Smoothed local adaptation of connectionist systems

Abstract

abbot is the hybrid connectionist hidden Markov model (HMM) large vocabulary continuous speech recognition system developed at Cambridge University Engineering Department. abbot makes eective use of the linear input network (LIN) adaptation technique to achieve speaker and channel adaptation. Although the LIN is eective at adapting to new speakers or a new environment (e.g. a dierent microphone), the transform is global over the input space. In this paper we describe a technique by which the transform may be made locally linear over dierent regions of the input space. The local linear transforms are combined by an additional network using a non-linear transform. This scheme falls naturally into the mixtures of experts framework.

Classifying Unprompted Speech by Retraining LSTM Nets

“…Although we cannot compare the framewise phoneme error of BLSTM directly with the phoneme error of the HMM we expect that a BLSTM-HMM hybrid (under construction) will outperform both plain BLSTM on frame by frame and plain HMMs on the phoneme level, inheriting the best of both worlds, namely reduction of training material and training time (BLSTM), as well as more built-in structural bias (HMMs).This expectation is encouraged by experiments on read speech by Chen and Jamieson [3], Shire [14], Waterhouse, Kershaw and Robinson [16], and Elenius and Blomberg [4]. They all achieved better results on the phoneme level using an ANN-HMM hybrid approach, as shown in As can be seen from table 4 the framewise errors are quite high for noisy input sequences (several microphones or enriched with background noise) as opposed to clean speech.…”

“…Retraining both significantly reduced both time costs and training set size and improved recognition results. An extrapolation based on previous work on read speech [16,3,14,4] promises significant additional improvements on the phoneme level through a BLSTM-HMM hybrid, which we are currently implementing.…”

The 1995 ABBOT LVCSR system for multiple unknown microphones

“…Recent improvements to the ABBOT system include training of the recurrent networks for effective use of the SI284 training corpus [2], and local speaker-adaptation approaches [12], while application of state-based context-dependent phone modelling is planned for the near future.…”