Sixth European Conference on Speech Communication and Technology
Incorporating discriminative strengths from alternative acoustic models is an important topic of recent increasing interest. Multi-resolution sub-band models and a novel phonetic segmental model independently achieve improvements on HMMs with standard MFCCs of 70.21% and 70.63% respectively from a baseline TIMIT classification score of 66.4%. Discriminatively trained weighted combination of the log likelihood scores from these acoustic modelling strategies is shown to successfully extend the performance to 72.6%.
Full Paper (PDF) Gnu-Zipped Postscript
Bibliographic reference. McCourt, Paul / Harte, Naomi / Vaseghi, Saeed (1999): "Combined temporal and spectral multi-resolution phonetic modelling", In EUROSPEECH'99, 1111-1114.