This paper presents a probabilistic phone mapping model (PPM) that makes possible automatic speech recognition using a foreign phonetic system. We formulate the training of the phone mapping model in the framework of maximum likelihood estimation. The model can be learned automatically from the reference phonetic transcript and the phonetic transcript resulting from a foreign phonetic recogniser using the Expectation Maximisation algorithm. This paper also compares the use of temporal and spatial contexts to enchance the phone mapping performance. A decision tree clustering technique is used to tie unseen contexts for robustness. We evaluate the PPM method on cross-lingual phone and isolated word recognition tasks, using the Hungarian and Russian phone recognisers to recognise Czech speech. Consistent improvement is obtained by using context-dependent phone mapping.
Bibliographic reference. Sim, Khe Chai / Li, Haizhou (2008): "Context-sensitive probabilistic phone mapping model for cross-lingual speech recognition", In INTERSPEECH-2008, 2715-2718.