Re-evaluation of LVQ-HMM hybrid algorithm

The LVQ-HMM hybrid algorithm was one of the first algorithms proposed in a recent approach aiming to integrate a highly discriminative artificial neural network-based classifier with an HMM capable of representing temporal structure effectively. The high phoneme classification capability of LVQ-HMM has already been demonstrated. However, the performance of LVQ-HMM has been less striking in more difficult, large scale speech recognition situations, making evaluation of the algorithm controversial and suggesting a more detailed investigation of the properties of the algorithm in such situations. This technical report is thus devoted to re-evaluation of the hybrid algo rithm, evaluated for word and phrase recognition tasks. Specifically, recognition ex perimentsare conducted under rather difficult, speaker-independent and large-vocabu lary conditions. Our recognizer uses a phoneme-based strategy; in particular, the predictive LR-parser is incorporated for efficient recognition. Experimental results alone are unfortunately insufficient to cease the controversy. However, possible contribu tions and aspects of the algorithm needing further improvement are brought to light.