Pseudo-articulatory speech synthesis for recognition using automatic feature extraction from X-ray data

Describes a self-organising pseudo-articulatory speech production model (SPM) trained on an X-ray microbeam database, and present results when using the SPM within a speech recognition framework. Given a time-aligned phonemic string, the system uses an explicit statistical model of co-articulation to generate pseudo-articulator trajectories. From these, parametrised speech vectors are synthesised using a set of artificial neural networks (ANNs). We present an analysis of the articulatory information in the database used, and demonstrate the improvements in articulatory modelling accuracy obtained using our co-articulation system. Finally, we give results when using the SPM to re-score N-best utterance transcription lists as produced by the Cambridge University Engineering Department (CUED) HTK hidden Markov model (HMM) speech recognition system. Relative reductions of 18% in the phoneme error rate and 15% in the word error rate are achieved.