Demos for REVERB challenge

Authors: Scott Wisdom*, Thomas Powers*, Les Atlas*, and James Pitton† *

*Electrical Engineering Department, University of Washington
† Applied Physics Lab, University of Washington

Here are a few demos of our submission to the 2014 REVERB challenge, where the goal is to dereverberate speech and remove background noise.

SimData

For SimData utterances, clean speech is convolved with a room impulse response (RIR) measured using a 8-channel circular microphone array with a 20cm diameter. Background noise is artificially added. Our algorithm takes advantage of all 8 channels to perform dereverberation and noise reduction.

Original (noisy and reverberant) Enhanced
Female 1 Female 1
Male 1 Male 1
Female 2 Female 2
Male 2 Male 2

RealData

For RealData, utterances are read from prompts in a reverberant room and recorded by an 8-channel circular microphone array with a 20cm diameter. Again, our algorithm takes advantage of all 8 channels to perform dereverberation and noise reduction.

Original (noisy and reverberant) Enhanced
Male 1 Male 1
Male 2 Male 2
Female 1 Female 1