JP2010061151A - Voice activity detector and validator for noisy environment - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an improved voice activity detector and validator for noisy environments. <P>SOLUTION: A communication unit 100 includes an audio processing unit 109 having a voice activity detection mechanism 130, 135. The voice activity detection mechanism 130, 135 measures an energy acceleration of a signal input to the communication unit 100 and determines whether the input signal is speech or noise, based on the measurement. A method of detecting voice and a method of deciding whether an input signal is voice or noise are also described. Using an energy acceleration based voice activity detector and validator, particularly for noisy environments, provides the advantages of noise robustness, fast response and independence of the level of input speech. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

[Field of the Invention]
The present invention relates to voice detection (typically known as voice activity detection (VAD) in noisy environments). The present invention is applicable to energy acceleration measurement of a sound signal in a sound detection system, but is not limited thereto.

[Background of the invention]
Many voice communication systems, such as the Global System for Mobile Communications Cellular Telephone Standard (GSM), and the Terrestrial Trunk Radio (TETRA) system for personal mobile radio users, use voice processors to encode and decode voice patterns. ing. In such a voice communication system, the voice encoder converts the analog voice pattern into an appropriate digital format for transmission. The audio decoder converts the received digital audio signal into an audible analog audio pattern.

  Methods and apparatus for detecting voice activity are known in the art. A voice activity detector (VAD) operates under the assumption that speech is present only in a portion of the audio signal. This assumption is usually correct. This is because there are many audio signal intervals that exhibit only silence or background noise (background noise).

  Voice activity detectors can be used for many purposes. This includes suppressing overall transmission activity in the transmission system when there is no speech, thus potentially saving power and channel bandwidth. When VAD detects that voice activity has resumed, VAD can resume transmission activity.

  A voice activity detector can also be used in connection with a voice storage device by distinguishing an audio part containing voice from an audio part that is "no voice". Thus, the part containing the voice is stored in the storage device and the "no voice" part is discarded.

  Conventional methods for detecting speech are based, at least in part, on methods for detecting and evaluating the power of speech signals. The estimated power is compared to some constant or adaptive threshold to determine whether the signal was speech. The main advantage of these methods is their low complexity, which makes them suitable for the realization of low processing resources. The main drawback of such a method is that background noise can erroneously result in detecting “speech” when it is not actually present. On the other hand, since the existing “voice” is not clear, the “voice” may not be detected and may be difficult to detect due to background noise.

  Some methods of detecting speech activity are directed to noisy mobile environments and are based on adaptive filtering of speech signals. This reduces the noise component from the signal before final determination. The frequency spectrum and noise level can vary as the method is used for different speakers and in different environments. Thus, input filters and thresholds are often adaptive to track these variations.

Examples of these methods are given for each of half-rate, full-rate and augmented full-rate voice traffic channels in the GSM specification 06.42 voice activity detector (VAD). Another such method is described in ITUG. 729 Appendix B, “Multi-boundary voice activity detection algorithm”. These methods are more accurate in noisy environments, but are significantly more complicated to implement.

  All these methods require inputting an audio signal. Some applications that employ voice recovery schemes (voice decompression schemes) require performing voice detection during the voice recovery process.

European patent application no. EP-A-075419 includes the following steps: (i) extracting a predetermined set of parameters for each frame from the incoming speech signal;
(Ii) directed to voice activity detection including the step of making a frame speechization of the incoming speech signal for each frame according to a set of difference measurements extracted from a predetermined set of parameters.

  A cellular system VAD allows a wireless device, including a voice codec and RF circuitry, to send voice to other third parties in the presence of background noise and other obstacles when the third party emits voice. In order to ensure that it is in an active state (active state) for transmission, it is in a biased state. However, this leads to sending data when the third party is not speaking. This loss slightly reduces battery life and slightly increases interference to co-channel users in other cells of the system. These are essentially secondary (or higher order) effects.

