Properties of Auditory Model Representations.

摘要: We address the problem of robustness auditory models as front ends for speech recognition. Auditory have been referred superior when is corrupted by noise or linear filtering, but there not yet a deep understanding its functioning. analyze some commonly used and show that they present interesting properties which are useful robust In our view, short-time adaptation provided hair cell key factor this robustness. A disadvantage distributions obtained features well represented gaussian pdfs. discuss parameter transformation in order to use standard recognizer based on CDHMMs with pdfs digit recognition experiments.