HMM-based breath and filled pauses elimination in ASR

Abstract

The phenomena of filled pauses and breaths pose a challenge to Automatic Speech Recognition (ASR) systems dealing with spontaneous speech, including recognizer modules in Interactive Voice Reponse (IVR) systems. We suggest a method based on Hidden Markov Models (HMM), which is easily integrated into HMM-based ASR systems and allows detection of those disturbances without incorporating additional parameters. Our method involves training the models of disturbances and their insertion in the phrase Markov chain between word-final and word-initial phoneme models. Application of the method in our ASR shows improvement of recognition results in Polish telephonic speech corpus LUNA.