JP2009229583A - Signal detection method and device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To improve detection accuracy of a signal by automatically determining a voice activity detection (VAD) threshold. <P>SOLUTION: A calculation means calculates an intensity signal that indicates the intensity of a signal. A filter means filters the intensity signal. A tracking means tracks the noise level of the intensity signal. A discrimination means discriminates the signal, by using the filtered intensity signal and the noise level which is output by the tracking means. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a signal detection method and apparatus.

  VAD (Voice Activity Detection) is a technique for detecting sound in an environment with background noise.

  There are two typical scenes where VAD is required.

  The first is a voice communication device, which is shown in FIG. In this case, the audio signal input from the audio input unit 10 is encoded by the encoder 13 and transmitted to another location including the decoder 16. It is assumed that the communication channel 15 is restricted from the viewpoint of cost and capacity to be used. The VAD 12 allows the device to transmit only when the user is speaking. If the user is not speaking, the VAD 12 can prevent transmission, thus saving money and allowing another user on this network to transmit. Although it is convenient for the VAD 12 and the encoder 13 to have the same front-end processing 11, it is not necessary. Also, since the encoder 13 does not require all words, the encoder 13 does not need to distinguish between pauses in speech utterance and long silence.

  The second is an automatic speech recognition (ASR) system, which is shown in FIG. The voice input unit 20 is the same as the voice input unit 10 of FIG. 1, but in this case, the VAD 22 prevents the voice recognition unit 24 from trying to determine background noise as voice. This not only helps prevent errors (discriminating background noise from speech) but also helps manage system resources (ASR is usually computationally expensive). An example is a mobile device that can be controlled by voice. VAD distinguishes between when the user is not speaking and when the user is speaking. Thereby, when the ASR is not performed, the apparatus can be concentrated on other functions, and when the user is speaking, the apparatus can be concentrated on the ASR. As described above, it is convenient that the front end processing 21 is the same for the VAD 22 and the voice recognition unit 24. In this example, the end point detection unit 23 uses the VAD signal to distinguish between the start or end of the utterance and the pause period between the words being uttered. This is because the voice recognition unit 24 needs to receive voices corresponding to all words without interruption.

  The present invention can also be applied to communication, but in the description of embodiments of the present invention described later, a case where the present invention is applied to ASR will be described. In this specification, the terms “noise”, “silence”, and “non-speech” are interchangeable.

  A VAD generally comprises the following components as shown in FIG.

  1. VAD calculation unit 38. A VAD scale value 39 for the audio signal (instantaneous signal) 31 is formed. Examples of the scale include spectral expressions such as a zero crossing rate, amplitude, power (energy), Fourier transform, and periodogram of the signal. The scale may also be based on front end characteristics. As shown in FIG. 3, the front end unit 42 includes an amplitude framer 32 that generates an amplitude signal 33. A VAD scale value is calculated from the amplitude signal 33 as the intensity signal.

  2. Noise tracking unit 36. Track the steady state of the signal representing background noise. This is usually the average value of the signal measure over past observations.

  3. The VAD calculation unit 38 further compares the instantaneous signal scale with the steady state signal scale to generate a VAD scale value 39. The VAD scale value may take a continuously changing value, but may be threshold-determined (40) so as to obtain a logical operation value VAD discriminant value 41. The assumption here is that if the instantaneous scale is significantly different from the steady state scale, it is likely that the speech is being measured rather than just background noise.

  The ASR system VAD is usually combined with the following configuration.

  4). An end point detection unit 44. An endpoint discrimination value 45 is obtained from the VAD scale value. For example, the VAD scale value is set to take a larger value for voice than for non-voice. When the VAD scale value exceeds a certain value for a certain period, it can be determined that there is a voice. Conversely, if the VAD scale value is equal to or less than a certain value for a certain period, it is determined as non-speech.

  In addition to the above, it is common to feed back the result of the VAD discriminant value 41 or the end point discriminant value 45 to the noise tracking unit 36. The purpose is to indicate to the noise tracker whether the current signal is speech or non-speech. As a result, the noise tracking unit can update the noise estimation more reliably.

  Conventionally, methods for performing background noise tracking and automatic threshold determination have been disclosed. For example, the inventor's patent application discloses an automatic threshold determination method based on background noise and probability distribution. However, using a statistical distribution would be inappropriate for VAD. The noise tracking method of our previous application is based on Non-Patent Document 1.

  In Patent Document 1, a fixed threshold value VAD is used to control automatic gain control that adjusts the signal level, not the VAD threshold value. This complements the approach used in the present invention.

