JP2005157086A - Speech recognition device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a speech recognition device capable of reducing misrecognition even when a plurality of speaker's voices are present. <P>SOLUTION: The speech recognition device comprises: microphones 101a and 101b which pick up a speaker's voice that a speaker speaks; amplifiers 102a and 102b which amplify it; a speaker's voice emphasis part 403 which emphasizes the speaker's voice; power calculators 107a and 107b provided to judge whether a detected speaker's voice is a portion of a conversation; speech detectors 108a and 108b; a speech recognition part 113 which recognizes a detected speaker's voice when a conversation detector 109 or the speech detector 108a or 108b detects the speaker's voice and outputs a predetermined signal when the word recognized by the speech recognition match a predetermined keyword; and a recognition controller 110 which deters the speech recognition part 113 from performing keyword detection when the conversation detector 109 detects a conversation. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to a voice recognition apparatus for recognizing voices of unspecified speakers and controlling other connected apparatuses.

  2. Description of the Related Art Conventionally, as a speech recognition apparatus that assumes a plurality of speakers, a device that uses a plurality of microphones in the vicinity of each speaker and uses signals collected through these microphones is known (for example, patents). Reference 1). FIG. 6 is a conceptual diagram showing a configuration of a conventional speech recognition apparatus disclosed in Patent Document 1. The conventional speech recognition apparatus shown in FIG. 6 operates as described below.

  First, the speaker voice, which is the voice of the speaker, is collected by the microphones 601a to 601c and converted into a sound collection signal that is an electric signal by the microphones 601a to 601c. The collected sound signals obtained by the microphones 601a to 601c are amplified by the amplifiers 602a to 602c and then band-limited by the filters 603a to 603c, respectively. Thereafter, the comparator 604 detects the signal of the filter (hereinafter referred to as filter 603a) that outputs the highest level signal (signal obtained by band limitation) among the filters 603a to 603c, and the CPU (Central Prosecting Unit) 607 controls the voice switching unit 605 in accordance with the detection result to set “voice input” as a voice recognition target, and the voice recognition unit 606 recognizes the voice input as voice. The signals from the microphones 601b and 601c other than the microphone corresponding to the filter 603a that outputs the voice input (in this case, the microphone 601a) are determined as “noise input” that is a signal that is not subject to voice recognition. .

Further, after the noise input spectrum is subtracted from the above-described voice input spectrum to reduce the noise component included in the voice input, the signal after the noise component deletion is input to the voice recognition unit 606. The voice recognition unit 606 activates voice recognition without using the voice recognition start switch by detecting whether or not the voice input matches a predetermined keyword.
JP-A-11-52976 (page 1-5, FIG. 1) The Institute of Electronics, Information and Communication Engineers, "Sound System and Digital Processing", Chapter 7, Signal Processing in Microphone, pp. 173-218

  However, in such a conventional speech recognition apparatus, since keyword detection is always performed during speech recognition standby, an error occurs because keyword detection is performed even for conversations between speakers that originally do not require keyword detection. There was a problem that recognition occurred.

  Moreover, during speech recognition, misrecognition caused by superimposing other speaker speech on the speaker speech to be recognized occurs, and the cause of the misrecognition is superimposed on other speaker speech. Since it cannot be determined whether or not it is due to the situation, it has not been possible to give an instruction to the speaker that speech recognition cannot be performed because other speaker's voice is superimposed. For this reason, there has been a problem that erroneous recognition occurs even if the speech recognition is repeated again.

  The present invention has been made to solve such a problem, and provides a speech recognition apparatus capable of reducing the occurrence of erroneous recognition even when a plurality of speaker voices exist.

  The speech recognition apparatus according to the present invention includes a plurality of sound collecting means for collecting sound including a speaker sound uttered by a speaker, and a plurality of speaker sound candidates for the sound collected by each of the sound collecting means. Is included, speaker voice emphasizing means for emphasizing speaker voice candidates of a plurality of speakers, power calculating means for calculating power of speaker voice candidates emphasized by the speaker voice emphasizing means, Based on the power calculated by the power calculation means, a voice detection means for detecting a speaker voice among the candidate speaker voices, and the voice detection means detects speaker voices of a plurality of speakers within a predetermined time in the past. A conversation detecting means for detecting whether or not the voice is detected, and when the voice detecting means detects a speaker voice, the detected voice of the speaker is recognized, and the words recognized by the voice recognition are predetermined keywords. Voice recognition means for outputting a predetermined signal when the If the speech detection means detects that the voice detection means has detected speaker voices of a plurality of speakers within a certain past time, the speech is spoken within the predetermined time when the speaker voices of the plurality of speakers are detected. And a recognition control means for suppressing the voice recognition means from performing voice recognition of the speaker's voice.

