Laboratoire de Linguistique Formelle, CNRS/Université Paris 7, 2 place Jussieu, case 7003, 75251, Paris, FRANCE,
last revised: 6 feb 98
I present a few ideas from Hearing Science applicable to Speech Recognition. One has to do with the choice of time and frequency resolution for feature analysis, especially in relation to periodicity. Another has to do with source separation and interference rejection (the cancellation principle). A third has to do with how to compensate for damaged or missing evidence in the recognition stage. This last idea should be useful in a variety of contexts, including blind separation, multimodal speech recognition, intonation, etc..