, a time sample sequence in which the portion corresponding to the N samples before the start time to the M sample after the end time is 1 and the other portions are 0) subjected to this margin processing The main speaker's voice can be extracted by multiplying the signal every time sample.

In addition, when a signal with the flag of the main speaker voice section added to the noise-suppressed voice digital signal is used as section information, margin processing is performed on the signal (that is, noise suppression of N samples and M samples at the start and end, respectively) The main speaker voice section flag is added to the voice digital signal), and the noise-suppressed voice digital signal corresponding to the portion to which the main speaker voice section flag is added is extracted. In addition, when a signal with the flag of the main speaker outside sound section added to the noise-suppressed voice digital signal is used as section information, margin processing is applied to the noise-suppressed voice digital signal without the flag of the main speaker outer sound section. Then, a noise-suppressed speech digital signal corresponding to a portion to which the flag of the main speaker outside sound section is not given is extracted.
<Effect>
With respect to mixed speech from a plurality of speakers to a single microphone in an actual environment, it is possible to recognize the main speaker speech with high accuracy not only in a quiet environment but also in a high noise environment. As a result, the number of microphones can be reduced, and the hardware configuration can be reduced.
<Modification>
The main speaker voice segment extraction unit 160 may use the original voice digital signal or the noise dereverberation voice digital signal in place of the noise suppressed voice digital signal. Even in that case, the main speaker voice can be extracted. However, when speech recognition processing is performed in the subsequent stage, it is considered that the recognition accuracy is most enhanced when a noise-suppressed speech digital signal is used.

  The speech section detection device 10 does not necessarily need to include the speech signal acquisition unit 100 when receiving a speech digital signal as an input signal.

  The speech section detection device 10 does not necessarily include the gain adjustment unit 130. In this case, the main speaker voice feature emphasizing unit 140 uses a reverberation signal that has not been gain adjusted.

  The speech segment detection device 10 does not necessarily include the main speaker speech extraction unit 170. The output value (section information) of the main speaker identification unit 150 is output as the output value of the speech section detection device 10.

  The main speaker identification unit 150 is not necessarily a part of the voice section detection noise suppression unit 110. In short, it is only necessary that the speech signal section estimation technology and the noise suppression technology can be handled in an integrated manner by using a model similar to the speech model at the time of noise suppression processing.

<Second embodiment>
FIG. 8 is a diagram for explaining the arrangement of the speech segment detection device 10 and the speech recognition device 800. The speech segment detection device 10 is arranged in the front stage of the speech recognition device 800.

  The voice recognition apparatus 800 performs voice recognition using a signal obtained by the voice section detection apparatus 10 described above using a voice signal as an input. The speech signal is a concept including a speech analog signal, a speech digital signal, a noise-suppressed speech digital signal, and a noise reverberation-suppressed speech digital signal.

  For example, it receives a noise-suppressed speech digital signal corresponding to the speech of the main speaker obtained by the speech section detection device 10 described above, and outputs the speech recognition result.

  Further, for example, section information is output as an output value of the voice section detection device 10 (see the modification of the first embodiment), and as the section information, at least one start time of the main speaker voice section or the main speaker outside sound section and When the end time or the like is used, speech recognition is performed on the noise-suppressed speech digital signal corresponding to the section information, and the speech recognition result is output.

  Further, for example, section information is output as an output value of the voice section detection device 10 (see the modification of the first embodiment), and at least a main speaker voice section or a main speaker outside sound section is added to the voice digital signal as section information. When a signal or the like with one flag is used as section information, speech recognition is performed on the noise-suppressed speech digital signal to which the flag of the main speaker speech section is added, and the speech recognition result is output.

As described above, the speech recognition processing is performed by using the speech digital signal (noise-suppressed speech digital signal, noise-reverberation-suppressed speech digital signal) corresponding to the speech of the main speaker obtained by the speech segment detecting device 10 and the segment information. It is possible to remove the main speaker external sound / silence / noise from the input voice (speech signal) to be used, and perform voice recognition processing only on the main speaker's voice, thereby improving the accuracy. Normally, the main speaker's external sound and noise are recognized as voice without being determined as non-speech, and the voice recognition result is generated as misrecognition. Only with high accuracy, it is possible to reduce adverse effects on speech recognition systems caused by speech and noise that are not recognized.
<Other variations>
The present invention is not limited to the above-described embodiments and modifications. For example, the various processes described above are not only executed in time series according to the description, but may also be executed in parallel or individually as required by the processing capability of the apparatus that executes the processes. In addition, it can change suitably in the range which does not deviate from the meaning of this invention.
<Program and recording medium>
In addition, various processing functions in each device described in the above embodiments and modifications may be realized by a computer. In that case, the processing contents of the functions that each device should have are described by a program. Then, by executing this program on a computer, various processing functions in each of the above devices 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.

  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. Further, 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 storage unit. When executing the process, this computer reads the program stored in its own storage unit and executes the process according to the read program. As another embodiment of this program, a computer may read a program directly from a portable recording medium and execute processing according to the program. Further, each time a program is transferred from the server computer to the computer, 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 includes information provided for processing by the electronic computer and equivalent 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 addition, although each device is configured by executing a predetermined program on a computer, at least a part of these processing contents may be realized by hardware.

Claims (4)

A noise suppression unit that obtains a noise-suppressed voice digital signal by suppressing noise included in the voice digital signal including voice, noise, and reverberation using a voice model;
A reverberation estimation unit that estimates a reverberation component included in the noise-suppressed speech digital signal and obtains a reverberation signal;
A main speaker speech feature enhancement unit for obtaining a noise reverberation suppressed speech digital signal that is a difference between the noise suppressed speech digital signal and the reverberation signal;
Using the voice model, including a main speaker identification unit for identifying a main speaker voice section that is a section in which the main speaker is speaking from the noise reverberation suppression voice digital signal,
Voice segment detection device.   A speech recognition device that performs speech recognition on the speech signal using a signal output from the speech section detection device according to claim 1 with the speech signal as an input. A noise suppression step in which a noise suppression unit obtains a noise-suppressed speech digital signal by suppressing noise included in the speech digital signal including speech, noise, and reverberation using a speech model;
A reverberation estimation unit that estimates a reverberation component included in the noise-suppressed speech digital signal to obtain a reverberation signal; and
A main speaker voice feature enhancement step for obtaining a noise dereverberation suppressed voice digital signal that is a difference between the noise-suppressed voice digital signal and the reverberation signal;
The main speaker identifying unit includes a main speaker identifying step of identifying a main speaker voice section, which is a section in which the main speaker is speaking, from the noise reverberation-suppressed voice digital signal using the voice model.
Voice segment detection method.   A program for causing a computer to function as the speech section detection device according to claim 1 or the speech recognition device according to claim 2.

Images