Alex Waibel
Connectionist Large Vocabulary Word
Recognition
Abstract:In this paper we discuss recent research aimed at extending connectionist models to large vocabulary
word recognition. We describe the problem and the properties a successful large vocabulary system
must satisfy. While a number of different methods and ideas have recently been proposed and are under
investigation we will limit the discussion here to only one particular hybrid approach, i.e., the combination
of TDNN-based phoneme recognition/spotting nets with classical techniques for sequence management
(such as DP-matching and HMMs). We implement a baseline system using the best recent TDNN
phoneme spotting nets and evaluate its performance over a 500 and a 2620 word vocabulary, not used
during training. In both of these vocabulary independent evaluations high word recognition rates were
measured despite the large vocabulary size and perplexity in this otherwise unconstrained task. We then
describe, exploratory experiments that illustrate the importance and effectiveness of integral training, i.e.,
the integration of sequential management or alignment with phoneme network optimization. Significant
performance improvements were found with this technique over a system using decoupled training and
alignement. Finally, we offer a critical discussion and observations for further research.