  With this configuration, when the conversation detection means is used to determine whether or not the detected speaker voice is a part of the speaker voice constituting the conversation, and if it is determined that the detected speaker voice is the speaker voice constituting the conversation, the keyword Therefore, it is possible to realize a speech recognition apparatus that can reduce the occurrence of misrecognition even when there are a plurality of speaker voices.

  In the speech recognition apparatus of the present invention, the speech recognition means further detects the speaker voice of a plurality of speakers within the predetermined time in the past. When the voice recognition is not performed, there is provided a configuration including guidance output means for generating a signal for notifying that the voice recognition is not performed due to the presence of speaker voices of a plurality of speakers.

  With this configuration, a guidance output unit is provided to detect that a conversation has been detected and voice recognition has not been performed. Therefore, a voice recognition device capable of improving operability can be realized.

  The speech recognition apparatus according to the present invention also includes a plurality of sound collection means for collecting a sound including a speaker sound uttered by a speaker to generate a sound collection signal, and a sound collected by each of the sound collection means. Correlation between speaker voice enhancement means for emphasizing speaker voice candidates for a plurality of speakers and sound pickup signals generated by the plurality of sound pickup means when speaker voice candidates for a plurality of speakers are included Correlation calculating means for calculating a value by changing a delay time between the collected sound signals, and whether or not a delay time at which the correlation value between the collected sound signals is maximum is within a predetermined time range. A speaker direction detecting means for identifying a speaker by detecting a direction in which the speaker is present and detecting a speaker voice in a candidate speaker voice of the specified speaker; and the speaker direction detecting means in the past A conversation detecting means for detecting whether or not speaker voices of a plurality of speakers are detected within a predetermined time, and the speaker Speech recognition that performs speech recognition of the detected speaker speech when the direction detection means detects the speaker speech, and outputs a predetermined signal when the words recognized by speech recognition match a predetermined keyword And when the conversation detecting unit detects that the speaker direction detecting unit has detected speaker voices of a plurality of speakers within a predetermined time in the past, And a recognition control unit that suppresses the voice recognition unit from recognizing the voice of the speaker spoken in time.

  With this configuration, a conversation detecting means is provided, and it is determined whether or not the detected speaker voice is a part of the speaker voice constituting the conversation. Since the detection process is not performed, it is possible to reduce the occurrence of misrecognition even when a plurality of speaker voices exist, and to realize a voice recognition device capable of more accurately specifying a speaker. be able to.

  In the speech recognition apparatus of the present invention, the speech recognition means further detects the speaker voice of a plurality of speakers within the predetermined time in the past. When the voice recognition is not performed, there is provided a configuration including guidance output means for generating a signal for notifying that the voice recognition is not performed due to the presence of speaker voices of a plurality of speakers.

  With this configuration, a guidance output unit is provided to detect that a conversation has been detected and voice recognition has not been performed. Therefore, a voice recognition device capable of improving operability can be realized.

  The present invention can provide a speech recognition apparatus capable of reducing the occurrence of erroneous recognition even when there are a plurality of speaker voices.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings.

(First embodiment)
FIG. 1 is a conceptual diagram showing a block configuration of the speech recognition apparatus according to the first embodiment of the present invention. In FIG. 1, a voice recognition device 100 is configured to perform voice recognition in a vehicle 200 and output a corresponding signal corresponding to a voice-recognized keyword to the outside. A plurality of sound collecting means for collecting included sounds (in FIG. 1, a sound collecting means including a microphone 101a and an amplifier 102a, and a sound collecting means including a microphone 101b and an amplifier 102b) are shown. When the sound collected by the sound collecting means includes speaker voice candidates of a plurality of speakers, the speaker voice enhancing means for emphasizing the speaker voice candidate of one speaker for each sound collecting means (FIG. 1). The speaker voice emphasizing unit 103 (hereinafter referred to as a speaker voice emphasizing unit) 103 and power calculating means (FIG. 1) for calculating the power of the speaker voice candidate emphasized by the speaker voice emphasizing unit 103. This is referred to as a power calculator. 107a and 107b, and voice detecting means for detecting speaker voices among speaker voice candidates based on the powers calculated by the power calculators 107a and 107b (shown as voice detectors in FIG. 1). 108a and 108b, and speech detectors 108a and 108b that detect whether or not a plurality of speaker voices have been detected within a predetermined time in the past (FIG. 1 shows a conversation detector). The term “detector” (hereinafter referred to as “conversation detector”) 109 and the speech detectors 108a and 108b detect the speech of the speaker, and perform speech recognition of the detected speaker speech. Is a voice recognition means (shown as a voice recognition unit in FIG. 1; hereinafter referred to as a voice recognition unit) 113 and voice detectors 108a and 108b. If the conversation detector 109 detects that the speaker voices of a plurality of speakers have been detected within a certain period of time, the voice of the speaker voices uttered within a certain period of time when the speaker voices of the plurality of speakers are detected It is comprised so that the recognition control means (it shows as a recognition controller in FIG. 1 and hereafter is called a recognition controller) 110 which suppresses that the speech recognition part 113 performs recognition.