  In these systems, there is no notion that a finite resource is available for duplex calls. It is possible and consistent with the uplink and downlink as a whole. Note that the uplink and downlink are usually on different carriers since they use the full bandwidth simultaneously.

  In the field of this invention, some voice activity or voice start detectors (VAD / VOD) use voice characteristics such as harmonic structures (eg, via autocorrelation) to distinguish voiced voices. It is known that they are trying to do that. However, in noise, these structural indicators may fail due to speech structure collapse or due to structure in noise. This may be vehicle engine, tire or air conditioning noise. Finally, these methods are weak to detect unvoiced speech.

  An alternative method is simply to use the frame energy level to detect speech. This is sufficient for speech at high signal-to-noise ratio (SNR) conditions, where any threshold above the noise level can be set to indicate speech. However, this approach fails with more realistic noise conditions.

  For a denormalized database, or in real-world applications, it is likely that the noise level in one example is greater than the speech level in another example, which makes it impossible to set a threshold value To. The conventional method for overcoming this is that the first 100 milliseconds of the utterance represents noise, and the first 100 milliseconds of the utterance is averaged, and the utterance is ad hoc. This is a method for generating the threshold. However, again, this is not the case if the noise diverges rapidly from the initial estimate, or if the noise has a high variance, or if the first few frames actually contain speech rather than the estimated noise. Inadequate regarding.

Accordingly, there is a need for improved voice activity detectors and enablers for noisy environments that ameliorate the aforementioned drawbacks.

[Statement of invention]
According to a first aspect of the present invention, a communication device according to claim 1 is provided.

  According to a second aspect of the present invention, there is provided a method for detecting an audio signal input to a communication device according to claim 11.

  According to a third aspect of the present invention, there is provided a method for determining whether a signal input to a communication device is speech or noise according to claim 14.

  Further aspects of the invention are claimed in the dependent claims.

  In summary, the present invention aims to address the case of non-stationary noise of arbitrary amplitude by using energy acceleration measurements rather than energy amplitude measurements to indicate the presence or absence of speech.

  Exemplary embodiments of the present invention will now be described with reference to the accompanying drawings.

FIG. 1 shows a block diagram of a communication device adapted to perform voice activity detection and validation of a preferred embodiment of the present invention. FIG. 2 shows a flow chart of an energy acceleration based voice activity detector for a noisy environment, according to a preferred embodiment of the present invention. FIG. 3 shows a flow chart of enabling energy acceleration based voice activity for a noisy environment according to a preferred embodiment of the present invention. FIG. 4 illustrates the buffer operation according to the preferred embodiment of the present invention.

[Description of Preferred Embodiment]
Voiced speech (speech) has a relatively high energy acceleration value because it depends on vocal cord activity whose onset is either oscillating or stationary. Similarly, an unvoiced start (eg, a pop) also has a high energy acceleration value.

  The inventor has confirmed that the resulting energy acceleration is significantly higher than non-stationary noise in a representational domain that emphasizes voicing, such as a narrowband power spectrum or a mel spectrum. The only significant exception is impulse noise (eg applause).

  Thus, in accordance with a preferred embodiment of the present invention, the inventor can additionally distinguish against these noises by collecting frequency domain energy that is likely to include the fundamental pitch of the audio signal. I understood that I could do it. In particular, the inventors of the present invention propose to use unstructured characteristics of speech, ie energy acceleration (or acceleration of some metric representing speech energy or its components).

In particular, a preferred application for the inventive concepts described herein is the Distributed Speech Recognition (DSR) standard currently being defined by the European Telecommunications Standards Institute (ETSI): “Speech Processing, Transmission and Quality Aspects ( STQ); distributed speech recognition; front-end feature extraction algorithm; compression algorithm ”(ETSIES 201 108 vl.1.2 (20
00-04), April 2000).

  Referring now to FIG. 1, there is shown an audio subscriber unit 100 that is adapted to support the inventive concept of the preferred embodiment of the present invention.

