CN1623186A - Voice activity detector and validator for noisy environments - Google Patents

Abstract

A communication unit (100) comprising an audio processing unit (109) with a voice activity detection mechanism (130, 135). A voice activity detection mechanism (130, 135) measures an energy acceleration rate of a signal input into the communication unit (100) and determines from the measurement whether the input signal is speech or noise. A method of detecting speech and a method of deciding whether an input signal is speech or noise are also described. Using energy acceleration rate based voice activity detectors and verifiers offers the advantages of noise robustness, fast response and input speech level independence especially for noisy environments.

Description

Voice Activity Detector and Authenticator for Noisy Environments

technical field

The present invention relates to the detection of speech in noisy environments, commonly referred to as voice activity detection (VAD). The invention is applicable to (but not limited to) energy acceleration rate measurement of speech signals in speech detection systems.

Background technique

Many voice communication systems, such as the Global System for Mobile Communications (GSM) cellular telephone standard and Terrestrial Trunked Radio (TETRA) systems for personal mobile radio users, use a voice processing unit to encode and decode voice patterns. In such voice communication systems, a vocoder converts the analog speech pattern into a suitable digital format for transmission. The voice decoder converts the received digital voice signal into audio analog voice mode.

Methods and apparatus for detecting voice activity are known in the art. A Voice Activity Detector (VAD) works under the assumption that speech is only present in a part of the audio signal. This assumption is usually correct because many audio signal intervals have only silence or background noise.

Voice activity detectors can be used for many purposes. These include suppressing the entire transmission activity in the transmission system when there is no speech, thereby potentially saving power and channel bandwidth. When the VAD detects that voice activity is continuing, transmission activity can be restarted.

Voice activity detectors can also be used in conjunction with speech storage devices to distinguish portions of the audio that include speech from "no speech" portions. The portion including speech is later stored in a storage device and the "no speech" portion is discarded.

Existing methods for detecting speech are based at least in part on methods for detecting and estimating the power of speech signals. The estimated power is compared to a constant or an adaptive threshold to make a decision whether the signal is speech. The main advantage of these methods lies in their low complexity, which makes them suitable for implementation with low processing resources. The main disadvantage of this approach is that background noise may inadvertently cause "speech" to be detected when there is actually no "speech". Also, actual "speech" may go undetected because of ambiguity and be difficult to detect due to background noise.

Some methods for detecting voice activity are aimed at noisy mobile environments and are based on adaptive filtering of voice signals. This reduces the noise content from the signal before final decision. As the method is used for different speakers and different environments, the frequency spectrum and noise level may change. Therefore, input filters and thresholds are usually adaptive to track these changes.

Examples of these methods are provided in the GSM specification 06.42 Voice Activity Detector (VAD) for half-rate, full-rate and enhanced full-rate voice traffic channels respectively. Another such method is "Multi-Boundary Voice Activity Detection Algorithm (multi-boundary voice activity detection algorithm)" suggested in Appendix B of ITU G.729. These methods are accurate in noisy environments, but are complex to implement.

All of these methods require an input speech signal. Some applications employing speech decompression schemes need to perform speech detection during the speech decompression process.

European Patent Application No. EP-A-0785419 by Benyassine et al. relates to a method for voice activity detection comprising the following steps:

(i) extracting a predetermined set of parameters from each frame of the incoming speech signal, and

(ii) making a frame speech decision for each frame of the incoming speech signal based on a set of deviation measures extracted from a predetermined set of parameters.

VADs in cellular systems are biased to ensure that when one party speaks, the wireless equipment, including speech codecs and RF circuitry, is activated to transmit that speech to the other party amid background noise and other impairments. However, this results in data transmissions occurring when one party is not speaking. This approach comes at the cost of slightly reduced battery life and slightly increased interference to co-channel users in other elements of the system. These are basically second (or higher) order effects.

