JP2007264132A - Voice detection device and its method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a voice detection device for surely detecting the contents of utterance. <P>SOLUTION: When outputs of both an air conductive microphone 1 and a bone conductive microphone 3 exceed a predetermined level for each unit time, the correlation value is calculated. When it is determined that correlation is high, time information is sent to a voice section detection section 7, and thereby, a section which is surely assumed to be a voice section, can be utilized at the time of voice detection. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to an audio detection apparatus and method using a plurality of microphones.

  As a method for improving the interval detection accuracy for speech recognition by using a bone conduction microphone and a normal air conduction sound microphone together, the technique of Patent Document 1 has been proposed. This method performs a separate voice segment detection process for each microphone output, decides which detection result is given priority according to the noise level, and finalizes one of the voice segment detection results based on that priority. It is a method of obtaining a result.

Patent Document 2 also discloses a technique for obtaining the degree of correlation between a sound signal received by a unidirectional microphone and a sound signal received by an omnidirectional microphone and detecting a detection target sound based on the correlation result. Proposed.
JP-A-4-276799 JP 2005-227511 A

  In the technique of Patent Document 1, when the background noise level is low, priority is given to the sound section detection result of the air conduction microphone. However, even when the background noise level is low, for example, when non-stationary noise such as the neighbor's speech is added before and after the target speech, it is erroneous that the section including the added section is detected as the speech section. There is a case where a voice section is detected. In such a case, there is a problem that the error cannot be corrected, and the voice signal in the wrong voice section is input to the voice recognition device, resulting in a recognition error.

  In addition, when all the voice section detection processes for the voice signals obtained from the bone conduction microphone and the air conduction sound microphone are all finished, it is determined for the first time which voice section is provided to the voice recognition device. Therefore, there is a problem that real-time processing of voice recognition cannot be performed.

  Furthermore, in the technique of Patent Document 2, since an utterance is received by an omnidirectional microphone, for example, a neighbor's speaking voice is detected as a detection target sound and input to the voice recognition processing, and accurate recognition is performed. There is a problem that cannot be performed.

  Therefore, the present invention provides a voice detection apparatus and method for detecting only a target sound that should be reliably detected.

  The present invention provides a correlation between a plurality of microphones having different characteristics, correlation value calculation means for obtaining a correlation value between input signals respectively input from the microphones, and correlation between the input signals based on the correlation values. A determination means for determining whether or not there is a correlation and outputting time information when it is determined that there is a correlation; and one input signal among the input signals in a section determined to have a correlation based on the time information A voice section detecting means for outputting a voice signal as a voice signal.

  According to the present invention, the content of the utterance can be reliably detected.

  A voice detection device 100 according to an embodiment of the present invention will be described with reference to FIGS.

(1) Configuration of Voice Detection Device 100 FIG. 1 is a block diagram showing the configuration of the voice detection device 100 of the present embodiment.

  The voice detection device 100 will be described in the case of using two microphones, an air conduction microphone 1 and a bone conduction microphone 2. A person who wants to recognize voice (hereinafter referred to as a microphone wearer) wears the bone conduction microphone 2 near the body, particularly in the vicinity of the face, and the air conduction sound microphone 1 can receive the voice of the microphone wearer. Suppose that

  The voice detection device 100 includes an air conduction sound microphone 1, a level determination unit 2, a bone conduction microphone 3, a level determination unit 4, a correlation value calculation unit 5, a determination unit 6, a voice section detection unit 7, a noise feature estimation unit 8, a noise And a cancel unit 9. In this embodiment, it is assumed that a voice recognition device is connected to the voice detection device 100 and that the output of the voice detection device 100 is used for voice recognition processing.

(2) Air conduction sound microphone 1 and level determination unit 2
An acoustic signal transmitted in the air is input to the air conduction microphone 1, this acoustic signal is converted into an electric signal, further AD-converted, and then converted into a digital signal (hereinafter referred to as an input acoustic signal). Sent to.

  In the level determination unit 2, for example, the power is calculated by the square sum of the amplitude of the waveform of the input acoustic signal every 10 milliseconds. The power of the input acoustic signal is compared with a preset power threshold, and if the threshold is exceeded, the input acoustic signal for 10 milliseconds is input to the correlation value calculation unit 5.

(3) Bone conduction microphone 3 and level determination unit 4
On the other hand, in synchronization with time, vibration sound transmitted through the body is also input to the bone conduction microphone 3, and this vibration sound is converted into an electric signal, and further converted into a digital signal (hereinafter referred to as an input vibration signal) after AD conversion. It is converted and sent to the level determination unit 4.

