EUROSPEECH 2003 - INTERSPEECH 2003
A new data driven approach for phonetic broad class generation is proposed. The phonetic broad classes are used by tree based clustering procedure for node questions during the context dependent acoustic models generation for speech recognition. The data driven approach is based on phoneme confusion matrix, which is produced with the phoneme recogniser. Such approach enables the data driven method independency from particular language or phoneme set found in a database. Data driven broad classes generated with this method were compared to expert defined and randomly generated broad classes. The experiment was carried out with the Slovenian SpeechDat(II) database. Six different test configurations were included in the evaluation. Analysis of speech recognition results for different acoustic models showed that the proposed data driven method gives comparable or better results than standard method.
Bibliographic reference. Zgank, Andrej / Kacic, Zdravko / Horvat, Bogomir (2003): "Data driven generation of broad classes for decision tree construction in acoustic modeling", In EUROSPEECH-2003, 2505-2508.