JP2011081322A - Voice recognition system and voice recognition method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To detect utterance by discriminating noise without presuming a sound source position. <P>SOLUTION: A voice recognition system includes a plurality of directive microphones and a voice recognition section for performing voice recognition on a signal from at least one microphone in the plurality of directive microphones. The indirective microphones and an utterance detection section for detecting an utterance section by the signal from the indirective microphone are provided, and voice recognition is performed on the signal in the utterance section by the voice recognition section. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to speech recognition including a plurality of microphones, and more particularly to speech detection.

  Patent Document 1 (Japanese Unexamined Patent Application Publication No. 2009-210956A) discloses a speech recognition apparatus using a microphone array. In Patent Document 1, the position of a sound source is estimated from the difference in the arrival time of sound to individual microphones, and separately from this, the position of a speaker is stored in advance. Then, an acoustic signal having the speaker's position as a sound source is set as a speech recognition target. However, this method requires constant estimation of the sound source position, and the signal processing becomes heavy.

JP2009-210956A

  An object of the present invention is to detect an utterance by distinguishing it from noise without estimating a sound source position.

  The present invention is a speech recognition system comprising a plurality of directional microphones and a speech recognition unit that performs speech recognition on a signal from at least one microphone among the plurality of directional microphones, An omnidirectional microphone; and an utterance detection unit that detects an utterance period based on a signal from the omnidirectional microphone, wherein the voice recognition unit is configured to perform voice recognition on a signal in the utterance period. It is characterized by being.

  Further, the present invention is a method for performing speech recognition by a speech recognition device on a signal from at least one microphone among a plurality of directional microphones, wherein the speech recognition device is a non-directional microphone. The speech section is detected from the signal of, and speech recognition is performed on the signal of the speech section.

  In this specification, the description related to the speech recognition apparatus is also applied to the speech recognition method as it is, and the description related to the speech recognition method is applied to the speech recognition apparatus as it is. As the directional microphone and the omnidirectional microphone, for example, a plurality of omnidirectional microphones may be provided, and a plurality of directional microphones may be realized by combining these. Alternatively, for example, one omnidirectional microphone and a plurality of directional microphones may be provided separately.

  When there is a noise source in the directivity direction of a directional microphone, weak noise is regarded as a large signal and it is easy to mistake the noise as an utterance. In addition, when the noise source moves or when the speaker moves, it is difficult to arrange the directional microphone so as to avoid the noise source. On the other hand, when an utterance period is detected with a non-directional microphone signal, fluctuations in noise intensity are reduced and the noise level approaches a constant level, making it easy to distinguish between noise and speech. Therefore, it is possible to detect an utterance section without processing such as estimating the position of a sound source or arranging a directional microphone so as to avoid a noise source.

  Preferably, a headset having the plurality of directional microphones and the omnidirectional microphone is provided. Since the posture or position of the worker wearing the headset is not fixed, it is particularly problematic that the directional microphone faces the noise source. Even in such a case, if an omnidirectional microphone is used, speech can be detected separately from noise.

  Preferably, the voice recognition unit is configured to determine the strength of the signal in the utterance section and other sections, for example, the S / N ratio, and the frequency band of the signal in the utterance section with respect to the plurality of directional microphones. And a selection unit that determines which directional microphone to recognize a signal based on at least one of. These signals represent whether individual directional microphones are picking up speech or noise. Therefore, it recognizes a signal from a directional microphone that is likely to pick up the voice. Further, there is no need to recognize a signal from a directional microphone that has a low possibility of recognizing the sound.

  Preferably, the utterance detection unit detects an utterance by comparing a signal from an omnidirectional microphone with at least a threshold value, and determines the utterance according to the strength of the signal from the omnidirectional microphone outside the utterance section. The threshold value is learned, and the threshold value is changed according to the position of the voice recognition device. By learning the threshold according to the noise level, it is possible to detect the utterance more reliably. When each microphone of the voice recognition device is provided in a headset, for example, if an operator wearing the headset moves between a quiet environment and a noisy environment, learning of the noise level cannot follow. Moving to a noisy environment tends to misidentify noise as an utterance, and moving to a quiet environment tends to overlook utterances because people tend to make their voices quieter. Therefore, by forcibly changing the threshold according to the position, it is possible to compensate for the fact that learning is not in time.