  In the level determination unit 4, for example, the length of 10 ms of the input vibration signal is defined as one frame, and the power is calculated by the square sum of the amplitude of the waveform of the input vibration signal for each frame. The power of the input vibration signal is compared with a preset power threshold, and if the threshold is exceeded, the input vibration signal for 10 milliseconds is input to the correlation value calculation unit 5.

  Note that the threshold values in the level determination units 2 and 4 are set to different values because the levels of the input acoustic signal and the input vibration signal are different.

(4) Correlation value calculation unit 5 and determination unit 6
In the correlation value calculation unit 5, the input acoustic signal from the air conduction microphone 1 and the input vibration signal from the bone conduction microphone 3 that have passed the level determination process are input. A correlation value is calculated when both signals are input.

  For correlation, for example, the Euclidean distance between frequency patterns obtained by frequency-resolving each signal may be used, or a correlation coefficient may be obtained.

  Here, in the case of the voice actually uttered by the microphone wearer, the input acoustic signal obtained from the level determination unit 2 is the voice transmitted in the air, and the input vibration signal obtained from the level determination unit 4 depends on the voice that is uttered. The signal is a vibration of the bone. Since both have the same signal source, each correlation value is considered to indicate a high value.

  As shown in FIG. 2, a case where a person who is not a microphone is uttered is considered. The input acoustic signal obtained from the level determination unit 2 is the sound that travels in the air. On the other hand, since there is no input to the bone conduction microphone, it is not output from the level determination unit 4.

  As shown in FIG. 3, a case is considered where an input vibration signal is input only to the bone conduction microphone 33 due to body friction or the like, without the microphone wearer speaking. There is no input to the air-conduction sound microphone 1, or even if it exists, it becomes a very low level signal, and there is no output from the level determination unit 2.

  As shown in FIG. 4, a case is considered in which both the voice of a person who is not a microphone wearer and the vibration generated from the body of the microphone wearer overlap in time. The level determination units 2 and 4 respectively output an input acoustic signal of human voice and an input vibration signal. Since these signal sources are different, the correlation value is considered to be very small.

  When the correlation value calculated by the correlation value calculation unit 5 exceeds a threshold set in advance by the determination unit 6, it is determined that the signals input to both microphones 1 and 3 have a correlation, and the time at that time Information is output to the speech section detection unit 7 and the noise feature estimation unit 8.

  The above processing describes a phenomenon that can be considered when there is no background noise. In any case, it is possible to prevent a signal other than the voice of the microphone wearer from being input to the voice section detection unit 7, and the time information of the section considered to be uttered by the microphone wearer is reliably detected. Part 7 can be notified. In other words, it can be considered that the microphone wearer is speaking at the time.

  However, among the voices uttered by the person wearing the microphone, especially the consonant at the beginning of the word hardly oscillates in the bone and the like, and it is difficult to pick up with the bone conduction microphone 3, so that the correlation value is actually high. Audio may have started at a time earlier than the previous time. Moreover, these phenomena may occur not only at the beginning of a word but also at the end of a word or at the end of a word. These problems are dealt with by the speech recognition device connected to the speech section detection unit 7 and the speech detection device 100.

  On the other hand, considering the case where there is background noise, the input acoustic signal that has passed through the air conduction microphone 1 is assumed to be a high-level signal due to the influence of noise even when the microphone wearer does not utter. Then, it passes through the level determination unit 2 and is sent to the correlation value calculation unit 5. However, since the bone conduction microphone 3 is not affected by the background noise, there is no output from the level determination unit 4 in the section of only the background noise where the microphone wearer does not utter. Therefore, the influence of background noise can be eliminated. In this way, whether or not there is background noise, the level determination process and the determination process based on the correlation value are performed without being affected by the noise, so that the information of the time when the microphone wearer is surely uttered is voiced. It can be output to the section detection unit 7.

(5) Voice section detection unit 7
Although the process in the audio | voice area detection part 7 is not specifically limited, The example is shown based on FIG.

  In addition to the time information assumed to be the above-described speech section, an input acoustic signal output from the air conduction microphone 1 is input to the speech section detection unit 7 via the noise feature estimation unit 8.

  Further, the signal after the noise cancellation processing by the noise cancellation unit 1 is also input via the noise feature estimation unit 8.

  The voice section detector 7 detects the correct voice section of the microphone wearer using these three types of information and signals.

  As described above, since the time information of the speech section output from the determination unit 6 may not include a section such as a consonant at the beginning of the word, for example, a time point 30 frames (= 300 milliseconds) before the time The voice section detection process is started from the beginning. Therefore, a storage area for storing, for example, 30 frames of the input acoustic signal is built in the voice section detection unit 7. The voice section detection unit 7 uses both the input acoustic signal before noise cancellation and the input acoustic signal after noise cancellation for each start frame. For example, the frame is uttered by the microphone wearer based on the power and frequency pattern for each frame. It is determined whether or not it corresponds to the selected voice.

