In this paper we present a study of automatic speech recognition systems using context-dependent phonemes and graphemes as sub-word units based on the conventional HMM/GMM system as well as tandem system. Experimental studies conducted on three different continuous speech recognition tasks show that systems using only context-dependent graphemes can yield competitive performance on small to medium vocabulary tasks when compared to a context-dependent phoneme-based automatic speech recognition system. In particular, we demonstrate the utility of tandem features that use an MLP trained to estimate phoneme posterior probabilities in improving grapheme based recognition system performance by implicitly incorporating phonemic knowledge into the system without having to define a phonetically transcribed lexicon.
展开▼