  Here, the microphones 101a and 101b constituting the sound collecting means are provided in the vehicle 200, for example, and are easy to pick up the speaker voice of the person sitting in each seat. And the sound containing the speaker voice which each speaker uttered is picked up, and a sound pickup signal is generated. The amplifiers 102a and 102b constituting the sound collection means amplify the sound collection signal output from the microphone 101a and generate an amplified sound collection signal. However, when the level of the collected sound signal output from the microphones 101a and 101b is sufficiently high, the amplifiers 102a and 102b may be omitted.

  Here, the speaker voice emphasizing unit 103 further delays the amplified sound pickup signal output from the delay device 104a and the amplifier 102b that generates a delay signal by delaying the amplified sound pickup signal output from the amplifier 102a. A delay unit 104b that generates a delay signal, an adder 105a that adds an amplified sound pickup signal from the amplifier 102a and a delay signal from the delay unit 104b, an amplified sound pickup signal from the amplifier 102b, and a delay signal from the delay unit 104a And an equalizer 106a that changes the frequency spectrum of the output signal of the adder 105a according to a predetermined rule, and an equalizer 106b that changes the frequency spectrum of the output signal of the adder 105b according to a predetermined rule. The The output from the equalizer 106a is input to the power calculator 107a and the buffer 112, and the output from the equalizer 106b is input to the power calculator 107b and the buffer 112.

  Here, the adders 105 a and 105 b add the delayed signals generated by the delay units 104 a and 104 b to the amplified sound collection signals generated by the amplifiers 102 b and 102 a by inverting the signs (hereinafter referred to as addition processing). ) Can remove a delay component and emphasize a specific speaker voice (for example, a speaker voice of a speaker closest to the microphone). Further, the equalizers 106a and 106b are provided so as to restore (close to) the low frequency component whose level has been reduced by performing the addition process to the original frequency spectrum.

  As shown in FIG. 1, the speech recognition apparatus 100 may further include a guidance output unit 114 that outputs a predetermined message or the like under the control of the recognition controller 110. The guidance output unit 114 detects that the conversation detector 109 has detected the speaker voices of a plurality of speakers within a certain period of time in the past (that is, the conversation has occurred). When speech recognition is not performed, a signal may be generated to notify that speech recognition has not been performed due to the presence of speaker speech of multiple speakers. In this case, the amplifier 115 and the speaker 116 may be connected to the speech recognition apparatus 100, and the output signal from the guidance output unit 114 may be amplified by the amplifier 115 and output via the speaker 116.

  As shown in FIG. 1, the speech recognition apparatus 100 further selects the data stored in the buffer 112 under the control of the buffer 112 that stores the output signal from the speaker speech enhancement unit 103 and the recognition controller 110. A guidance output unit 114 that outputs a predetermined message or the like under the control of the input selector 111 and the recognition controller 110, selects the data stored in the buffer 112 under the control of the recognition controller 110, and performs voice recognition Voice recognition may be performed on the data selected by the unit 113.

  Hereinafter, the operation of the speech recognition apparatus 100 according to the first embodiment of the present invention will be described with reference to FIGS. As shown in FIG. 2, the speech recognition apparatus 100 is assumed to operate in two modes: a standby mode and a speech recognition mode. In the standby mode S210, it is detected that a keyword serving as a trigger for starting speech recognition is uttered, and a signal for starting speech recognition by means or methods other than speech such as pressing a speech recognition start button is given. Voice recognition can be started only by speaking without requesting the user.

  The voice recognition mode S220 is a mode in which a voice command is recognized when the keyword that becomes a trigger for starting voice recognition is detected in the standby mode S210. The voice recognition device 100 returns to the standby mode S210 when the operation in the voice recognition mode S220 ends.

  In the standby mode S210, first, the recognition controller 110 determines whether or not an end condition is satisfied (S211). Here, the term “end condition” refers to a condition that should be satisfied when the end process is performed in advance, such as a power-off command for the speech recognition apparatus. If it is determined in step S211 that the end condition is not satisfied, the process proceeds to step S212.

  If it is determined in step S211 that the end condition is not satisfied, the speaker voice candidates in the collected sound signals generated by the microphone 101a and the microphone 101b are emphasized (S212). Hereinafter, the operation for performing the process in step S212 is referred to as “speaker speech enhancement operation”, and details of the speaker speech enhancement operation will be described with reference to FIG.

  First, a speaker voice is collected and a collected signal is generated by the microphones 101a and 101b (S301). Speaker voices spoken by speaker A and speaker B in vehicle 200 are collected by being superimposed on microphone 101a and microphone 101b, respectively. However, when there is a sound source other than the speaker voice such as vehicle running noise, the sound output from the sound source other than the speaker voice is also collected by the microphone 101a and the microphone 101b.

