engleski

A Method for Estimating Variations in Speech Tempo from Recorded Speech

In this paper we describe a method for measuring variations in speech tempo from speech samples recorded from Croatian news channels, where the text of what was spoken is available through subtitles. For speech recognition we use a feed- forward neural network trained with about 150 seconds of speech. To extract word boundaries, we created a speech-to-text aligner capable of finding an acceptable match between text and sequence of phonemes classified by the neural network. The aligner takes into consideration certain categories of phonemes for which the neural network has higher accuracy. Preliminary experiments show average alignment miss of about one to three phonemes, depending on the speaker, the content, and recording quality.

speech recognition, text-to-speech alignment, speech tempo, neural network

nije evidentirano