superhgq1 / vad-1

Voice Activity Detection System

VAD

Voice Activity Detection System

VoiceActivityDetector implements a Vocal Activity Detection algorithm. It works by using voiced band energy ratios, periodicity measures, running and dynamic minimum and maximum RMS energy estimates, adaptive and noise resistant threshold computation, and hangover smoothing to fade in and out speech/noise boundaries. The final output is a reconstructed voice only audio file, together with a visual representation showing the VAD marked sections and the voice plot

Usage

VoiceActivityDetector(input, noise) - Performs VAD on the input audio signal 'input'. 'noise' is an optional wav of noise to mix into the input audio file.

For input, the following voice audio files are present in the folder:

male1.wav - sample male voice male2.wav - sample male voice male3.wav - sample male voice female1.wav - sample female voice female2.wav - sample female voice

For noise, the following files are present in the folder:

'white.wav' - standard white noise, constant level 'pink.wav' - standard pink noise, constant level 'babble.wav' - generic people talking, dynamically changing levels 'music.wav' - music used as noise, dynamically changing levels

The noise file is optional. You can simply provide your own already noisy signal directly as 'input' if you'd like, or you can also provide your own individual voice and noise files separately.

NOTE: If you are using only a single noisy input file, please make sure the noise is reasonably softer than the voice input. The VAD works best when the noise is around 10of the signal amplitude. Also, ensure the first ~300ms of the audio is only noise (no voice).

Finally , to play around with the accuracy of the VAD, you can try tweaking the following variables directly in the code itself:

frameLength - ~10 to 20ms (if you want to go below 10, you will need to change the upperLagMS value in Periodicity.m to be lower)

hangOverThreshold - ~3 to 10 (# of unvoiced frames to pad
unconditionally after the end of a voiced frame

noiseScalingFactor - ~0.05 to ~0.2 (amplification of noise compared to voice signal expressed as a percentage