  The collected sound signal is amplified to a predetermined signal level by the amplifier 102a and the amplifier 102b. The signal immediately after the microphones 101a and 101b pick up the speaker's voice is generally an analog signal. However, it is more efficient to perform the signal processing after converting the signal into a digital signal. Therefore, in the following, it is assumed that the collected sound signal is output as a digital signal. Here, for conversion from an analog signal to a digital signal, a so-called AD (Analogue to Digital) converter can be used, and the AD converter is publicly known and description thereof is omitted. The sampling frequency at the time of conversion to a digital signal only needs to cover the frequency band of audio, and generally a range of 10 to 48 kHz is used.

  When the collected sound signal of the speaker voice is generated in step S301, a delay addition process for giving directivity to a specific speaker direction is performed (S302). In the case of the configuration of the present invention, the delay adding process for providing directivity is such that the delay units 104a and 104b delay the collected sound signal (or amplified collected signal) to generate a delayed signal, and the delayed signal is inverted in sign. Then, it is performed by superimposing on the collected sound signal (or amplified collected sound signal).

  Note that the method of imparting directivity is not limited to the above method, and for example, a plurality of methods shown in Non-Patent Document 1 can be applied. Specifically, a so-called delay-and-sum array is also included in which a plurality of microphones are arranged and voices that arrive from a specific direction are added in phase and emphasized. In the first embodiment of the present invention, the delay time of the collected signals from a plurality of microphones is set so that signals arriving from a specific direction cancel each other, and sound from other directions is emphasized, The operation will be described using a subtraction type array.

  The adder 105a constituting the speaker voice enhancement unit 103 arrives from the direction of the speaker A by canceling a signal derived from the voice coming from the direction of the speaker B (hereinafter referred to as a B speaker arrival signal). Processing for emphasizing a signal derived from speech (hereinafter referred to as A-speaker arrival signal) is performed. That is, since the distance between the speaker B and the microphone 101a is longer than the distance between the speaker B and the microphone 101b, the speech uttered by the speaker B reaches the microphone 101b first, and arrives at the microphone 101a with a delay. .

  Therefore, the time difference between the voices of speakers B collected by the microphones 101a and 101b is set by setting the time difference between the voices uttered by the speaker B to the microphones 101a and 101b in the delay unit 104b. After setting so as not to be the input terminal of the adder 105a, the adder 105a adds the signal obtained by inverting the polarity of the output signal of the delay unit 104b and the output signal of the amplifier 102a. As a result, the B speaker arrival signal is canceled from the output signal of the adder 105a, but the A speaker arrival signal is less canceled than the B speaker arrival signal. As a result, the arrival signal of the speaker A is emphasized. The implementation method of the delay device may be a method using a FIR (Finite Impulse Response) filter described in Non-Patent Document 1, or a method using a first-order or second-order all-pass IIR (Infinite Impulse Response) filter. Therefore, it is advantageous.

  The adder 105b performs processing for enhancing the B speaker arrival signal by canceling the A speaker arrival signal. The processing is the same as the processing performed by the adder 105a with the signal to be canceled as the A speaker arrival signal. Thus, the A speaker arrival signal is canceled from the output signal of the adder 105b, but the B speaker arrival signal is not relatively canceled, and therefore the B speaker arrival signal is emphasized with respect to the A speaker arrival signal. Will be. In addition, although the process after step S302 is possible also by the sample process implemented for every sample point, since the method by the frame process which performs a process for every fixed frame length has a small calculation amount, it is advantageous below. The description will be made assuming that frame processing is performed.

  When delay addition is performed in step S302 and the speaker voices from the speakers A and B are emphasized, the frequency spectrum of the speaker voices of the speakers A and B changed by the delay addition is returned to the original frequency spectrum. The process (hereinafter referred to as “equalizing” process) is performed by the equalizer 106a (S303). For example, the change in the frequency spectrum of the A speaker arrival signal is corrected by the equalizing process.

  This is because the adder 105a subtracts the output signal of the microphone 101b from the output signal of the microphone 101a, so that the A-speaker arrival signal has a frequency spectrum in which the sensitivity of the output signal of the adder 105a decreases with decreasing frequency. Is needed by to become. In order to correct the change in the frequency spectrum, the equalizer 106a corrects the level higher as the frequency is lower, and the frequency characteristic of the sensitivity to the voice coming from the direction of the speaker A is flattened. The equalizer 106a may be realized by dividing the input signal into a plurality of frequency band signals, adjusting the gain of each frequency band signal, and then adding the signals in each frequency band. A method of correcting a frequency spectrum using a filter is common.

  Similarly to the equalizer 106a, the equalizer 106b is used to perform an equalizing process for correcting a change in the frequency spectrum of the signal arriving from the direction of the speaker B by the above process performed by the adder 105b. By the correction, the frequency spectrum of the sensitivity with respect to the voice coming from the direction of the speaker B is flattened.