In these systems, there is no conception that limited resources are available for duplex calls. Usually the uplink and downlink on different carriers can use the entire bandwidth in unison at the same time.

It is known in the field of the invention that some voice activity or voice onset detectors (VAD/VOD) attempt to use speech characteristics such as harmonic structure (eg by autocorrelation) to distinguish voiced speech. However, in noise, these structural indicators may fail due to breakdown of the speech structure or due to structure in noise. This can be, for example, engine, tire or air conditioning noise in a car. Finally, these methods are weak in detecting unvoiced speech.

An alternative to this is simply to use the frame energy levels to detect speech. This is satisfactory for speech in high signal-to-noise ratio (SNR) conditions, where an arbitrary threshold above the noise level can be set to represent speech. However, this approach fails in many real noise conditions.

For non-normalized databases or in practical applications, the noise level in one example set is likely to be higher than the speech level in another example set, which makes it impossible to set the threshold value. Existing methods to overcome this problem are to take an average of about the first 100 milliseconds of an utterance, assuming this represents noise, thereby creating a specific threshold for that utterance. However, moreover, this is not sufficient for non-stationary noise, where the noise may deviate rapidly from the initial estimate, where the noise has high variance or where the first few frames actually contain speech that is not supposed to be noise.

Therefore, there is a need for an improved voice activity detector and verifier for noisy environments that can alleviate the above disadvantages.

Contents of the invention

According to a first aspect of the present invention there is provided a communication unit as claimed in claim 1 .

According to a second aspect of the present invention there is provided a method of detecting a speech signal input into a communication unit as claimed in claim 11.

According to a third aspect of the present invention, there is provided a method of determining whether a signal input into a communication unit is speech or noise as claimed in claim 14 .

Other aspects of the invention are described in the dependent claims.

In summary, the present invention aims to address the case of non-stationary noise of arbitrary magnitude by using an energy acceleration rate measurement, preferably an energy magnitude measurement, to indicate the presence or absence of speech.

Description of drawings

Exemplary embodiments of the invention will now be described with reference to the accompanying drawings, in which:

Figure 1 shows a block diagram of a communication unit suitable for performing voice activity detection and verification of a preferred embodiment of the present invention;

Fig. 2 shows the flow chart of the voice activity detector based on the energy acceleration rate for the noise environment according to the preferred embodiment of the present invention;

FIG. 3 shows a flow chart of energy acceleration rate-based voice activity verification for noisy environments according to a preferred embodiment of the present invention; and

Figure 4 illustrates buffer operation according to a preferred embodiment of the present invention.

Detailed ways

Voiced speech has a relatively high energy acceleration rate value because the onset of voiced speech depends on the activity of either vibrating or stationary vocal cords. Similarly, unvoiced onsets (eg plosives) also have high energy acceleration rates.

The inventors have realized that in domains typically characterized by distinct speech features, such as the narrowband power spectrum or the Mel spectrum, the resulting energy acceleration rates are much higher than for non-stationary noise. The only major exception is impact noise (e.g. clapping).

Therefore, according to a preferred embodiment of the present invention, the inventors have found that by concentrating the energy in the frequency region which may contain the fundamental pitch of the speech signal, it is additionally possible to distinguish from these noises. In particular, the inventors of the present invention propose to use a non-structural feature of speech, namely energy acceleration rate (or acceleration rate reflecting some measure of speech energy or its components).

In particular, a preferred application for the inventive concept described here 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 (speech processing, transmission and quality aspects (STQ); distributed speech recognition; front-end feature extraction algorithm; compression algorithm)", ETSI ES 201 108 vl.1.2 (2000-04), April 2000.

Referring now to FIG. 1, there is shown a block diagram of an audio subscriber unit 100 suitable for supporting the inventive concepts of the preferred embodiment of the present invention.

