Channel selection by class separability measures for automatic transcriptions on distant microphones.

摘要: Channel selection is important for automatic speech recognition as the signal quality of one channel might be significantly better than those other channels and therefore, microphone array or blind source separation techniques not lead to improvements over best single microphone. The mayor challenge, however, find this particular who leading most accurate classification. In paper we present a novel method, based on class separability, improve multi-source far distance speech-totext transcriptions. Class separability measures have advantage, compared methods such noise ratio (SNR), that they are able evaluate actual features system. We evaluated NISTs RT-07 development set observe significant in word accuracy SNR methods. also used technique evaluation.