Automatic Phoneme Segmentation with Relaxed Textual Constraints

Speech synthesis by unit selection requires the segmentation of a large single speaker high quality recording. Automatic speech recognition techniques, e.g. Hidden Markov Models (HMM), can be optimised for maximum segmentation accuracy. This paper presents the results of tuning such a phoneme segmentation system. Firstly, using no text transcription, the design of an HMM phoneme recogniser is optimised subject to a phoneme bigram language model. Optimal performance is obtained with triphone models, 7 states per phoneme and 5 Gaussians per state, reaching 94.4% phoneme recognition accuracy with 95.2% of phoneme boundaries within 70 ms of hand labelled boundaries. Secondly, using the textual information modeled by a multi-pronunciation phonetic graph built according to errors found in the first step, the reported phoneme recognition accuracy increases to 96.8% with 96.1% of phoneme boundaries within 70 ms of hand labelled boundaries. Finally, the results from these two segmentation meth...

• Type: Conference
• Year: 2008
• Where: LREC
• Authors: Pierre Lanchantin, Andrew C. Morris, Xavier Rodet, Christophe Veaux
• Keywords: Education, LREC 2008, Phoneme Boundaries, Phoneme Recognition Accuracy, Phoneme Segmentation System