  With the above processing, a signal in which the A speaker arrival signal is emphasized is obtained as the output signal of the equalizer 106a, and a signal in which the B speaker arrival signal is emphasized is obtained as the output signal of the equalizer 106b. As a result, a signal in which the A speaker arrival signal is emphasized and a signal in which the B speaker arrival signal is emphasized are output from the speaker voice enhancement unit 103.

  If the speaker voice candidate is emphasized in step S212, the speaker voice candidate is detected (S213). Specifically, the following processing is performed. First, the power of the output signal from the equalizer 106a is calculated by the power calculator 107a. As a method for calculating the power, a method for calculating all the frequency components of the output signal of the equalizer 106a can be used. However, a method in which the amplitude is squared after the band is limited to a band where the voice is dominant is preferable. The power calculator 107a outputs a signal after time smoothing after the amplitude is squared. The time smoothing time constant is suitably about 100 ms to 1 s. Similarly to the power calculator 107a, the power calculator 107b calculates power using the output signal of the equalizer 106b.

  Based on the power calculated by the power calculators 107a and 107b, the voice detector 108a detects the presence or absence of the voice of the speaker A. As a method for detecting the presence or absence of speech, for example, there is a method of observing the output of the power calculator 107a and determining that there is a speaker speech candidate when the output exceeds a predetermined reference value. As another detection method, smoothing power is calculated by further smoothing the output of the power calculator 107a, and the output of the power calculator 107a is compared with the smoothed power. There is a method of determining that there is a speaker voice candidate when the output exceeds a predetermined relative reference value from the smoothing power.

  The latter method is effective when there is ambient noise because the reference value is relatively determined. Further, as another method for detecting a speaker voice candidate, the output of the power calculator 107a and the output of the power calculator 107b are used to compare the output of the power calculator 107a with the output of the power calculator 107b. There is a method of determining that there is a speaker voice candidate when the output of the calculator 107a exceeds a predetermined relative reference value than the output of the power calculator 107b. This method uses the fact that the output of the power calculator 107a is relatively larger than the output of the power calculator 107b when only the speaker A speaks, and the case where only the speaker A speaks is used. This is effective because it can improve the accuracy of voice detection.

  The voice detector 108b performs the same processing as the voice detector 108a described above, and detects the presence or absence of the speaker B's speaker voice candidate.

  If a speaker voice candidate is detected in step S213, it is determined whether the detected speaker voice candidate is a speaker voice (S214), and at least one of the voice detector 108a and the voice detector 108b is determined. If the speaker voice is detected from any one of them, the process proceeds to step S215, and if neither is detected, the process returns to step S211.

  If it is determined in step S214 that the detected speaker voice candidate is a speaker voice, the conversation detector 109 performs a process of detecting a conversation between the speakers including the speaker voice (S215). . Specifically, when a speaker voice of a certain speaker (for example, speaker A) is detected, a speaker voice of another speaker (in this case, speaker B) is collected within a certain time. Detect. Here, 10 seconds to 30 seconds is appropriate as the “certain time”.

  If the process for detecting the conversation between the speakers is performed in step S215, it is determined whether or not the conversation is detected by the detection process (S216). If it is determined in step S216 that a conversation has been detected, the speech detected as the speaker voice in step S214 is part of the conversation between speaker A and speaker B, and thus keyword detection is performed. The processing returns to step S211 as not being the target voice. If it is determined in step S216 that no conversation is detected, it is determined that the speech is a single utterance by speaker A or speaker B, and the process proceeds to step S217.

  If it is determined in step S216 that the utterance is a single utterance, keyword detection is performed on the uttered voice (S217). The keyword detection is performed using speech recognition, and publicly known speech recognition may be used as speech recognition, and details thereof are omitted because they are not the essence of the present invention. Here, it is assumed that the keyword that is the target of keyword detection is set in advance, and the preset keyword means the start of voice recognition such as “start voice recognition”, “start voice recognition”, etc. It's okay.

  Keyword detection is controlled by the recognition controller 110. Specifically, when it is determined in step S214 that the speaker voice is detected by the voice detector 108a, the input selector 111 selects the output from the equalizer 106a stored in the buffer 112. Further, when sound is detected by the sound detector 108 b, the input selector 111 selects the output from the equalizer 106 b accumulated in the buffer 112. Further, the voice recognition unit 113 performs keyword detection on the output of the input selector 111.

  After performing keyword detection in step S217, it is determined whether or not a preset keyword is detected (S218). When a preset keyword is detected, the operation of the speech recognition apparatus 100 is performed in the speech recognition mode S220. If not, the process returns to step S211.

  The voice recognition device 100 performs processing for receiving a voice recognition command in the voice recognition mode S220. First, it is determined whether or not the mode end condition of the voice recognition mode S220 is satisfied (S221), and when it is determined that a mode end condition such as detection of a keyword representing the end of voice recognition is satisfied, The process proceeds to the standby mode S210. If the mode end condition is not satisfied, the process proceeds to step S222.