  Preferred embodiments of the present invention are wireless audio communication devices, eg, wireless audio communication capable of operating with the 3rd Generation Partnership Project (3GPP) standard for future cellular wireless communication systems and providing DSR capabilities. The description will be made with reference to the apparatus. However, the inventive concepts described herein relating to voice activity detection and its validation are equally applicable to any electronic device that responds to a voice signal and benefits from an improved voice activity detection circuit. It is within the spirit of the present invention.

  As is known in the art, the audio subscriber unit 100 preferably includes an antenna 102 coupled to a double filter, an antenna switch or a switch that separates the receive chain and the transmit chain within the audio subscriber unit 100. A circulator 104 is included.

  The receiver chain includes a receiver front-end circuit 106 (effectively performing reception, filtering and intermediate or baseband frequency conversion). The front-end circuit 106 is connected in series with a signal processing function 108 (generally realized by a digital signal processor (DSP)). The signal processing function 108 performs signal demodulation, error correction, and formatting. The recovered data from the signal processing function 108 is coupled in series to an audio processing function 109, which formats the received signal in an appropriate manner to an audio initiator / display 111. Send to.

  In various embodiments of the present invention, the signal processing function 108 and the audio processing function 109 may be provided in the same physical device. The controller 114 is configured to control the information flow and operating state of the components of the audio subscriber unit 100.

  With respect to the transmit chain, this essentially includes an audio input device 120 coupled in series through an audio processing function 109, a signal processing function 108, a transmitter / modulation circuit 122 and a power amplifier 124. The processor 108, transmitter / modulation circuit 122 and power amplifier 124 are operatively responsive to the controller 114. The output of the power amplifier 124 is coupled to a double filter, antenna switch or circulator 104 and antenna 102 to radiate the final radio frequency signal.

In particular, the audio processing function 109 includes a voice activity (or voice start) detection (VAD) function 130 operatively coupled to the voice activity determination function 135. In accordance with the preferred embodiment of the present invention, VAD function 130 and voice activity determination function 135 are adapted to provide an improved voice detection and determination mechanism, the operation of which is further described with respect to FIGS. . In particular, the voice activity detector function 130 includes a frame-by-frame detection stage consisting of three measurements. The three frequency range measurements are
(I) the entire spectrum;
(Ii) spectral subbands, and (iii) spectral dispersion.

  Subsequently, the voice activity determination function 135 performs a determination based on a buffer of measurements that are analyzed for voice likelihood. The final decision from the decision stage is applied retrospectively to previous frames in the buffer.

  In the preferred embodiment of the present invention, timer / counter 118 is also adapted to perform timing functions in the detection and determination process of FIGS.

  Signal processor function 108, audio processing function 109, VAD function 130 and voice activity determination function 135 may be implemented as separate processing components that are operatively coupled. Alternatively, one or more processors may be used to perform one or more of the corresponding processing operations. In yet another alternative embodiment, the functions described above use an application specific integrated circuit (ASIC) and / or processor, such as a digital signal processor (DSP), as a mix of hardware and software, or firmware components. Can be executed.

  Of course, the various components within the audio subscriber unit 100 can be implemented in individual or integrated component forms, and thus the final structure is merely an arbitrary choice.

  For this reason, there are several ways to achieve an indication of energy acceleration for use in the preferred embodiment of the present invention.

  (I) The theoretically ideal method is described in previously published US patent application no. As seen in 6009391, it is a method of literally differentiating energy levels over successive frames of speech. The disadvantage of this approach is that it is likely to introduce a delay because one needs to analyze a certain number of frames under analysis on each side of the frame.

(Ii) Zero delay estimation of energy acceleration can be achieved by comparing the short-term average ratio with the instantaneous value, for example:
Frame average

Or Rolling Average

Can be used.

In each case, the method returns a value that can be interpreted as ｀ deceleration '<｀ 1'<｀ acceleration '. The person then enters the empirical value for A ^~ (where "X ^~ " represents the symbol with ^~ on the symbol X), and the denominator length that best distinguishes speech from noise. ) Can be found.

