Novel Approaches to Speech Detection in the Processing of Continuous Audio Streams
Janez Žibert, Boštjan Vesnicer, France Mihelič · 2007
This chapter addresses the problem of speech detection in continuous audio streams and explores the impact of speech/non-speech segmentation on speech-processing applications. We proposed a novel approach for deriving speech-detection features based on phoneme transcriptions from generic speech-recognition systems. The proposed phoneme-recognition features were designed to be recognizer and language independent and could be applied in different speech/non-speech segmentation-classification frameworks. In our evaluation experiments two segmentation-classification frameworks were tested, one based on the Viterbi decoding of hidden Markov models, where speech/non-speech segmentation and detection were performed simultaneously, and the other framework, where segments were initially produced on the basis of acoustic information by using the Bayesian information criterion and then speech/non-speech classification was performed by applying Gaussian mixture models.