Knowledge-Based Analysis of Speech Mixed with Steady and Sporadic Environmental Sounds

点击翻译按钮获取中文摘要

This is a Small Grant for Exploratory Research (SGER) project. The project investigates the feasibility of using knowledge from the steady and sporadic background noise sources that occur in arbitrary environments not conditioned for speech recognition to isolate them from speaker-independent articulation of telephone numbers in a recognition task. The environmental noisy background could be found in home, offices, and automobiles; the types of steady-state noise include those originating from an air conditioner or from a car engine, while the sporadic noise sources are discrete events such as a person walking, a door slammed, a ringing telephone, or an appliance being switched on. The problem is addressed on the basis of a signal-understanding architecture with context-dependent reconfiguration for the feature extraction process. The results from an initial round of feature extraction are analyzed for time-frequency regions consistent with the system's knowledge regarding the correlates of various speech and non-speech sources. Those regions of higher certainty - or islands of certainty - are considered in the prediction of potentially masked speech by non-speech sources that may occur in other parts of the utterance. Those regions are used also as a basis in the reconfiguration of the feature-extraction process which is used in analyzing the masked region as well as in verifying, pruning, and refining hypotheses in all regions. In summary, while the goal of the system is to recognize speech, it must also perform the recognition on non-speech sounds to help the speech recognition process. //