The inventor of the present invention finds a denominator that can quickly track non-stationary noise, but it is too long to track speech start. The proposed sequence of values for the rolling average is a = 0.2, b = 0.8 * a, c = 0.8 * b, etc., which can be simply expressed as a regression method.

d _t = 0.2x _t +0.8 d _{t −1} [3]

Therefore,

A = x _t / d _t [4]

A preferred VAD and parameter initialization system within the detection stage is summarized in the flow chart of FIG. In non-stationary noise, the long-term energy threshold is not a reliable indicator of speech. Similarly, in high noise conditions, speech structure (eg, harmonics) relies entirely as an indicator because harmonics can be corrupted by noise or structured noise can disrupt the detector. I can't. Thus, the preferred voice activity detector uses a noise robust characteristic, ie energy acceleration associated with voice onset.

Referring now to FIG. 2, a flow chart 200 of a preferred detection process is shown. As indicated above, this detection process involves frame-by-frame analysis. A preferred VAD mechanism is associated with the “whole spectrum” measurement process. As shown in step 205, the frame counter is first evaluated to determine if it is less than "N". “N” defines the number of frames to be buffered. As an example of the preferred embodiment, assuming that each frame is incremented by, for example, 10 milliseconds, “N” is set to “15”. In step 205, if the frame counter is less than “N”, the rolling average for the initial acceleration test is updated, as shown in step 210. In step 205, if the frame counter is not smaller than "N", step 210 is skipped.

  A determination is then made to evaluate whether the energy acceleration measurement is within one or more specified margins, as shown in step 235. In step 235, if the energy acceleration measurement is within one or more specified margins, the rolling average is updated with the results of the further energy acceleration test, as in step 240. If, in step 235, the energy acceleration measurement is not within one or more specified margins, step 240 is skipped.

  A determination is then made to evaluate whether the energy acceleration measurement is greater than a specified threshold, as shown in step 260. In step 260, if the energy acceleration measurement is greater than the specified threshold, the frame is considered a speech frame, as in step 265. If, in step 260, the energy acceleration measurement is not greater than the specified threshold, the frame is considered a noise frame, as in step 270.

  The frame counter is then incremented, as in step 275, and the process is repeated from step 205.

  As an improvement to this process, instead of or in addition to the entire spectrum measurement process, the sub-region measurement process shown in optional steps 215 and 245 may be performed. A particular subregion of the spectrum is selected because that subregion is most likely to contain the basic pitch.

  In the sub-region measurement process, once the rolling average for the initial acceleration test is updated in step 210 in the overall spectrum measurement, to check whether the energy acceleration measurement is greater than the threshold value, as shown in step 220. Make a decision. In step 220, if the energy acceleration measurement is greater than the threshold value, the process of initializing other parameters is interrupted, as shown in step 225. In step 220, if the energy acceleration measurement is not greater than the threshold value, the initialization of other parameters is updated as in step 230. The process then returns to step 235 as shown.

  In step 235, after the above determination, a more suitable determination is made to evaluate whether the energy acceleration measurement is within one or more specified margins. In step 250, the deceleration value is evaluated to determine if the deceleration value is "high", and if so, the rolling average for the energy acceleration test is slowly updated, as shown in step 255. Then, in step 260, the process returns to the full spectrum method.

In this way, the generally higher signal-to-noise ratio (SNR) of the sub-region detector makes it very robust against noise. However, it is vulnerable to adverse microphone and speaker changes and band-limited noise. Therefore, in all situations, you should not rely on measurements. Accordingly, the preferred embodiment of the present invention incorporates a sub-region detector to increase the overall spectral measurement.

  The further measurement process is preferably performed, for example, using the “acceleration” of the variance of values within the lower half of the spectrum of each frame. The dispersion measurement detects the structure in the lower half of the spectrum and makes it very sensitive to voiced speech. Dispersion measurements follow the subregion process approach, and the lower half of the spectrum is the specific subregion selected. This dispersion measurement further complements the whole spectrum measurement approach, which can better detect unvoiced and bursting speech.