A preferred embodiment of the invention is described in terms of a wireless audio communication unit, such as a wireless audio communication unit capable of operating under the 3rd Generation Partnership Project (3GPP) standard for future cellular wireless communication systems and providing DSR capabilities. However, the inventive concepts described herein with respect to voice activity detection and verification are equally applicable to any electronic device that responds to voice signals and could benefit from improved voice activity detection circuitry and is within the scope of the present invention.

As is known in the art, the audio subscriber unit 100 includes an antenna 102 that is preferably connected to a duplex filter, antenna switch, or circulator 104 that provides a link between the receive chain and the transmit chain within the audio subscriber unit 100. isolation.

The receiver chain includes receiver front-end circuitry 106 (effectively providing reception, filtering, and IF or baseband frequency conversion). Front-end circuitry 106 is connected in series to a signal processing functional block (typically implemented by a digital signal processor (DSP)) 108 . Signal processing functional block 108 performs signal demodulation, error correction and formatting. The data recovered from signal processing function 108 is connected in series to audio processing function 109 which formats the received signal in a suitable manner for transmission to audio producer/display 111 .

In different embodiments of the present invention, the signal processing functional block 108 and the audio processing functional block 109 may be disposed in the same physical device. Controller 114 is positioned to control the information flow and operating status of the components of subscriber unit 100 .

As for the transmission chain, this basically consists of an audio input device 120 , which connects an audio processing function 109 , a signal processing function 108 , a transmitter/modulation circuit 122 and a power amplifier 124 in series. Processor 108, transmitter/modulation circuit 122, and power amplifier 124 are operatively responsive to the controller. The power amplifier output is connected to a duplex filter, antenna switch or circulator 104 and antenna 102 to transmit the final radio frequency signal.

Specifically, the audio processing functional block 109 includes a voice activity (or voice onset) detection (VAD) functional block 130 , which is operatively connected to a voice activity decision functional block 135 . According to a preferred embodiment of the present invention, VAD functional block 130 and voice activity decision functional block 135 are adapted to provide an improved voice detection and decision mechanism, the operation of which will be further described with reference to FIGS. 2 and 3 . It should be noted that the voice activity detector functional block 130 includes a frame-by-frame detection stage consisting of three measurements. The three frequency range measurements include:

(i) the entire spectrum;

(ii) spectrum sub-bands; and

(iii) Spectral variance.

Next, the voice activity decision function block 135 performs a decision based on the measured buffer, analyzing its speech likelihood. The application of the final decision of the decision stage goes back to the earliest frame in the buffer.

In a preferred embodiment of the present invention, timer/counter 118 is also adapted to perform timing functions in the detection and determination processes of FIGS. 2 and 3 .

The signal processor functional block 108, the audio processing functional block 109, the VAD functional block 130 and the voice activity decision functional block 135 may be implemented as distinct, operatively connected processing components. Additionally, one or more processors may be utilized to implement one or more corresponding processing operations. In another alternative embodiment, the functional blocks described above may be implemented as a mix of hardware, software or firmware components, using application specific integrated circuits (ASICs) and/or processors such as digital signal processors (DSPs).

Of course, the various components within audio subscriber unit 100 may be implemented as separate or integrated components, so that the final structure is only a matter of arbitrary choice.

To achieve this, there is a method of obtaining an indication of the energy acceleration rate used in the preferred embodiment of the present invention.

(i) A theoretically ideal approach would be to accurately double-differentiate the energy levels over successive frames of the utterance, as shown in the prior published application US 6009391. The downside of this approach is that it can cause delays as multiple frames on each side of the frame need to be analyzed when analyzing.

