Acoustic Unit Discovery
While most aspects of ASR are heavily data-driven, the lexicon and acoustic units are typically derived by experts. For low-resource languages these expert-defined lexicons may not be available. We propose methods for both discovering acoustic units and learning pronunciations using the discovered units. Our initial work relies on the use of an initial grapheme-based lexicon. Acoustic models representing the graphemes are clustered to generate a new set of acoustic units. Pronunciations are generated using a statistical machine translation based approach.
Related Publications
-
“Acoustic unit discovery and pronunciation generation from a grapheme-based lexicon”, William Hartmann, Anindya Roy, Lori Lamel, Jean-Luc Gauvain, in Proceedings of IEEE ASRU, pp. 380-385, 2013. [publication] [poster] [post]
-
“Efficient rule scoring for improved grapheme-based lexicons”, William Hartmann, Lori Lamel, Jean-Luc Gauvain, in Proceedings of IEEE EUSIPCO, pp. 1477-1481, 2014. [publication] [poster]