All three measurements were taken from the raw inputs of U.S. Patent Application No. No. 5, filed by Motorola and inventor Yan-Ming Chen. Extract from the spectral representation of the filter gain generated by the first stage of the double Wiener filter as described in 09/427497. As described above, each measurement uses a different aspect of this data.

  In particular, the overall spectrum detector uses a known Mel-filtered spectral representation of the filter gain generated by the first stage of the double Wiener filter. A single input value is obtained by squaring the sum of mel filter banks.

  In the preferred embodiment of the present invention, the full spectrum detector applies the following process to every frame, as described below.

Step 1 initializes a noise estimate tracker as follows.

When frame <15 and acceleration <2.5, tracker = MAX (tracker, input)

The energy acceleration measurement prevents the tracker from being updated if the sound occurs within 15 frames of lead-in time.

Step 2 updates the tracker value as follows if the current input is similar to the noise estimate.

If input <tracker * upper limit and input> tracker * lower limit,
Tracker = a * tracker + (1−a) * input Step 3 provides a fail-safe mechanism for those instances where speech or non-characteristically large noise components in the first few frames are present. This results in the resulting incorrect high noise estimate for attenuation. Step 3 preferably functions in the following manner.

If input <tracker * floor,
Tracker = b * tracker + (1−b) * input Step 4 returns as a “true” voice decision in the following manner if the current input is 165% greater than the tracker.

If input> tracker * threshold,
Outputs true, otherwise it outputs false.

The ratio of the instantaneous input to the short-term average tracker is a function of the energy acceleration of the continuous input.

Here, in the above,
a = 0.8 and b = 0.97
The upper limit is 150% and the lower limit is 75%
The floor is 50% and the threshold is 165%.

In particular, there is no update if the value is greater than the upper limit or between the lower limit and the floor. Furthermore, as indicated above, the energy acceleration input is
It can be calculated either as a double-differentiation of successive inputs or as an estimate by tracking the ratio of the two rolling averages of the inputs.

  In particular, the ratio of the fast adaptive rolling average to the slow adaptive rolling average reflects the continuous input energy acceleration.

As an example, the contribution to the average used above is
(I) 0 * average + 1 * input, and (ii) ((frame-1) * average + 1 * input) / frame, making energy acceleration measurements increasingly sensitive over the first 15 frames.

  The subband detector preferably uses the average of the second, third and fourth mel filter banks derived for the “overall spectrum” measurement. The detector then applies the following process to all frames as described below.

(I) input = p * current input + (1-p) * previous input (ii) if frame <15,
Tracker = MAX (tracker, input)
(Iii) If input <tracker * upper limit and input> tracker * lower limit,
Tracker = a * tracker + (1-a) * input (iv) If input <tracker * floor,
Tracker = b * Tracker + (1-b) * Input (v) If Input> Tracker * Threshold,
Outputs true, otherwise it outputs false.

Here, in the sub-region measurement, p = 0.75.

  All other parameters are the same for the overall spectrum measurement, except for the threshold, which is equal to 3.25.

  For the measurement of the spectral variance, the variance of the value having the lower half of the narrowband spectral representation of the gain for each frame is used as input. The detector then applies exactly the same process for the measurement of the whole spectrum.

  The variance is calculated as follows:

here,
N = the FFT length / 4, and w _i is a narrow-band spectrum display of the value of the gain.

  In accordance with the preferred embodiment of the present invention, the three measurements described above are provided to the VAD determination algorithm as shown in the flow chart of FIG. Sequential input is provided to the buffer, which provides context analysis. This introduces a frame delay equal to the length of the buffer minus one frame.

  Referring now to FIG. 3, a flow chart 300 of an acceleration-based voice activity activation process for a noisy environment is shown in accordance with a preferred embodiment of the present invention.

