HMM-based continuous speech recognition (CSR) systems have proved very successful despite the well-known limitations of HMMs as models of speech production. Recent work in pattern classification has demonstrated the performance advantages of combining sets of classifiers operating on complementary sources of input data. We have investigated this approach on a CSR task by combining conventional HMM processing with formant measurements. Our results show that this can significantly improve phone-level classification performance.
Bibliographic reference. Abberley, Dave / Green, Phil (1995): "Combining HMM processing and formant measurements in automatic speech recognition", In EUROSPEECH-1995, 2209-2212.