  In Patent Document 2 and Patent Document 3, the VAD threshold is automatically adjusted, but basically, this is performed considerably after the processing steps. In the present invention, the threshold value is adjusted using the information before the VAD processing, and therefore this adjustment is immediately performed.

  In Patent Document 4, the threshold value is automatically adjusted based on the signal-to-noise ratio (SNR) calculated for one frame. This technique is different in the present invention in that the moving average noise level is used instead of the noise level of one frame, and the threshold is an actual noise level instead of SNR.

Sohn and Sung, "A Voice Activity Detector using soft decision based noise spectrum adaptation", In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing, pages 365-368, May 1998 US Patent Application No. 2002/0118851 US Patent Application No. 2003/0009333 US Pat. No. 6,539,352 US Pat. No. 6,640,208 US Pat. No. 6,711,536

  Patent Document 5 describes an implementation example of VAD. In this VAD, a modulation signal is calculated for a frame sequence using the frame size. Next, the VAD feature is calculated by filtering the modulated signal with one discrete Fourier transform (DFT) bin. The DFT is arranged so that the filtered signal has an energy of 2-6 Hz (ideally about 4 Hz). To determine if the input signal is speech or noise, the filtered signal is compared to a threshold value. If the filtered signal exceeds the threshold for a predetermined period, it is determined that the sound is voice. Otherwise, it is determined that the signal is noise. This VAD has a problem of obtaining a threshold value in advance. Since the threshold must be above the noise level and below the audio level, making such a pre-determination requires knowledge of the signal levels of both background noise and audio input. This problem becomes significant in the following two cases.

  1. If the threshold is set in a quiet background noise environment, if the device is used in a noisy environment, the filtered background noise will exceed the threshold and the VAD will incorrectly classify the noise as speech.

  2. If the user speaks with a very low voice or speaks far away from the microphone, the filtered audio signal may not reach a predetermined threshold. In this case, a valid utterance is not detected.

  It is an object of the present invention to provide a novel method that automatically determines the VAD threshold and minimizes the occurrence of the above-mentioned problems.

  According to an aspect of the present invention, a calculation step for calculating an intensity signal indicating the intensity of a signal divided into frames, a filtering step for filtering the intensity signal, and a tracking step for tracking the noise level of the intensity signal; There is provided a signal detection method comprising: a discrimination step for discriminating the signal using the intensity signal filtered in the filtering step and the noise level output in the tracking step.

  The present invention has two major advantages over prior art methods. First, compared to our previously disclosed invention, the automatic threshold is more accurate for unknown background noise levels. For example, when the background noise increases, the threshold value is also increased to compensate for the background noise rather than being fixed in the previously disclosed invention.

  Second, the baseline threshold value is calculated so that the noise level is surely below the threshold value and the sound level is surely above the threshold value under normal conditions as compared to other prior art. No further subjective adjustment is necessary. In the prior art method, the threshold value must be set subjectively so as to exceed the noise level.

  DESCRIPTION OF EMBODIMENTS Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the drawings. In addition, this invention is not limited to the following embodiment, It shows only the specific example advantageous for implementation of this invention. In addition, not all combinations of features described in the following embodiments are indispensable as means for solving the problems of the present invention.

  The present invention is based on background noise tracking. During the non-speech period, an average value of the signal level is taken and an estimate of the background noise level is calculated. By considering the two filtered and unfiltered versions of the signal, the algorithm determines a suitable baseline value for the threshold. Although this reference line can be experimentally adjusted with knowledge of the input signal level, such adjustment need not necessarily be performed under normal conditions. Specifically, after the filtering is performed, the background noise averaged before the filtering is performed is used as a threshold for the signal.

  A main feature of the VAD of the present embodiment is an amplitude signal 33 calculated from an overlap frame of audio samples. Using standard PCM sampling, the audio signal 31 is sampled at 11025 Hz and the sample stream is divided into 256 sample frames each. A new frame starts every 110 samples. That is, the frames are overlapped and the frame rate is about 100 Hz. For each frame, the average value of 256 samples is subtracted to remove the DC component, and the magnitude is calculated as the sum of the absolute values of each sample. Amplitude is closely related to frame energy, but does not require multiplication. This amplitude signal is then processed in two separate processing stages as shown in FIG. FIG. 4 shows the configuration of the signal detection apparatus in this embodiment.

The first processing stage is a process of filtering with a filter realized as a 25-point DFT 51. The first non-DC bin of the DFT is calculated using the square root 50 of the current frame size and the square root of the past 24 frames. This is done, for example, using the sliding DFT algorithm described in US Pat. No. 4,723,125. The real part and the imaginary part of the DFT bin are squared and added to obtain one real value D _t ² and a VAD scale value (4 Hz component value) 52.