  For N = 7 frame buffers, the most recent true / false audio input is stored at location N in the data buffer, as shown in step 305. Decision logic provides a number of the following steps, preferably one of them.

Step 1:
V _N = measurement 1 or measurement 2 or measurement 3
Input V _N is defined as “true” if any of the three measurements returns a true voice indication.

  Step 2:

  The algorithm looks for the longest continuous sequence of “true” values in the buffer, as in step 310. Thus, for example, for the sequence “T T F T T T F”, M would be equal to “3”.

Step 3:
If M ≧ S _P and T _L <L _S , then T = L _S
Here, _{S P} is handled equated with the first threshold in step 315. FIG. In step 315, if the longest sequence of true (T) speech values is equal to or exceeds the first threshold, ie, S _P = 3 or more consecutive “true” values, the buffer is It is determined to contain “possible” speech. If at step 320 it does not already exist (or has been exceeded) from the decision, at step 325 a short timer (time 1), eg L _S = 5 frames, is activated.

Step 4:
If M ≧ S _L and F> F _S , then T = _TL .
In other cases, it is T = _{L L.}

Here, S _L are handled regarded as equivalent to the second threshold in step 330. If S _L = 4 or more consecutive “true” values, the buffer is again determined to contain “likely” speech. If the current frame F is outside the initial lead-in safety period F _S , as determined in step 335, then in step 340, for example, an intermediate timer T of L _M = 22 frames is activated. Otherwise, in step 345, for example, a long fail-safe timer T of L _L = 40 frames is used. Such a configuration is used because the early presence of speech in the utterance can make the VAD initial noise estimate too high.

Step 5:
M In the case of _{<S P,} and T> 0, T--
If, at step 350, the process determines that it is a continuous “true” value less than S _P = 3, and at step 355, the timer is greater than zero, then at step 360, the timer is decremented.

Step 6:
If T> 0, “true” is output, otherwise “false” is output.

  In step 365, if the timer is greater than zero, the process outputs a "true" audio decision, as shown in step 370. Alternatively, if the timer is not greater than zero, the process outputs a “noise” determination, as shown in step 375.

Step 7:
For frame ++, shift the buffer to the left and return to step 1.

  In preparation for the next frame in step 380, the buffer is shifted left to accept the next input, as shown in FIG. The output audio decision is applied to the frame that is being emitted from the buffer. In step 305, the process repeats for the next true / false input to the data buffer.

  It is within the spirit of the present invention that an alternative mechanism for making speech or noise determinations can be implemented based on the energy acceleration process described above. For example, the decision mechanism may not be based on one or more timers and may simply determine whether one or more energy acceleration thresholds have been exceeded.

  Referring now to FIG. 4, an example of a buffer operation 400 according to a preferred embodiment of the present invention is shown in more detail. Suppose that the first threshold is set for three consecutive “true” values. Suppose that at time (time) “t” 410, only the current input (frame # 7) 425 and the previous input (frame # 6) 420 were “true”. Thus, when the buffer is shifted, the first frame (frame # 1) 415 is marked “false”.

  At time (time) “t + 1” 430, a third “true” input (frame # 8) 450 is received, supplementing two earlier “true” inputs 440, 445. Thus, when the buffer is shifted, the next output frame (frame # 2) 435 is marked “true”.

  It should be noted that in the above decision process, only the constraints are:

(I) Time 1 <Time 2 <Time 3 and (ii) Threshold 1 <Threshold 2.

  Assuming that only these three inputs (frame # 6, frame # 7 and frame # 8) are “true”, the full output sequence is:

Here, frames # 2 to # 5 indicate “true” due to the buffer lead-in function. Frames # 6 to # 8 indicate “true” as the actual original “true” audio input position. Frames # 9- # 12 indicate “true” due to the buffer lead-out function. Frames # 13- # 18 indicate “true” in response to the timer hangover used. Once all the frames in the utterance have been input, the buffer shifts the “false” input (frames # 19- # L _M ) until empty.