The figure which shows the external appearance of the speech recognition apparatus of an Example. Circuit diagram of the filter used to configure the directional microphone in the example Block diagram of speech recognition apparatus of embodiment The top view which shows arrangement | positioning of the directional microphone in a modification, and an omnidirectional microphone Block diagram of principal parts of a modified speech recognition apparatus The figure which shows the detection model of the utterance area in an Example Block diagram of the selector in the embodiment Flowchart showing detection of utterance interval in the embodiment Flowchart showing selection of directional microphone in embodiment

  In the following, an optimum embodiment for carrying out the present invention will be shown. The scope of the present invention should be determined in accordance with the understanding of those skilled in the art based on the description of the scope of claims, taking into consideration the specification and well-known techniques in this field.

  1 to 9 show a speech recognition apparatus 2 according to the embodiment and a speech recognition method according to the embodiment. In FIG. 1, 4 is a speaker, 6 is a microphone array, and includes a plurality of omnidirectional microphones 7-10. Reference numeral 12 denotes an arm, 13 denotes a cord, and 14 denotes a voice recognition apparatus body, which includes a power supply 15, a signal processing unit 16, and a communication unit 18. In the embodiment, the voice recognition device 2 performs voice recognition alone, communicates with a server (not shown) via the communication unit 18, and transmits a picking command from the server to the worker via the speaker 4, for example. The voice is recognized by the microphones 7 to 10 and the signal processing unit 16 and notified from the communication unit 18 to the server. The server also monitors the position of the worker using a camera, GPS, etc., and changes the threshold for speech detection. Note that voice recognition may be performed by a server, and the signal processing unit 16 may be removed from the voice recognition device 2 side.

  The voice recognition apparatus 2 of the embodiment recognizes voices of not only picking workers but also pilots such as airplanes, automobile drivers, doctors and dentists during surgery, factory workers, and call center operators. Suitable for The voice recognition device 2 of the embodiment is composed of a headset, and when the operator shakes his / her head, the direction of the microphone array 6 changes, so the direction to noise constantly changes. Further, when the worker moves, the position with respect to the surrounding noise source changes.

The microphone array 6 is shown enlarged on the left side of FIG. The microphone array 6 includes, for example, four microphones 7 to 10 having no directivity. The microphones 7 to 10 are arranged at the apexes of a regular tetrahedron, for example, and the microphone 7 protrudes upward from the microphones 8 to 10. And For example, one of the four microphones 7 to 10 is used as it is as an omnidirectional microphone. When two microphones are combined from the four microphones 7 to 10, six combinations of ₄ C ₂ are generated.

  Six virtual directional microphones are realized by six combinations. For example, a combination of the microphone 7 and the three microphones 8 to 10 provides three directional microphones. A combination of the microphone 8 and the microphone 9 provides a microphone directed rightward and a microphone directed leftward. Similarly, a total of, for example, nine (6 +3) directional microphones are obtained.

  FIG. 2 shows a directional microphone in which microphones 7 and 8 are combined. Here, the microphones 7 and 8 are taken as an example, but the same applies to other combinations of microphones. Reference numerals 20 and 21 schematically show sensitivity distributions (directivity) of the microphones 7 and 8, and the sensitivity decreases in a direction that is a shadow of the counterpart microphone. An amplifier 22 amplifies the audio signal from the microphones 7 and 8. Reference numeral 25 denotes a delay unit that amplifies the signal and delays the signal by the time required for the sound wave to travel the distance between the microphones 7 and 8. Since the microphones 7 to 10 are arranged at the apexes of the regular tetrahedron, the interval between the microphones is constant, and the time delayed by the delay unit 25 is constant regardless of the combination of the microphones. Therefore, one delay unit 25 is provided for each microphone 7-10.

  A difference unit 26 obtains a difference between the signal from the delay unit 25 and the undelayed signal from the microphone of the other party of the combination. For example, the subtractor 26a subtracts the signal of the microphone 8 from the signal of the microphone 7 to obtain a directional signal on the lower side of FIG. When a sound wave arrives at the microphone 7 in FIG. 2 from above, it is input to the difference unit 26a via the delay unit 25 and canceled by the signal from the microphone 8, so the output from the difference unit 26a becomes small. When a sound wave arrives at the microphone 8 from the lower side, it is immediately input to the differentiator 26a. Since the signal from the microphone 7 is delayed by the delay unit 25, the signal is not canceled out. At this time, the sound signal of one time is inverted in sign and output from the differentiator 26a twice with a slight time difference to become a kind of repetitive signal. However, the sound model 36 in FIG. The effect is small. If necessary, a filter for removing the repetitive signal may be provided between the delay unit 25 and the differentiators 26a and 26b. Similarly, in the case of the differentiator 26b in FIG. 2, the microphone 7 signal is subtracted from the microphone 8 signal to obtain a microphone directed upward in FIG.

