Phone-context specific gender-dependent acoustic-models for continuous speech recognition

Abstract

Gender-dependent systems are usually created by splitting the training data into each gender and building two separate acoustic models for each gender. This method assumes that every state of a sub-phonetic model is uniformly dependent on the gender. In this work we use the premise that the acoustic realizations of various sub-phonetic units are dependent on gender in varying degrees across phones and more particularly context dependent. We show that this is indeed the case by using gender as a question in addition to phone context questions in the context decision trees. Using these trees we build phone-specific gender-dependent acoustic models and demonstrate a novel method to pick between genders during decoding based on a measure of confidence of the decoded hypothesis. An improvement of 6.3% in word error is achieved relative to a Gender-independent system.