It is within the intent of the present invention that the buffer length and hangover timer can be adjusted dynamically to meet the requirements of the audio communication device. As such, the preferred embodiment using a buffer length “N” of 8 and a 5 frame hangover timer is used for illustrative purposes only. However, it should be noted that the buffer length “N” should always be determined such that N ≧ S _L.

  In addition to using it as its own right as VAD, it is the intent of the present invention that the initialization of other parameters can be validated using the energy acceleration measurement performed in the method step of FIG. Is within. For example, spectral subtraction schemes require an initial estimate of noise based on the first 10 frames of speech (typically 100 milliseconds). Even in stationary noise, several events can occur to invalidate the initial guess. Examples of such events include:

(A) Ramp-up of signal:
Due to various possible causes, the full start of recording can “ramp up” to full volume within a cycle under evaluation. The reasons behind such full ramp-up include buffer filling in digital systems, capacitance, or tape-head engagement in analog systems. The effect of such an event will invalidate the estimation. Thus, energy acceleration measurements can be used to detect such ramp-up and prevent errors there.

(B) Spikes in the initial signal:
A common "spike" occurs with the full arrangement of "press-to-talk" (PTT) buttons on the subscriber radio device, where electrical contact is , Just slightly ahead of the button hitting the back of the switch. When such an event occurs, the energy acceleration measurement described above can be used to interrupt the estimation process, as shown in step 225 of FIG.

(C) Audio in the initial signal:
Another common occurrence, especially in the case of PTT systems, is that the user begins to speak as soon as the user presses the PTT button. In this way, electrical contact is made after talking begins. The energy acceleration measurement can identify this and so interrupt the noise-based initialization or force the use of the default estimate, as shown in step 225 of FIG.

  In summary, a communication device including an audio processing device having a voice activity detection mechanism has been described. The voice activity detection mechanism provides an indication of energy acceleration of a signal input to the communication device, and determines whether the input signal is voice or noise based on the instruction.

  In addition, a method for detecting an audio signal input to a communication device has been described. The method includes instructing acceleration of an input signal to the communication device and determining based on the instructing step whether the input signal is speech or noise.

  Further, a method for determining whether a signal input to the communication apparatus is voice or noise has been described. The method includes determining whether the input signal is speech or noise based on energy acceleration, for example, using a frame average or rolling average of a number of input signals.

  Thus, the energy acceleration based voice activity detector and enabler described above for noisy environments provides the advantages of robustness and fast response to noise. Since the preferred embodiment uses energy acceleration dependent measurements instead of absolute measurements, the inventive concepts described herein can be applied to speech at any input level.

  While specific and preferred implementations of embodiments of the present invention have been described above, those skilled in the art will readily be able to apply such changes and modifications of the inventive concept that fall within the scope of the present invention. It is clear that it is deaf.

  Accordingly, an improved voice activity detector and enabler for a noisy environment has been described in which the aforementioned drawbacks associated with prior art configurations have been substantially improved.

Claims (18)