  The second processing stage is processing by the noise tracking unit 36 realized by a single pole recursive filter expressed by the following equation.

/ D _t = ρ (/ d _t−1 ) + (1−ρ) d _t

However, / d is a tracked noise level 37, t is a frame index, d is an amplitude, and ρ is a time constant set to 0.98. In order to prevent the impact of high energy audio signals on the tracked noise level, the tracking operation needs to be inhibited by VAD. This will be described later. Since the above expression is recursive, / d needs to be initialized. In a preferred embodiment, initialization is performed to be the same as the first measurement value d ₁ .

Since / d is the upper limit of D _t ² for stationary background noise, if D _t ² exceeds / d, there is a high possibility that the signal is speech. In this sense, / d is the preferred VAD threshold 53. In the preferred embodiment, / d is multiplied by a small fixed value e to increase or decrease the sensitivity of the VAD, respectively.

For ease of understanding, we have described without using logarithms so far, but in the preferred embodiment, in practice, / d and D _t ² are converted to logarithm before being applied to threshold determination. Therefore, the coefficient e is not multiplication but addition. In realization by fixed-point arithmetic, numerical stability increases by such logarithmization. It will be apparent to those skilled in the art that the difference between thresholding before and after the logarithmic operation is of little concern. Next, when the following equation is satisfied by the final VAD discriminating value 41, the input frame at time t is discriminated as speech, and otherwise it is discriminated as silence.

logD _t ² > log (/ d _t ) + e

  In this way, the signal can be determined using a region offset by a predetermined amount from the noise level output from the noise tracking unit. In logarithmic space, the coefficient e is closely related to the decibel (dB) measurement and can therefore be set very intuitively. It can be set objectively by a known noise environment, or it can be set subjectively by detecting perceptual error and eliminating it as an error.

  The signal detection method described above can be applied to speech recognition. For example, voice and background noise are discriminated by the above-described signal detection method to recognize the voice. For example, when used in an application example of speech recognition, an endpoint detection algorithm is necessary. The end point detection algorithm used in this embodiment will be described.

  FIG. 5 shows a configuration of the end point detection unit 44. The main feature is a state transition determination unit 90, which performs a state transition as shown in FIG. The state starts at “silent state” 80. When the transition 84 to the “voice confirmation waiting state” 81 is performed, the current frame index is stored in the frame index storage unit 91. Next, when the transition 86 to the “voice state” 82 is performed, the frame index stored as the start point of the end point is output. Similarly, when the transition 87 to the “silent determination waiting state” 83 is performed, the current frame index is stored in the frame index storage unit 91. Next, when the transition 89 to the “silent state” 80 is performed, the frame index stored as the end point of the end point is output.

  The transition is caused by the VAD discrimination value 41. First, in the “silent state” 80, when the VAD discriminating value indicates the presence of speech, a transition 84 to the “sound confirmation waiting state” 81 occurs. At this time, the sound count of the sound count storage unit 93 is set to zero. During the “voice confirmation waiting state” 81, the voice count in the voice count storage unit 93 is incremented every time the VAD determination value 41 indicates the presence of voice in each frame. When the voice count value in the voice count storage unit 93 exceeds a certain value (for example, 30), the transition 86 to the “voice state” 82 is performed. On the other hand, when the silence count value in the silence count storage unit 92 exceeds a certain value (for example, 26), the transition 85 to the “silence state” 80 is performed.

  In the “sound state” 82, when the VAD discriminating value indicates non-speech, a transition 87 to the “silent determination waiting state” 83 is performed. At this time, the silence count value in the silence count storage unit 92 is set to zero. During the “silence determination waiting state” 83, the silence count value in the silence count storage unit 92 is incremented every time the VAD determination value 41 indicates silence in each frame. If the VAD discriminant value indicates voice, a transition 88 is made to return to “voice state” 82. On the other hand, when the silence count value in the silence count storage unit 92 exceeds a certain value (for example, 26), a transition 89 to the “silence state” 80 is performed.

  In a preferred embodiment, the end point detection unit 44 may be configured using the maximum likelihood method described in US Pat. No. 6,711,536. However, it will be apparent to those skilled in the art that using the maximum likelihood method does not strongly affect the present invention.

  In a preferred embodiment, it is possible to perform the noise tracking operation only during a “silent state”. That is, when there is a possibility that the input signal is voice after the end point is detected, the operation of the noise tracking unit 36 is then prohibited. As a result, the noise tracking unit can always reliably hold the background noise level, not the voice level.