  In steps S222 to S224, the same processes as those in steps S212 to S214 are performed. If a speaker voice is detected, the process proceeds to step S225. If not, the process returns to step S221.

  If it is determined in step S224 that speaker voice has been detected, predetermined keyword detection processing such as voice recognition commands, place names, and device settings is performed (S225). The keyword detection process may be performed using publicly known speech recognition, and details of the speech recognition are not the essence of the present invention, and will be omitted. The keyword detection process is controlled by the recognition controller 110. When the speaker voice is detected by the voice detector 108a in step S223, the output of the equalizer 106a stored in the buffer 112 is output by the input selector 111. select. Further, when sound is detected by the sound detector 108b, the input selector 111 selects the output of the equalizer 106b accumulated in the buffer 112.

  When the keyword detection process is performed in step S225, the voice recognition unit 113 determines whether or not the speaker voice subjected to the keyword detection process matches a predetermined keyword. If so, the process proceeds to step S227, and if not, the process proceeds to step S228.

  If it is determined in step S226 that the keyword has been detected, a signal corresponding to the keyword (hereinafter referred to as a corresponding signal) is output or transmitted to a device mounted on the vehicle, a center connected via a network, or the like (S227). ), The process returns to step S221. If it is determined that no keyword is detected, the process proceeds to step S228.

  In step S228, the disturbing speaker is determined. If the voice is detected by both the voice detector 108a and the voice detector 108b in step S223, it is determined that a keyword cannot be detected because a plurality of speakers are speaking, and the processing is performed in step S229. Proceed to Otherwise, the process returns to step S221.

  In step S229, information indicating that speech recognition is impossible due to a plurality of speakers is transmitted to the user. As information transmission means, there are a method of displaying characters and pictures on a monitor or the like, and a method of transmitting by voice. In this embodiment, a method of transmitting by voice is used, and a voice signal is output from the guidance output unit 114, amplified by the amplifier 115, and then output from the speaker 116. The content of the voice to be transmitted may be, for example, “Please quietly except for the driver because voice recognition is performed”, and it is sufficient if the cause of the voice recognition cannot be transmitted to the vehicle occupant. By transmitting such information, even if speech recognition fails due to the presence of another speaker, only the speaker who operates the speech recognition speaks when performing speech recognition again after information transmission. Expectation and speech recognition will be possible. When step S229 ends, the process returns to step S221.

  As described above, the speech recognition apparatus according to the first embodiment of the present invention collects speaker speech with directivity in a plurality of speaker directions, and outputs each directional sound collection output. Conversation detectors that use the voice detection results determined by using them are provided, and when there is a conversation between passengers, keyword detection during voice recognition standby is not performed, thereby suppressing false detection during voice recognition activation. can do.

  In addition, since the guidance output unit is provided so that conversation is detected and voice recognition is not performed, operability can be improved.

(Second Embodiment)
FIG. 4 is a diagram showing a block configuration of a speech recognition apparatus according to the second embodiment of the present invention. In FIG. 4, the speech recognition apparatus 400 according to the second embodiment of the present invention performs speech recognition in the room of the vehicle 200 in the same manner as the speech recognition apparatus 100 according to the first embodiment of the present invention. A corresponding signal corresponding to the keyword is output to the outside, and a plurality of sound collecting means (refer to FIG. 4) for collecting a sound including a speaker voice uttered by a speaker and generating a collected signal. Shows a sound collecting means composed of a microphone 101a and an amplifier 102a, and a sound collecting means composed of a microphone 101b and an amplifier 102b.) And a speaker of a plurality of speakers in the sound collected by each sound collecting means. When speech candidates are included, speaker speech enhancement means for emphasizing the speaker speech candidate of one speaker for each sound collection means (shown as a speaker speech enhancement unit in FIG. 1). 403) and a plurality of sound collecting means are live. Correlation calculation means (shown as a correlation calculator in FIG. 1; hereinafter referred to as a correlation calculator) 401 for calculating the correlation value between the collected sound signals by changing the delay time between the collected sound signals, and between the collected sound signals The speaker is identified and the speaker's voice is identified by detecting the direction of the speaker based on whether or not the delay time that maximizes the correlation value is within a predetermined time range. Speaker direction detection means (in FIG. 1, referred to as a speaker direction detector; hereinafter referred to as a speaker direction detector) 402 and a speaker direction detector 402 in the past. Conversation detection means (shown as a conversation detector in FIG. 1; hereinafter referred to as a conversation detector) 409 for detecting whether or not speaker voices of a plurality of speakers are detected within a predetermined time, and a speaker direction detector When the speaker voice is detected by 402, the voice of the detected speaker voice is recognized, and the voice A speech recognition unit (shown as a speech recognition unit in FIG. 1; hereinafter referred to as a speech recognition unit) 113 that outputs a predetermined signal when a word recognized by the knowledge matches a predetermined keyword; When the conversation detector 409 detects that the direction detector 402 has detected the speaker voices of a plurality of speakers within a predetermined time in the past, the speaker voices of the plurality of speakers have been uttered within a predetermined time. It is configured to include recognition control means (represented as a recognition controller in FIG. 1; hereinafter referred to as a recognition controller) 110 that suppresses speech recognition of the speaker's voice by the voice recognition unit 113.

