On the Relation between Statistical Properties of Spectrographic Masks and Recognition Accuracy

Missing Data Techniques (MDT) can significantly improve the accuracy of automatic speech recognition (ASR) for speech corrupted by background noise. The increase in recognition accuracy obtained using MDT is largely dependent on the estimation of spectrographic masks used to distinguish speech from noise. We present an analysis technique which enables us to compare two mask estimation techniques. By contrasting a sound-class independent and a sound-class dependent distance measure, we show that we can directly relate differences between masks to their difference in recognition accuracy using the sound-class dependent distance measure. Experiments on AURORA2 using an oracle mask and an estimated mask show that modifying the estimated mask in order to reduce the statistical differences with the oracle mask leads to an increase in word recognition accuracy.