Synthetic triphones from trajectory-based feature distributions

Abstract

We experiment with a new method to create synthetic models of rare and unseen triphones in order to supplement limited automatic speech recognition (ASR) training data. A trajectory model is used to characterise seen transitions at the spectral level, and these models are then used to create features for unseen or rare triphones. We find that a fairly restricted model (piece-wise linear with three line segments per channel of a diphone transition) is able to represent training data quite accurately. We report on initial results when creating additional triphones for a single-speaker data set, finding small but significant gains, especially when adding additional samples of rare (rather than unseen) triphones.