(ii) A zero-delay estimate of the energy acceleration rate can be obtained by comparing the short-term average with the instantaneous value, for example:

Use frame averaging:

$\stackrel{<span\; class="notranslate">\; ~</span>}{\mathrm{<span\; class="notranslate">\; A</span>}}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; +</span><span\; class="notranslate">\; \&CenterDot;</span><span\; class="notranslate">\; \&Center\; Dot;</span><span\; class="notranslate">\; \&Center\; Dot;</span><span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; no</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; /</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ]</span>$

or use a rolling average:

$\stackrel{<span\; class="notranslate">\; ~</span>}{\mathrm{<span\; class="notranslate">\; A</span>}}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; ax</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; bx</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; +</span><span\; class="notranslate">\; \&CenterDot;</span><span\; class="notranslate">\; \&CenterDot;</span><span\; class="notranslate">\; \&CenterDot;</span><span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; no</span>}}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; ]</span>$

In each case, the method returns a value which can be interpreted as 'deceleration rate'<'1'<'acceleration rate'. can then be found The empirical value of and the denominator length that best distinguishes speech from noise.

The inventors of the present invention have realized that the preferred optimal solution is to find a denominator that can quickly track non-stationary noise, but it is too long for tracking speech onset. The suggested sequence of values for the rolling average is a=0.2, b=0.8xa, c=0.8xb, etc., which can be expressed simply as a recurrence:

d _t =0.2x _t +0.8d _t-1 [3]

but:

A=x _t /d _t [4]

The preferred VAD and parameter initialization system within the detection phase is outlined in the flowchart of FIG. 2 . In nonstationary noise, long-term energy thresholds are not reliable indicators of speech. Similarly, under high noise conditions, the structure of speech (eg, harmonics) cannot be relied upon entirely for indications, since it may be corrupted by the noise, or the structure noise may confuse the detector. Therefore, preferred voice activity detectors use the noise-robust characteristics of speech, ie the rate of acceleration of energy associated with speech onset.

Referring now to FIG. 2 , a flowchart 200 of a preferred detection process is shown. As noted above, this processing includes frame-by-frame analysis. Preferably the VAD mechanism involves measurement processing of the 'whole spectrum'. The frame counter is initially evaluated to determine if it is less than 'N', which defines the number of buffered frames, as shown in step 205. As an example of a preferred embodiment, 'N' is set to '15', assuming a setting of eg 10 milliseconds per frame increment. If the frame counter is less than 'N' in step 205, the rolling average of the initial acceleration test is updated as in step 210. If the frame counter is not less than 'N' in step 205, then step 210 is skipped.

A determination is then made as to whether the estimated energy acceleration rate measurement is within one or more specified limits, as shown in step 235 . If the energy acceleration rate measurement is within one or more specified limits in step 235 , the rolling average is updated with the results of further energy acceleration rate tests, as in step 240 . If the energy acceleration rate measurement is not within one or more specified limits in step 235, then step 240 is skipped.

A determination is then made as to whether the estimated energy acceleration rate measurement is greater than a specified threshold, as shown in step 260 . If the energy acceleration rate measurement is greater than the specified threshold in step 260, the frame is considered to be a speech frame, as in step 265. If the energy acceleration rate measurement is not greater than the specified threshold in step 260 , then the frame is considered to be a noisy frame, as in step 270 .

The frame counter is then incremented, step 275, and the process repeats from step 205 onwards.

As an improvement to this process, instead or in addition, an entire spectrum measurement process, such as the sub-region measurement process shown in optional steps 215 and 245, may also be performed. A specific subregion of the spectrum is selected as one that is likely to contain the fundamental pitch.

In this subfield process, when the rolling average of the initial acceleration test is updated across the spectrum measurements in step 210 , a determination is made to check whether the energy acceleration rate measurement is greater than a threshold value, as shown in step 220 . If the energy acceleration rate measurement is greater than the threshold value in step 220 , then the process of initializing other parameters is suspended, as shown in step 225 . If the energy acceleration rate measurement is not greater than the threshold value in step 220 , update the initialization of other parameters, as in step 230 . The process then returns to step 235, as shown.

A further preferred determination is made after the determination made in step 235 whether the estimated energy acceleration rate measurement is within one or more specified limits. This deceleration rate value is evaluated to determine if it was 'high' in step 250, and if so, the rolling average of the energy acceleration rate test is slowly updated, as shown in step 255. The process then returns to the full spectrum method at step 260 .