A communication device (100) comprising an audio processing device (109) having a voice activity detection mechanism (130, 135),
The voice activity detection mechanism (130, 135) measures an energy acceleration of a signal input to the communication device (100) by calculating a ratio between an average energy value and an instantaneous energy value, and Configured to determine frame by frame based on the measurement whether the input signal is speech or noise;
The communication device (100), wherein the input frame is determined to be an audio frame if the energy acceleration measurement yields an energy acceleration value greater than an energy acceleration threshold (265).   The voice activity detection mechanism includes a voice activity detector function (130) configured to perform per-frame detection of speech on a signal input to the voice activity detection mechanism (130, 135). Item 4. The communication device (100) according to Item 1.   The frame-by-frame detection includes: (i) performing an energy acceleration measurement on the entire spectrum for a signal input to the voice activity detection mechanism (130, 135); (ii) performing the energy acceleration measurement on a subband of the spectrum. And (iii) performing the energy acceleration measurement by using the acceleration of the variance of values within the low frequency half of the spectrum of each frame. The communication device (100) described. The voice activity detection mechanism is operatively coupled to the voice activity detector function (130), based on a buffering operation of an input frame of the input signal into a buffer, and one or more of the one or more of the A voice activity determination function (135) configured to determine whether the input signal is speech according to an energy acceleration measurement;
The voice activity determination function (135) is further configured to assign a true or false indication to each of the buffered input frames in the buffer, the one or more energy acceleration measurements for the input frames. A true indication is assigned if any one of the commands returns a voice indication, and the voice activity determination function (135) is further assigned for each of the buffered series of input frames in the buffer. 4. The communication device (100) of claim 3, wherein the communication device (100) is configured to determine that the input signal in the buffer is audio if the indication is true.   The communication device (100) according to any of claims 1 to 4, wherein the voice activity detection mechanism is configured to measure energy acceleration using a frame average or rolling average of a number of the input signals. .   The energy acceleration is estimated by tracking the ratio of the two rolling averages of the input signal using (0 × average + 1 × input) and ((Frame-1) × average + 1 × input) / Frame, respectively. The communication device (100) according to any one of claims 1 to 4, wherein Frame represents a value of a frame counter. The estimation of the energy acceleration using the frame average is

The communication device (100) according to claim 5, wherein: If the energy acceleration measurement is within one or more specified margins, the estimation of energy acceleration using a rolling average is:

The communication device (100) according to claim 5 or 6.   The buffer has a buffer length of N frames, and if a continuous input frame is passed to and released from the buffer and it is determined that the input frame in the buffer is a voice frame, the input frame The communication device (100) of claim 4, wherein the determination that is a speech frame (265) is applied retroactively to earlier frames in the buffer.   The communication device (100) according to any one of claims 3, 4 or 9, wherein the subband includes a basic pitch of an audio signal. A method of detecting an audio signal input to a communication device,
Measuring an energy acceleration of a signal input to the communication device (100) by calculating a ratio of an average energy value and an instantaneous energy value;
Determining whether the input signal is speech (370) or noise (375) on a frame basis based on the measuring step (315, 330, 350), wherein the energy acceleration measurement comprises: The method wherein the input frame is determined to be a speech frame if an energy acceleration value greater than the energy acceleration threshold is generated (265).   The frame-by-frame measurement includes (i) performing an energy acceleration measurement on the input signal over the entire spectrum, (ii) performing the energy acceleration measurement over a subband of the spectrum, and (iii) 12. The method of claim 11 including one or more of: performing the energy acceleration measurement by using an acceleration of a variance of values within the low frequency half of the spectrum of each frame.   13. A method according to claim 11 or 12, wherein the step of measuring energy acceleration uses a frame average or rolling average of a number of input signals.   The energy acceleration is estimated by tracking the ratio of the two rolling averages of the input signal using (0 × average + 1 × input) and ((Frame-1) × average + 1 × input) / Frame, respectively. 13. The method according to claim 11, wherein Frame represents a value of a frame counter. Measuring the energy acceleration comprises:

The method of claim 13, comprising estimating the energy acceleration using a frame average by calculating The step of measuring the energy acceleration comprises: the energy acceleration measurement being within one or more specified margins;

15. The method according to claim 13 or 14, comprising estimating the energy acceleration using a rolling average by calculating. The buffer has a buffer length of N frames, and successive input frames are passed to and released from the buffer, the method comprising:
12. The method of claim 11, further comprising: retroactively applying the determination to earlier frames in the buffer if it is determined that the input frame in the buffer is a speech frame. The determining step includes:
Buffering an input frame of the input signal in a buffer;
Assigning a true or false indication to each of the buffered input frames in the buffer, wherein a true indication is assigned if an energy acceleration measurement for the input frame returns a voice indication;
12. The method of claim 11, further comprising: determining that the input signal in the buffer is speech if the indication assigned to each of the buffered series of input frames in the buffer is true. Method.

Images