Knowledge-Based Speech Signal Representation

点击翻译按钮获取中文摘要

The goal of this study is to use a recently developed automatic optimization procedure to further the design of a knowledge based speech signal representation based on phonetic features, the message bearing components of the speech signal. The knowledge based signal representation consists of acoustic parameters that are exact measures performed on the speech signal or its time frequency representation to target the linguistic information in the speech signal. Such a representation can serve as the front end of a speech recognition system, or it can be used as a speech training for those having difficulty producing natural sounding speech (whether due to a speech impairment, hearing loss or learning of a second language). A comparison of the APs and the traditional cepstral based parameters in an HMM speech recognition system shows that the APs are better able to extract the phonetically relevant information from the speech signal and reduce speaker dependent effects. In previous research, APs were designed for the manner of articulation phonetic features using acoustic phonetic knowledge and histogram analysis to eye ball the data, a time consuming and subjective process. In contrast, the optimization procedure we developed is efficient and uses the objective Fisher criterion and classification trees. Furthermore, the optimization procedure allows us to explore many parameters to determine the one(s) that best characterizes a phonetic feature and separates it from its antonym(s). A comparison of the hand designed manner of articulation APs and the optimized ones in a speech recognition system shows that they yield comparable results. In this project, we plan to complete the knowledge based speech signal representation by using the optimization process to develop APs related to the place of articulation phonetic features and the remaining manner of articulation phonetic features.