In this way, the higher signal-to-noise ratio (SNR) of the subfield detector makes it more robust to noise. However, it is susceptible to adverse microphone and speaker variations and band-limited noise. Therefore, this measurement should not depend on all circumstances. Therefore, the preferred embodiment of the present invention incorporates subregion detectors to enhance overall spectral measurements.

A further measurement process is preferably performed using eg 'acceleration' of the variance of values within the lower half of the frequency spectrum for each frame. This variance measure detects structures within the lower half of the spectrum, making it highly sensitive to voiced speech. The variance measure follows the approach of subfield processing, with the lower half of the spectrum being the specific subfield chosen. This variance measure further complements the whole spectral measure method, which enables better detection of unvoiced and plosive speech.

All three measurements take their original input from the spectral representation of the filter gain produced by the first stage of the double Wiener filter, as in US 09/427497 to Motorola Corporation and inventor Yan-Ming Chen described in the application. As noted above, each measurement uses a different aspect of this data.

Specifically, the entire spectrum detector uses the known Mel-filtered spectral representation of the filter gain produced by the first stage of the double Weiner filter. A single input value is obtained by squaring the sum of the Mel filterbank.

In a preferred embodiment of the invention, the entire spectrum detector applies the following processing to all frames, as follows:

Step 1 initializes the noise estimation tracking value (Tracker) in the following way:

If the number of frames < 15 and the acceleration rate < 2.5,

Then tracking value = MAX(tracking value, input).

If speech occurs within the lead-in time of 15 frames, the energy acceleration rate measurement prevents the tracking value from being updated.

If the current input is the same as the noise estimate, step two updates the tracked value in the following way:

If input < tracking value × upper limit and

Input > Tracking Value × Lower Limit,

Then tracking value = a × tracking value + (1-a) × input

Step three provides a fail-safe mechanism for those instances where there is speech or featureless loud noise content within the first few frames. This results in a reduction in the resulting erroneous high-noise estimates. Step 3 is preferably carried out in the following manner:

If input<tracking value×lowest value (Floor),

Then tracking value=b×tracking value+(1-b)×input

If the current input is 165% greater than the tracked value, step four returns as the 'true' speech determination in the following manner:

If input>tracking value×threshold,

Then output 'true', otherwise output 'false'.

The ratio of the instantaneous input to the short-term mean tracking value is a function of the energy acceleration rate of the continuous input.

Among them, in the above:

a=0.8 and b=0.97;

The upper limit is 150% and the lower limit is 75%;

The minimum value is 50%; and

The threshold is 165%.

It should be noted that if the value is greater than the upper bound or between the lower bound and the lowest value, it is not updated. Additionally, as noted above, the energy acceleration rate input can be calculated as follows:

Quadratic derivative on continuous input or estimated by tracking the ratio of two rolling averages of the input.

It should be noted that the ratio of the fast and slow adaptive rolling averages reflects the rate of energy acceleration of the continuous input.

For example, the contribution to the average used above is:

(i) 0× mean + 1× input, and

(ii) ((frame number-1)×mean+1×input)/frame number,

Make energy acceleration measurement more and more sensitive to the first fifteen frames.

The sub-band detector preferably uses the average of the second, third and fourth Mel filter banks measured from the 'whole spectrum'. The detector then applies the following processing to all frames in the manner described below:

(i) input = p × current input + (1-p) × previous input;

(ii) If the number of frames < 15,

Then tracking value = MAX(tracking value, input);

(iii) If input < tracking value × upper limit and

Input > Tracking Value × Lower Limit,

Then tracking value=a×tracking value+(1-a)×input;

(iv) If input < tracking value × minimum value,

Then tracking value=b×tracking value+(1-b)×input

(v) If input>tracking value×threshold,

Then output 'true', otherwise output 'false'.

Among them, in the sub-region measurement:

p = 0.75.