(Other embodiments)
In the above-described embodiment, the present invention has been described with respect to voice and voice recognition. However, in addition to voice, the present invention can also be applied to acoustic signals such as animal calls and machinery sounds. It can also be applied to acoustic signals outside the normal human audible range, such as sonar and animal calls. Furthermore, the present invention can also be applied to electromagnetic signals such as radar or radio signals.

  In the above-described embodiment, the amplitude of the signal is used as the feature quantity. However, an arbitrary intensity signal such as energy or power may be used.

  In the above-described embodiment, the DFT is realized by the sliding DFT, but other DFTs that exist in large numbers can also be used. For example, but not limited to, the Goertzel algorithm, direct DFT calculation, and fast Hartley transform.

  As mentioned above, although embodiment of this invention was explained in full detail, this invention may be applied to the system comprised from several apparatuses, and may be applied to the apparatus which consists of one apparatus.

  In the present invention, a program for realizing each function of the above-described embodiments is supplied directly or remotely to a system or apparatus, and a computer included in the system or apparatus reads and executes the supplied program code. Can also be achieved.

  Accordingly, since the functions and processes of the present invention are implemented by a computer, the program code itself installed in the computer also implements the present invention. That is, the computer program itself for realizing the functions and processes is also one aspect of the present invention.

  In this case, the program may be in any form as long as it has a program function, such as an object code, a program executed by an interpreter, or script data supplied to the OS.

  Examples of the computer-readable recording medium for supplying the program include a flexible disk, a hard disk, an optical disk, a magneto-optical disk, an MO, a CD-ROM, a CD-R, and a CD-RW. Examples of the recording medium include a magnetic tape, a non-volatile memory card, a ROM, a DVD (DVD-ROM, DVD-R), and the like.

  The program may be downloaded from a homepage on the Internet using a browser on a client computer. That is, the computer program itself of the present invention or a compressed file including an automatic installation function may be downloaded from a home page to a recording medium such as a hard disk. Further, it is also possible to divide the program code constituting the program of the present invention into a plurality of files and download each file from a different home page. That is, a WWW server that allows a plurality of users to download a program file for realizing the functions and processing of the present invention on a computer may be a constituent requirement of the present invention.

  The program of the present invention may be encrypted and stored in a computer-readable storage medium such as a computer-readable CD-ROM and distributed to users. In this case, only the user who cleared the predetermined condition is allowed to download the key information to be decrypted from the homepage via the Internet, decrypt the program encrypted with the key information, execute it, and install the program on the computer May be.

  Further, the functions of the above-described embodiments may be realized by the computer executing the read program. Note that an OS or the like running on the computer may perform part or all of the actual processing based on the instructions of the program. Of course, also in this case, the functions of the above-described embodiments can be realized.

  Furthermore, the program read from the recording medium may be written in a memory provided in a function expansion board inserted into the computer or a function expansion unit connected to the computer. Based on the instructions of the program, a CPU or the like provided in the function expansion board or function expansion unit may perform part or all of the actual processing. In this way, the functions of the above-described embodiments may be realized.

It is a figure explaining the example of application to the audio | voice communication apparatus of VAD. It is a figure explaining the example of application to ASR of VAD. It is a figure which shows the specific structural example of VAD38. It is a figure which shows the structure of the signal detection apparatus in embodiment. It is a figure which shows the structure of the endpoint detection part in embodiment. It is a figure which shows the state transition determination process in embodiment.

Claims (8)

A calculation step for calculating an intensity signal indicating the intensity of the signal divided into frames;
A filtering step for filtering the intensity signal;
A tracking step for tracking a noise level of the intensity signal;
A signal detection method comprising: a determination step of determining the signal using the intensity signal filtered in the filtering step and the noise level output in the tracking step.   2. The signal detection method according to claim 1, wherein in the determination step, the signal is determined using a threshold value offset by a predetermined amount from the noise level output in the tracking step.   The signal detection method according to claim 2, wherein the signal is an audio signal.   A speech recognition method, wherein the speech is recognized by discriminating speech and background noise using the signal detection method according to claim 3. Calculating means for calculating an intensity signal indicating the intensity of the signal divided into frames;
Filter means for filtering the intensity signal;
Tracking means for tracking the noise level of the intensity signal;
A signal detection apparatus comprising: a determination unit configured to determine the signal using the intensity signal filtered by the filter unit and the noise level output from the tracking unit.   The program for making a computer perform the signal detection method of any one of Claims 1 thru | or 3.   A program for causing a computer to execute the speech recognition method according to claim 4.   A computer-readable storage medium storing the program according to claim 6 or 7.

Images