Informed vs. Blind beamforming in AD-HOC Acoustic Sensor Networks for Meeting Transcription

We consider a network of initially unsynchronized microphone arrays to be used to capture a meeting which is afterwards transcribed. Beamforming is applied to exploit the spatial diversity of the setup for signal enhancement. We propose and compare two approaches to compute the beamformer coefficients. The first, informed beamforming, localizes speakers and sensors and computes speaker activity information, from which beamformer filter coefficients are derived, while the second, blind beamforming, estimates the beamforming coefficients in an unsupervised manner employing a spatial mixture model. We discuss the pros and cons of the two approaches and experimentally assess their sensitivity to synchronization errors, localization errors, erroneous activity information, etc. Simulations show that the informed beamforming achieves a promising performance as measured by the word error rate of a downstream speech recognizer.