Example-based Realization of Isolated Words Recognizer under Limited Training Data Constraint


Today’s speech recognition systems are mostly based on Hidden Markov Models which are known to have good modeling ability based on a training database. Such a database is required to be very large, often many hours of recordings and it must contain all speech units of interest with proper transcriptions. This article describes an isolated words small vocabulary speaker-dependent recognizer which is designed to avoid the need of excessive amount of training data. The recognizer uses example-based approach built around the dynamic time warping technique and knearest neighbors classification with additional preprocessing via recurrent artificial neural network and gaussian mixture model-based endpointer.

Year: 2009
In session: Erkennung
Pages: 48 to 55