System and method for performing voice activity detection
A Voice Activity Detection (VAD) algorithm provides a simple binary signal indicating the presence or absence of speech in a microphone signal. The VAD algorithm includes a first step of noise suppression which both estimates and removes (i.e., filters) ambient noise from the microphone signal to create a filtered signal. The magnitude of the filtered signal is then compared to a threshold in order to produce a VAD output signal. The threshold is dynamic and may be derived either from the filtered signal itself, or from a noise spectrum estimate calculated by the noise suppression step.
1. A method for Voice Activity Detection (VAD), the method comprising the steps of:
producing a microphone signal including speech and ambient noise by a microphone;
providing the microphone signal to noise suppression processing;
processing the microphone signal in the noise suppression processing to generate an ambient noise estimate, comprising:
buffering and windowing the microphone signal to create time domain frames;
transforming the time domain frames into frequency domain frame; and
computing noise estimates of each of the frequency domain frames;
providing the ambient noise estimate to threshold calculation;
calculating a threshold as the square root and mean of the noise estimates;
provided the threshold to the VAD logic;
computing a noise suppressed signal from the microphone signal and the ambient noise estimate by the noise suppression processing, comprising:
computing trim values from the noise estimates; and
scaling the frequency domain frame by the trim values;
comparing the noise suppressed signal to the threshold in the VAD logic;
setting a VAD signal to 0 when the noise suppressed signal is below the threshold; and
setting a VAD signal to 1 when the noise suppressed signal is above the threshold.
2. The method of claim 1 , wherein calculating trims from the frequency domain noise estimate frames includes using over subtraction factor and a noise floor factor to limit the trims.
3. A method for Voice Activity Detection (VAD), the method comprising the steps of:
producing a microphone signal including speech and ambient noise by a microphone;
providing the microphone signal to noise suppression processing;
processing the microphone signal in the noise suppression processing to generate an ambient noise estimate, comprising:
buffering and windowing the microphone signal to create time domain frames:
sequentially transforming the time domain frames into microphone signal frequency domain frames; and
computing frequency domain noise estimate frames of each of the microphone signal frequency domain frames;
providing the ambient noise estimate to threshold calculation;
calculating a threshold;
provided the threshold to the VAD logic;
computing a noise suppressed signal from the microphone signal and the ambient noise estimate by the noise suppression processing, comprising:
calculating trims from the frequency domain noise estimate frames; and
scaling the microphone signal frequency domain frames using the trims to obtain the noise suppressed signal;
calculating a signal energy estimate as the square of the magnitude of the noise suppressed signal;
comparing the signal energy estimate to a threshold in the VAD logic;
setting a VAD signal to 0 when the signal energy estimate is below the threshold;
setting a VAD signal to 1 when the signal energy estimate is above the threshold.
4. The method of claim 3 , wherein comparing the noise suppressed signal to a threshold in the VAD logic comprises comparing the current signal energy estimate to a threshold determined by prior signal energy estimates.
5. The method of claim 4 , wherein the threshold is determined by:
during the most recent non-speech frames, setting the threshold to the scaled result of a first order smoother applied to the prior signal energy estimates.
6. The method of claim 5 , wherein calculating trims from the frequency domain noise estimate frames includes using over subtraction factor and a noise floor factor to limit the trims.
7. A method for Voice Activity Detection (VAD), the method comprising the steps of:
producing a microphone signal including speech and ambient noise by a microphone;
buffering and windowing the microphone signal to create time domain frames;
sequentially transforming the time domain frames into microphone signal frequency domain frames;
computing frequency domain ambient noise estimate frames of each of the microphone signal frequency domain frames;
providing the frequency domain ambient noise estimate to threshold calculation;
calculating a threshold by taking the square root of each bin of the frequency domain noise estimate frames and then taking the mean of the square roots of each bins;
provided the threshold to the VAD logic;
calculating trims from the frequency domain noise estimate frames;
scaling the microphone signal frequency domain frames using the trims to obtain a frequency domain noise suppressed signal;
inverse transforming the frequency domain noise suppressed signal into a time domain noise suppressed signal;
buffering and windowing the time domain noise suppressed signal;
provided the buffered and windowed time domain noise suppressed signal to the VAD logic;
comparing the buffered and windowed time domain noise suppressed signal to a threshold in the VAD logic;
setting a VAD signal to 0 when the windowed time domain noise suppressed signal is below the threshold; and
setting a VAD signal to 1 when the windowed time domain noise suppressed signal is above the threshold.
8. A method for Voice Activity Detection (VAD), the method comprising the steps of:
producing a microphone signal including speech and ambient noise by a microphone;
buffering and windowing the microphone signal to create time domain frames;
sequentially transforming the time domain frames into microphone signal frequency domain frames;
computing frequency domain noise estimate frames of each of the microphone signal frequency domain frames;
calculating trims from the frequency domain noise estimate frames;
scaling the microphone signal frequency domain frames using the trims to obtain a frequency domain noise suppressed signal;
inverse transforming the frequency domain noise suppressed signal into a time domain noise suppressed signal;
buffering and windowing the time domain noise suppressed signal;
squaring the buffered and windowed time domain noise suppressed signal to generate a signal energy estimate;
determining a threshold using prior signal energy estimate;
comparing the present signal energy estimate to the threshold in the VAD logic;
setting a VAD signal to 0 when the signal energy estimate is below the threshold; and
setting a VAD signal to 1 when the signal energy estimate is above the threshold.
9. The method of Claim 8 , wherein the VAD signal is held to “0” during a threshold initialization period.