  Frames determined to correspond to speech are sequentially sent from the speech start frame to the speech end frame to the speech recognition device, and feature extraction and speech recognition processing for speech recognition are executed.

(6) Noise feature estimation unit 8
The noise feature estimation unit 8 uses the signal of the frame for noise feature estimation based on the determination that the frame is a noise interval until the determination unit 6 determines that the frame is a speech interval.

  Specifically, when a plurality of continuous noise frames continue, the noise feature signal N (n−1) before the frame is changed by a certain update coefficient α using the frequency feature signal F (n) of the frame. By updating, the noise feature signal N (n) up to the frame is obtained. If the next frame is determined to be a noise frame, the same update is performed. In other words, the update formula is as follows.

N (n) = (1-α) · N (n−1) + α · F (n)

This noise feature signal is used by the noise canceling unit 9.

(7) Noise canceling unit 9
In the noise canceling unit 9, based on the estimated noise feature signal sent from the noise feature estimating unit 8 and determination information on whether the frame is noise or speech, if the frame is a speech frame, the air conduction microphone 1 To extract only the audio component from the input signal output from the.

  For example, the spectral subtraction method or the like is used to perform processing such as subtracting the noise feature signal from the frequency feature signal of the frame, canceling the noise component, and extracting only the speech component.

  This voice component signal is used for voice section detection, and is also sent to the voice recognition device via the voice section detector 7 and used for voice recognition processing.

(8) Effects The voice detection device 100 according to the present embodiment can accurately detect a voice section without being affected by stationary noise such as ambient background noise and non-stationary noise such as human speech, telephone sound, and horn. As a result, the speech recognition performance can be improved.

(9) Modification Examples The present invention is not limited to the above-described embodiments, and various modifications can be made without departing from the gist thereof.

  For example, in the above embodiment, the case where two microphones, the air conduction sound microphone 1 and the bone conduction microphone 2 are used, has been described. Instead of this, a non-directional air conduction sound microphone and a directional air conduction microphone are used. Three microphones, a sound microphone and a bone conduction microphone, may be used. In this case, whether the voice or noise is determined is determined to be voice when there is a correlation among at least a majority of combinations among a plurality of input signals. Further, when there is no correlation in the majority of combinations, it is determined that the noise is present.

It is a block diagram of the audio | voice detection apparatus of one Embodiment of this invention. It is a figure which shows the example when audio | voices other than the utterance of a microphone wearer are input into the air conduction sound microphone. It is a figure which shows the example at the time of vibrations other than the voice of a microphone wearer being input into the bone conduction microphone. It is a figure which shows the example when the sound and vibration other than a microphone wearer's utterance are simultaneously input into both an air conduction sound microphone and a bone conduction microphone. It is a figure which shows an example of an audio | voice area detection process.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Air conduction sound microphone 2 Level judgment part 3 Bone conduction microphone 4 Level judgment part 5 Correlation value calculation part 6 Judgment part 7 Voice area detection part 8 Noise feature estimation part 9 Noise cancellation part

Claims (6)

Multiple microphones with different characteristics,
Correlation value calculation means for obtaining a correlation value between input signals respectively input from the microphones;
A determination means for determining whether or not there is a correlation between the input signals based on the correlation value, and for outputting time information when it is determined that there is a correlation;
Voice section detecting means for outputting one input signal among the input signals in the section determined to be correlated based on the time information as a voice signal;
A voice detecting device characterized by comprising: The speech detection device according to claim 1, wherein at least two of the plurality of microphones are an air conduction microphone and a bone conduction microphone. The voice detection device according to claim 2, wherein the input signal output as a voice signal in the voice section means is an input signal of the air conduction microphone. The voice detection device according to claim 3, further comprising noise canceling means for removing noise from the input signal of the air conduction microphone and outputting the removed input signal to the voice section detection means. Noise feature estimation means for estimating a noise feature signal from an input signal of the air conduction microphone in a section determined to have no correlation in the determination means;
The voice detection device according to claim 4, wherein the noise cancellation unit removes noise from an input signal of the air conduction microphone based on the noise characteristic signal. Find the correlation value between the input signals respectively input from multiple microphones with different characteristics,
It is determined whether there is a correlation between the input signals by the correlation value, and outputs time information when it is determined that there is a correlation,
A voice detection method, wherein one input signal among the input signals in a section that is correlated based on the time information is output as a voice signal.

Images