A series of experiments was performed to evaluate the effectiveness of several pre-processing techniques in speaker-independent recognition of Putonghua finals using hidden Markov models. The evaluated techniques include phonemic labeling, vector quantization and a hybrid approach derived from the two. Whilst previous research (ref 1) showed that phonemic labeling was fast and effective in labeling Putonghua vowels, its performance was found to be inferior to that of vector quantization. However, as the training and recognition of vector quantization demands an excessive processing time, the major speed advantage of hidden Markov models is greatly offset. On the other hand, preliminary results showed that the hybrid approach made a promising compromise between speed and performance.