Except for the threshold equal to 3.25, all other parameters are the same for the entire spectrum measurement.

For the spectral variance measurement, the variance of the values of the lower frequency half of the narrowband spectral representation comprising the gain per frame is used as input. The detector then applies the same processing to the entire spectrum measurement.

This variance is calculated as:

$\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; N</span>}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; -</span>{<span\; class="notranslate">\; (</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; /</span>{\mathrm{<span\; class="notranslate">\; N</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; ]</span>$

in:

N = FFT length/4, and

w _i is the value represented by the narrowband spectrum of the gain.

According to a preferred embodiment of the present invention, these three measurements detailed above are provided to the VAD decision algorithm, as shown in the flowchart of FIG. 3 . Continuous input is provided to a buffer, which provides contextual analysis. This makes the frame delay equal to the buffer length minus one frame.

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

For an N=7 frame buffer, the most recent true/fake speech input is stored at position N in the data buffer, as shown in step 305 . The decision logic applies several of the following steps, and preferably applies each step:

step 1:

V _N = measure 1 or measure 2 or measure 3

Input V _N is defined as 'true' (T) if any of these three measures returns a true speech indication.

Step 2:

The algorithm searches the buffer for the longest consecutive sequence of 'true' values, step 310. Thus, for example, for the sequence 'TTFTTTF' M is equal to 3.

Step 3:

If M≥S _P and T<L _S , T=L _S ;

Wherein, S _P is equal to the first threshold in step 315 . If in step 315 the longest sequence of true (T) speech values equals or exceeds the first threshold, i.e. S _P =3 or more consecutive 'true' values, the buffer is judged to contain 'possible' voice. If it is determined in step 320 that it has not already existed (or exceeded), then in step 325 a short timer T(time_1) of eg L _S =5 frames is started.

Step 4:

If M≥S _L and F>F _S , T=L _M , otherwise T=L _L ;

Wherein, _SL is equal to the second threshold in step 330 . If there are _SL = 4 or more consecutive 'true' values, the buffer is again judged to contain 'likely' speech. If the current frame F, as determined in step 335, is outside the initial lead-in security period _FS , then in step 340 an intermediate timer T of eg L _M =22 frames is started. Otherwise, in step 345 a failsafe long timer T such as _LL = 40 frames is used. Using this arrangement when speech occurs early in the utterance would overstate the initial noise estimate of the VAD.

Step 5:

If M<S _P and T>0, T--;

If the process determines in step 350 that there are consecutive 'true' values less than S _P =3 and the timer is greater than zero in step 355 , then the timer is decremented in step 360 .

Step 6:

If T>0, output 'true', otherwise output 'false';

If the timer is greater than zero in step 365, then the process outputs a 'true' speech decision, as shown in step 370. Additionally, if the timer is not greater than zero in step 365, then the process outputs a 'noise' decision, as shown in step 375.

Step 7:

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 to the left to accommodate the next input, as shown with respect to FIG. 4 . The outgoing speech decisions are applied to frames coming out of the buffer. The process is then repeated in step 305 for the next true/false input into the data buffer.

Alternative mechanisms for implementing speech or noise decisions based on energy acceleration rate processing as described above are also within the contemplation of the present invention. For example, the decision mechanism may not be based on one or more timers, but may be based entirely on whether one or more energy acceleration rate thresholds are exceeded.

Referring now to FIG. 4 , an example of buffer operation 400 is shown in greater detail in accordance with a preferred embodiment of the present invention. We assume that the first threshold is set to three consecutive 'true' values. At "t" 410, only the current input (frame #7) 425 and previous input (frame #6) 420 are assumed to be 'true'. Therefore, when the buffer is shifted, the first frame (frame #1) 415 will be marked false.

At 't+1' 430, a third 'true' input (frame #8) 450 has been received to supplement the previous two 'true' inputs 440 and 445. Therefore, when the buffer is shifted, the next output frame (frame #2) 435 will be marked 'true'.

