A Method for Dealing with Out-of-Vocabulary Items in Speech Recognition

In this paper, we describe a method to enhance the readability of out-of-vocabulary items (OOVs) in the textual output in a large vocabulary continuous speech recognition system. The basic idea is to indicate uncertain words in the transcriptions and replace them with phoneme recognition results that are post-processed using a phoneme-to-grapheme (P2G) converter. We concentrate on the final step, P2G conversion: we show that the phoneme recognition results can be reasonably reliably transcribed orthographically using machine learning techniques. More specifically, (i) we present experimental results of a machine learning approach to P2G conversion, and compare these results with an estimation of the upper and lower baseline performance, (ii) we give an error analysis and list some examples of the converter’s output, (iii) we investigate spelling correction as post-processing of the orthographic transcriptions, and (iv) we report on the interaction of the P2G converter with a speech recognizer.