  FIG. 3 shows the overall configuration of the speech recognition apparatus 2. An amplifier 22 is connected to each of the microphones 7 to 10, and an arbitrary one of the microphones 7 to 10 having no directivity, for example, a signal from the microphone 7 is uttered. Processing is performed by the detection unit 30 to detect an utterance section. The learning unit 32 changes the threshold for speech detection based on the signal from the amplifier 22 connected to the microphone 7 outside the speech detection section. In addition to this, the threshold value is changed depending on the position of the voice recognition device 2, and the position signal is generated by providing the voice recognition device 2 with GPS, for example. Alternatively, the position signal may be input from the server by recognizing the position of the voice recognition device 2 with a server (not shown).

  The threshold for speech detection changes, for example, in the order of seconds to hours. For example, for a section other than the speech section, for example, every predetermined time such as 1 second to 1 minute, the current threshold and the past 1 second to about 1 minute. The weighted average with the noise level at is used as a new threshold. For the weight, for example, the current threshold is set to about 99% to 80%, and the ambient noise level is set to about 1% to 20%. In this way, the threshold value is learned (changed) at a predetermined speed based on the time interval for changing the threshold value and the weight for the surrounding noise level. On the other hand, the threshold value is changed by the position signal because, for example, the picking operator passes through the door from a substantially silent section such as a refrigerator compartment or a freezer compartment, and the section where the automatic guided vehicle travels is noisy. Such as when moving to. The threshold value is changed with priority over learning by passing through the door.

  As shown in FIG. 2, the filter 24 generates nine directional microphone signals, the selector 34 selects the two directional microphone signals, and the selected two microphone signals are added to the adder 35. Are added and converted to phonemes by the acoustic model 36. Then, the phoneme is converted into a language by the language model 38, and the speech recognition is completed. The selector 34, the acoustic model 36, and the language model 38 constitute a speech recognition unit.

  In the embodiment, a headset is used as an example. However, as shown in FIG. 4, a directional microphone 41 and an omnidirectional microphone 42 are arranged on a fixed table 40, and an utterance period is determined by a signal from the omnidirectional microphone 42. May be detected. For example, it is assumed that the noise source 44 passes around the table 40. When an utterance section is detected from the signal of the directional microphone 41, the passage of the noise source 44 is detected by the specific directional microphone 41 and is easily misidentified as the utterance section. On the other hand, in the omnidirectional microphone 42, even if the position of the noise source 44 changes, the change in the noise level is small and it is difficult to be influenced by the specific noise source 44. it can.

  FIG. 5 shows a modification of the selector, in which sound signals from nine virtual microphones generated by the filter 24 are converted into phonemes by the acoustic model 36, respectively. The acoustic model 36 may generate likelihood at the time of conversion to a phoneme. For example, the top two signals having the highest likelihood may be selected by the selector 50, and the top two phoneme signals may be input to the language model 38. However, in this case, since the signal not processed by the language model 38 is also processed by the acoustic model 36, the processing becomes heavy.

  5 to 9 show the operation of the embodiment. For example, an operator who performs picking wears the voice recognition device 2 as a headset, and performs picking while communicating with a server (not shown) through the communication unit 18. It is assumed that the server recognizes the position of the worker using, for example, a camera or GPS (not shown). The server instructs the worker to work from the speaker 4, and input from the worker to the server is performed by the voice recognition device 2. When the orientation of the headset is changed, the orientation with respect to the noise source is changed. The degree of noise changes.

  Here, four omnidirectional microphones 7 to 10 are combined to virtually constitute nine directional microphones, and any one of the four microphones 7 to 10 is used for speech detection. . In the detection of utterances, as shown in FIG. 6, thresholds are set on, for example, both sides of the 0 level of the signal, and the presence or absence of utterances is detected from the number of times the signal has passed the 0 level after exceeding the threshold. This leads to the evaluation of signal strength and frequency.

  Since the noise level changes depending on the position of the worker and the like, the threshold value is changed by learning based on a signal from the microphone in a section where speech is not detected. The noise level changes when you enter another room through the door. Since learning generally does not follow a change in noise level due to a change in position, the threshold is forcibly changed based on position information from the server.