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

(i) time_1<time_2<time_3, and

(ii) threshold_1<threshold_2.

Assuming only these three inputs (frame #6, frame #7 and frame #8) are 'true', the entire output sequence is:

F T T T T T T T T T T T T T

1 2 3 4 5 6 7 8 9 10 11

T T T T T T T F F F F F F F

12 13 14 15 16 17 18 19 20 21 22

Among them, frames #2-#5 indicate 'true' due to the buffer import function. Frames #6-#8 indicate 'true' as the location of the actual initial 'true' speech input. Frames #9-#12 indicate 'true' due to the buffer export function. Frames #13-#18 indicate 'true' in response to the timer delay used. When all frames in the utterance have been input, the buffer is shifted out of 'false' entries (frames #19-#L _M ) until empty.

It is also within the scope of the present invention that the buffer length and delay timer can be dynamically adjusted to meet the needs of the audio communication unit. Again, using the preferred embodiment of a buffer length of 8 for 'N' and a delay timer of 5 frames is for explanatory purposes only. However, it should be noted that the buffer length 'N' should always be determined as N≥S _L .

In addition to being used as a VAD itself, the energy acceleration rate measurement performed in the method step of Fig. 2 can be used to verify the initialization of other parameters, which is also within the scope of the present invention. For example, spectral extraction schemes require an initial estimate of noise based on the first ten frames of speech (typically 100 milliseconds). Even in stationary noise, several events may occur that invalidate the initial estimate. Examples of such events include:

(a) Upslope of the signal:

For various possible reasons, at valuation, the start of recording may 'ramp up' to full value within the period. Reasons for full ramp up include: buffer filling in digital systems, capacity or lead connections in analog systems. The impact of these events invalidates this valuation. Therefore, energy acceleration measurements can be used to detect such ramp-ups and prevent such errors.

(b) Glitch in the initial signal:

A common 'glitch' occurs with the complete actuation of the push-to-talk (PTT) button on the subscriber wireless unit, where electrical contact rarely occurs before the button strikes the back of the switch. As mentioned above, when such an event occurs, the energy acceleration rate measurement may be used to suspend the evaluation process, as shown in step 225 of FIG. 2 .

(c) Speech in the initial signal:

Another common occurrence, particularly with PTT systems, is that the user immediately starts speaking when pressing the PTT button. In this way, the electrical contact is made after speech starts. The energy acceleration rate measurement can identify this and suspend noise-based initialization, as shown in step 225 of Figure 2, or force the use of fault estimates.

In summary, a communication unit comprising an audio processing unit with a voice activity detection mechanism has been described. A voice activity detection mechanism provides an indication of the energy acceleration rate of a signal input to the communication unit and determines from the indication whether the input signal is speech or noise.

Furthermore, a method of detecting a voice signal input into a communication unit has been described. The method includes the steps of: indicating an acceleration rate of an input signal input to the communication unit; and determining whether the input signal is speech or noise based on the indicating step.

Furthermore, the method of judging whether the signal input into the communication unit is speech or noise has been described. The method includes the following steps: judging whether the input signal is speech or noise according to the energy acceleration rate, for example, using frame average or rolling average of several input signals.

Therefore, it should be appreciated that the energy acceleration rate based voice activity detector and verifier for noisy environments as described above provides the advantages of noise robustness and fast response. Since the preferred embodiment uses an energy acceleration rate dependent measure, rather than an absolute measure, the inventive concepts described herein can be applied to speech at any input level.

While specific and preferred implementations of embodiments of the invention have been described above, it should be apparent that those skilled in the art can readily apply changes and modifications of such inventive concepts that fall within the scope of the invention.

Accordingly, an improved voice activity detector and verifier for use in noisy environments has been described in which the above-mentioned disadvantages associated with prior art arrangements are substantially eliminated.

