Channel selection measures for multi-microphone speech recognition

摘要: Automatic speech recognition in a room with distant microphones is strongly affected by noise and reverberation. In scenarios where the signal captured several arbitrarily located degree of distortion differs from one channel to another. this work we deal measures extracted given distorted that either estimate its quality or measure how well it fits acoustic models system. We then apply them solve problem selecting (i.e. channel) presumably leads lowest error rate. New selection techniques are presented, compared experimentally reverberant environments other approaches reported literature. Significant improvements rate observed for most measures. A new based on variance intensity envelope shows good trade-off between accuracy, latency computational cost. Also, combination allows further improvement