The joint multigram model is a statistical model, which achieves a many-to-many mapping between two, or more, streams of symbols. It relies on a fully unsupervised training phase, during which variable-length sequences of symbols from each stream are matched together according to a maximum likelihood criterion. Then the model can be used to perform sequence-by-sequence decoding. It is evaluated for a task of automatic orthographic-phonetic transcription, and results in a 95 % phoneme accuracy on the French corpus BDLEX. Preliminary experiments to test the ability of the model for a task of continuous speech recognition are also reported.
Bibliographic reference. Deligne, Sabine / Yvon, Francois / Bimbot, Frédéric (1995): "Variable-length sequence matching for phonetic transcription using joint multigrams", In EUROSPEECH-1995, 2243-2246.