An attribute detection based approach to automatic speech processing

Abstract

State-of-the-art automatic speech and speaker recognition systems are often built with a pattern matching framework that has proven to achieve low recognition error rates for a variety of resource-rich tasks when the volume of speech and text examples to build statistical acoustic and language models is plentiful, and the speaker, acoustics and language conditions follow a rigid protocol. However, because of the “blackbox” top-down knowledge integration approach, such systems cannot easily leverage a rich set of knowledge sources already available in the literature on speech, acoustics and languages. In this paper, we present a bottom-up approach to knowledge integration, called automatic speech attribute transcription (ASAT), which is intended to be “knowledge-rich”, so that new and existing knowledge sources can be verified and integrated into current spoken language systems to improve recognition accuracy and system robustness. Since the ASAT framework offers a “divide-and-conquer” strategy and a “plug-andplay” game plan, it will facilitate a cooperative speech processing community that every researcher can contribute to, with a view to improving speech processing capabilities which are currently not easily accessible to researchers in the speech science community.