Claims (15)

1. a communication unit (100), it comprises and has voice activity detection mechanism (130,135) audio treatment unit (109), described communication unit (100) is characterised in that, described voice activity detection mechanism (130,135) measure the energy rate of acceleration that is input to the signal in the described communication unit (100), and determine that according to described measurement described input signal is voice or noise.

2. communication unit as claimed in claim 1 (100), wherein, described voice activity detection mechanism comprises speech activity detector functional block (130), it is to being input to the detection frame by frame of the signal execution speech in the described voice activity detection mechanism (130,135).

3. communication unit as claimed in claim 2 (100), wherein, the described detection frame by frame comprises at one or more in the following frequency range and carries out the energy rate of acceleration and measure being input to signal in the described voice activity detection mechanism (130,135):

(i) entire spectrum

(ii) frequency spectrum frequency sub-band; And

(iii) frequency spectrum variance.

4. communication unit as claimed in claim 3 (100), wherein, described voice activity detection mechanism comprises voice activity decision function piece (135), whether it may be operably coupled to described speech activity detector functional block (130), be voice to adjudicate described input signal according to the buffer operation of one or more described measurements.

5. communication unit as claimed in claim 4 (100), wherein, described voice activity decision function piece (135) uses the frame of a plurality of described input signals average or roll whether on average adjudicate input signal be voice.

6. as each the described communication unit (100) in the claim 2 to 5, wherein,, think that then incoming frame is speech frame (265) if described energy rate of acceleration measures the energy rate of acceleration value greater than energy rate of acceleration thresholding.

7. communication unit as claimed in claim 6 (100) wherein, determines that incoming frame is the frame that the application of the judgement (265) of speech frame can trace back to the front in the impact damper of input signal.

8. as claim 6 or the described communication unit of claim 7 (100), wherein, if for a plurality of successive frames, described energy rate of acceleration measures the energy rate of acceleration value greater than energy rate of acceleration thresholding, thinks that then incoming frame is speech frame (370).

9. when depending on claim 3, as each the described communication unit (100) in the claim 3 to 8, wherein, if select the subarea of input signal spectrum, then this selection is based on the basic fundamental tone that the subarea most possibly comprises voice signal and makes.

10. as the described communication unit of each claim (100) of front, wherein, described voice activity detection mechanism (130,135) uses the rate of acceleration of the correlated characteristic of speech energy to verify the parameter initialization of the correlated measure of other speech or noise, for example frequency spectrum extraction scheme.

11. a detection inputs to the method for the voice signal in the communication unit, it is characterized in that, comprises following steps:

Measurement inputs to rate of acceleration or the variation in the energy of the input signal in the described communication unit; And

Determine that according to described measuring process (315,330,350) described input signal is voice (370) or noise (375).

12. voice signal detection method as claimed in claim 11 is characterized in that, further comprises following steps:

To inputing to the detection frame by frame of the signal execution speech in the described communication unit.

13. voice signal detection method as claimed in claim 12, wherein, the described detection frame by frame may further comprise the steps:

At one or more following frequency ranges, described input signal is carried out the energy rate of acceleration measures:

(i) entire spectrum

(ii) frequency spectrum frequency sub-band; And

(iii) frequency spectrum variance.

14. the signal that a judgement inputs in the communication unit is the voice or the method for noise, preferably according to each claim in the claim 11 to 13 of front, the method is characterized in that, further comprises following steps:

According to the energy rate of acceleration in the energy measurement of described input signal or change that to adjudicate (315,330,350) described input signal be voice (370) or noise (375), for example use the frame of a plurality of input signals average or roll on average.

15. the signal that judgement as claimed in claim 14 inputs in the communication unit is the voice or the method for noise, wherein, described decision steps comprises:

If described energy rate of acceleration measures energy rate of acceleration value greater than energy rate of acceleration thresholding, determine that then incoming frame is speech frame (265); And

The frame of the front in the described impact damper of determining to be applied to input signal with reviewing.

Images