After the second year, the general status of this project is quite satisfactory. The collection of the Portuguese database is complete. A pronunciation dictionary was developed, and bigram language models were extracted from the on-line version of the Portuguese newspaper PÚBLICO, with the CMU-Cambridge SLM toolkit. A French recognizer was trained on the BREF database. A text-to-phonemes system provided a lexicon and bigram and trigram language models were extracted from a large text corpus.
A software package for automatic learning of pronunciation dictionaries has been developed at ICSI. Work on a variety of novel language modelling techniques is in progress, and these being incorporated into our existing speech recognition systems. Language modelling techniques, at both word and morpheme level, have been developed for Portuguese and evaluated on text.
FPMs has developed a vocabulary independent isolated word recognizer. CUED has investigated interesting techniques to assign a confidence level to phone probabilities. Sub-band based speech recognition showed preliminary encouraging results. Assumptions at the origin of multi-band ASR have been investigated. A ``Forward-Backward'' training of an MLP has been implemented. Experiments on dimensionality reduction have been performed.
Multi-SPERT systems have been built. Due to difficulties with fixed point, we were unable to port the recurrent network training software to the SPERT board. To enable future large RNN training runs at CUED, ICSI's RAP boards are being sent to Cambridge so that their RAP capacity can be doubled. For feed-forward MLPs, though training on SPERT boards is now stable, and prototype multi-SPERT systems have been built and used for training of large networks.
The STRUT software package has been enhanced. A stable version is now available and has already been distributed. A new version of AbbotDemo has also been released.