  When an utterance is detected, speech recognition is performed using signals from the two microphones with the highest likelihood among, for example, nine directional microphones. A configuration for this is shown in FIG. 7, and 76 is one of the virtual directional microphones. The average value of the signal level outside the speech section is stored in the averaging unit 71, and the divider 70 is used. The ratio between the signal from the microphone 76 and the average value is obtained as the S / N ratio. Further, the counter 72 obtains the number of 0-crossings in the utterance section. Furthermore, the frequency spectrum corresponding to the audio signal is obtained by obtaining the strength of the signal in the frequency band corresponding to the audio signal, preferably obtaining the strength of the signal at a plurality of frequencies by a band pass filter or other simple frequency conversion unit. Ask whether or not.

  The S / N ratio represents the degree to which speech is recognized against noise, the number of 0-crossings represents the average frequency of the detected signal, and the signal of the frequency converter 73 is also the input signal. Represents the spectral shape. By evaluating these signals by the evaluation unit 74, the likelihood of recognizing speech is output as likelihood. Since both the signal of the counter 72 and the signal of the frequency conversion unit 73 relate to the frequency of the signal, only one of them may be used. Then, by processing the signals from the microphones with the two most likely directivities with the acoustic model 36, speech recognition can be performed accurately. Instead of setting the top two likelihoods, only the signal with the highest likelihood or only the signal with the top three likelihoods may be used.

  The processing of the embodiment is shown in FIGS. 8 and 9, and the contents thereof are as described above. In the detection of the speech section in FIG. 8, in steps 1 and 2, the signal of the omnidirectional microphone is zero-crossed from the outside of the threshold value. From the number of times, it is identified whether it is a speech segment or a non-speech segment. In the non-speech section, as in steps 3 and 4, the threshold for speech detection is changed by learning, and after passing through the door and entering another room, the threshold is corrected according to the position information (step 5).

  Speech recognition is performed on the speech section, and in steps 10 and 11, the directional microphones with the highest likelihood are determined by the S / N ratio of the signal from each directional microphone, the number of zero crossings, the frequency distribution, and the like. And the voice is recognized in step 12.

  In the embodiment, it is possible to detect an utterance by distinguishing between noise and voice without estimating the position of the sound source. In addition, it is possible to determine which directional microphone signal is recognized as a voice without estimating the sound source position. Furthermore, when the voice recognition device 2 is provided in the headset, the speech section is detected without being affected by the change of the microphone direction with respect to the noise source or the change of the relative position with respect to the noise source due to the movement of the operator. it can. By changing the threshold for utterance detection by learning, it is possible to set a threshold according to the strength of noise at each time point, and by changing the threshold by position information, without waiting for learning when moving to a new environment The threshold can be changed.

2 Speech recognition device 4 Speaker 6 Microphone array 7 to 10 Microphone 12 Arm 13 Code 14 Speech recognition device body 15 Power supply 16 Signal processing unit 18 Communication unit 20, 21 Sensitivity distribution 22 Amplifier 24 Filter 25 Delay unit 26 Differentiator 30 Speech detection unit 32 learning unit 34 selector 35 adder 36 acoustic model 38 language model 40 table 41 directional microphone 42 omnidirectional microphone 44 noise source 50 selector 70 divider 71 averaging unit 72 counter 73 frequency conversion unit 74 evaluation unit 76 directivity Sex microphone

Claims (5)

A system comprising a plurality of directional microphones and a voice recognition unit that performs voice recognition on a signal from at least one of the plurality of directional microphones,
An omnidirectional microphone; and an utterance detection unit that detects an utterance period based on a signal from the omnidirectional microphone, wherein the voice recognition unit is configured to perform voice recognition on a signal in the utterance period. A voice recognition system characterized by The speech recognition system according to claim 1, comprising a headset including the plurality of directional microphones and the omnidirectional microphone. The voice recognition unit may determine which directivity to the plurality of directional microphones based on at least one of a signal strength level in an utterance section and other sections and a signal frequency band in the utterance section. The speech recognition system according to claim 1, further comprising: a selection unit that determines whether to recognize a signal from a sexual microphone. The utterance detection unit detects an utterance by comparing at least a signal from an omnidirectional microphone with a threshold, and learns the threshold according to the strength of the signal from the omnidirectional microphone outside the utterance section. The voice recognition system according to any one of claims 1 to 3, wherein the threshold value is changed in accordance with the position of the voice recognition device. A method of performing speech recognition from a speech recognition device on a signal from at least one microphone among a plurality of directional microphones,
A speech recognition method, wherein the speech recognition device detects an utterance section from a signal from an omnidirectional microphone and performs speech recognition on a signal in the utterance section.

Images