  Of the constituent means of the speech recognition apparatus 400 according to the second embodiment of the present invention, the same reference numerals are given to the same constituent means as the constituent means of the speech recognition apparatus 100 according to the first embodiment of the present invention. A description thereof will be omitted. The speaker voice emphasizing unit 403 according to the second embodiment of the present invention except that the delay time of the delay units 404a and 404b constituting the speaker voice emphasizing unit 403 can be set during the operation of the speech recognition apparatus 400. For example, the function is the same as that of the speaker voice emphasizing unit 103 according to the first embodiment of the present invention, and a description thereof will be omitted.

  As shown in FIG. 4, the speech recognition apparatus 400 further includes a delay time setting unit 408 that sets the delay times of the delay units 404 a and 404 b, according to the direction of the speaker detected by the speaker direction detector 402. Alternatively, the time set in the delay units 404a and 404b may be updated.

  Hereinafter, the operation of the speech recognition apparatus 400 according to the second embodiment of the present invention will be described with reference to FIG. As shown in FIG. 5, it is assumed that the speech recognition apparatus 400 operates in two modes, a standby mode and a speech recognition mode, as with the speech recognition apparatus 100 according to the first embodiment of the present invention. In the standby mode S510, it is detected that a keyword serving as a trigger for starting speech recognition is uttered, and a signal for starting speech recognition by means or methods other than speech such as pressing a speech recognition start button is given. Voice recognition can be started only by speaking without requesting the user. Moreover, the same number is attached | subjected about the step which performs the process similar to the process in each step demonstrated in the 1st Embodiment of this invention, and the description is abbreviate | omitted.

  After the speaker voice in the collected sound signal is emphasized in step S212, a cross-correlation function between the output signal of the amplifier 102a and the output signal of the amplifier 102b is calculated by the correlation calculator 401 (S5131). As a method of calculating the cross-correlation function, in the time domain, the output signal of the amplifier 102a is shifted from the output signal of the amplifier 102b by a time delay (time lag) τ, and a correlation coefficient between the shifted signal and another signal is calculated. There is a way. As another method of calculating the cross-correlation function, the output signal of the amplifier 102a and the output signal of the amplifier 102b are converted into the frequency domain by a known FFT (Fast Fourier Transform) method or the like, and the cross spectrum of both is calculated. Thereafter, there is a method of calculating by returning to the time domain by a known method such as inverse FFT. This method is characterized by a small amount of calculation.

  When the correlation function is calculated in step S5131, the speaker direction is detected to identify the speaker, and the speaker voice among the speaker voice candidates identified is detected (S5132). The detection of the speaker direction is performed depending on whether or not the time lag τ that takes the maximum value of the cross-correlation function calculated in step S5131 is within the allowable time difference corresponding to the predetermined speaker direction. When it is within the predetermined allowable range, the speaker direction detector 402 determines that there is a speaker in that direction. As an allowable range, it is appropriate to set a time difference corresponding to a geometrically calculated speaker direction ± 10 ° to 30 °. Here, the time lag τ can be information specifying the direction in which the speaker is present.

  When the maximum value of the time lag (hereinafter referred to as correlation time) when the speaker voice is detected in step S214 is different from the preset correlation time, the newly detected correlation time is used as the speaker direction detector 402. Are set in the delay units 404a and 404b (S530). Setting the correlation time in the speaker direction detector 402 is equivalent to setting the speaker direction information. This processing need not be performed for each processing frame, and may be performed at an appropriate time interval longer than the processing frame length using the long-term average value in the detected speaker direction. The configuration may be implemented only at the time of termination or startup. Moreover, since this process is not necessarily essential, a configuration in which step S530 is omitted is also possible.

  When the speaker voice candidate is emphasized in step S222, a correlation function between the output of the amplifier 102a and the output of the amplifier 102b is calculated in the same manner as the processing in step S5131 (S5231). When the correlation function is calculated in step S5231, the speaker direction is detected in the same manner as in step S5232 (S5132). When the speaker direction is detected, the processing in steps S224 and above is performed as described in the first embodiment of the present invention.

  As described above, the speech recognition apparatus according to the second embodiment of the present invention detects the speaker direction using the correlation of signals collected using a plurality of microphones, and uses the speaker direction. It is determined whether or not there is a conversation between passengers, and if a conversation between passengers is being performed, the keyword detection at the time of voice recognition standby is not performed, so that false detection at the time of voice recognition activation is performed. Can be suppressed.

  The speech recognition apparatus according to the present invention has an effect of reducing misrecognition when there are a plurality of speakers, and is useful as an operation device that performs speech recognition and operates an in-vehicle device or the like.

The conceptual diagram which shows the block configuration of the speech recognition apparatus which concerns on the 1st Embodiment of this invention. The flowchart for demonstrating operation | movement of the speech recognition apparatus which concerns on the 1st Embodiment of this invention. The flowchart for demonstrating the speaker audio | voice emphasis operation | movement of the speech recognition apparatus which concerns on the 1st Embodiment of this invention. The conceptual diagram which shows the block structure of the speech recognition apparatus which concerns on the 2nd Embodiment of this invention. The flowchart for demonstrating operation | movement of the speech recognition apparatus which concerns on the 2nd Embodiment of this invention. Conceptual diagram showing a block configuration of a conventional speech recognition apparatus

Explanation of symbols

100, 400, 600 Speech recognition apparatus 101a, 101b, 601a, 601b, 601c Microphone 102a, 102b, 115, 602a, 602b, 602c Amplifier 103, 403 Speaker speech enhancement unit 104a, 104b, 404a, 404b Delay devices 105a, 105b Adder 106a, 106b Equalizer 107a, 107b Power calculator 108a, 108b Speech detector 109, 409 Conversation detector 110 Recognition controller 111 Input selector 112 Buffer 113, 606 Speech recognition unit 114 Guidance output unit 116 Speaker 200 Vehicle 401 Correlation Calculator 402 Speaker direction detector 408 Delay time setting unit 603a, 603b, 603c Filter 604 Comparator 605 Voice switching unit 607 CPU

Claims (4)

  A plurality of sound collecting means for collecting sounds including speaker voices spoken by a speaker, and a plurality of speaker sound candidates of the plurality of speakers included in the sound collected by each of the sound collecting means. Speaker voice emphasizing means for emphasizing speaker's speaker voice candidates, power calculating means for calculating power of speaker voice candidates emphasized by the speaker voice emphasizing means, and power calculated by the power calculating means And a voice detecting means for detecting a speaker voice in the speaker voice candidates based on the voice, and a conversation for detecting whether or not the voice detecting means has detected speaker voices of a plurality of speakers within a predetermined past time. Detecting means and when the voice detecting means detects speaker voice, performs speech recognition of the detected speaker voice, and when a word recognized by voice recognition matches a predetermined keyword, a predetermined signal And voice recognition means for outputting the past and the voice detection means When the conversation detecting unit detects that the speaker voices of a plurality of speakers are detected within a predetermined time, the voice recognition of the speaker voices uttered within the predetermined time when the speaker voices of the plurality of speakers are detected A speech recognition apparatus comprising: a recognition control unit that suppresses the speech recognition unit from performing the operation.   The voice recognition device further detects the voice of the speaker by the voice recognition means when the conversation detection means detects that there is a plurality of speaker voices within the predetermined time in the past. 2. The voice recognition according to claim 1, further comprising a guidance output means for generating a signal notifying that voice recognition has not been performed due to the presence of speaker voices of a plurality of speakers when not performed. apparatus.   A plurality of sound collecting means for collecting a sound including a speaker voice uttered by a speaker to generate a sound collecting signal, and a plurality of speaker voice candidates for the sound collected by each of the sound collecting means If included, speaker voice emphasizing means for emphasizing speaker voice candidates of a plurality of speakers, and a correlation value between the collected sound signals generated by the plurality of sound collecting means is used as a delay time between the collected sound signals. And calculating the direction of the speaker based on whether or not the delay time within which the correlation value between the collected sound signals is maximum is within a predetermined time range. And a speaker direction detecting means for detecting a speaker voice in a speaker voice candidate specified by the speaker, and the speaker direction detecting means is capable of detecting a plurality of speakers within a predetermined time in the past. A conversation detecting means for detecting whether or not a speaker voice is detected, and the speaker direction detecting means detects a speaker voice. Sometimes, speech recognition of the detected speaker voice is performed, and when a word recognized by the speech recognition matches a predetermined keyword, a voice recognition means for outputting a predetermined signal, and the speaker direction detection means When the conversation detecting means detects that the speaker voices of a plurality of speakers have been detected within a predetermined time in the past, the voices of the speakers spoken within the predetermined time when the speaker voices of the plurality of speakers are detected are detected. A speech recognition apparatus comprising: a recognition control unit that suppresses speech recognition by the speech recognition unit.   The voice recognition device further detects the voice of the speaker by the voice recognition means when the conversation detection means detects that there is a plurality of speaker voices within the predetermined time in the past. 4. The voice recognition according to claim 3, further comprising a guidance output means for generating a signal notifying that voice recognition has not been performed due to the presence of speaker voices of a plurality of speakers when not performed. apparatus.

Images