An improved two-stage mixed language model approach for handling out-of-vocabulary words in large vocabulary continuous speech recognition

This paper presents a two-stage mixed language model technique for detecting and recognizing words that are not included in the vocabulary of a large vocabulary continuous speech recognition system. The main idea is to spot the out-of-vocabulary words and to produce a transcription for these words in terms of subword units with the help of a mixed word/subword language model in the first stage, and to convert the subword transcriptions to word hypotheses by means of a look-up table in the second stage. The performance of the proposed approach is compared to that of the state-of-the-art hybrid method reported in the literature, both on in-domain and on out-of-domain Dutch spoken material, where the term 'domain' refers to the ensemble of topics that were covered in the material from which the lexicon and language model were retrieved. It turns out that the proposed approach is at least equally effective as a hybrid approach when it comes to recognizing in-domain material, and significantly more effective when applied to out-of-domain data. This proves that the proposed approach is easily adaptable to new domains and to new words (e.g. proper names) in the same domain. On the out-of-domain recognition task, the word error rate could be reduced by 12% relative over a baseline system incorporating a 100k word vocabulary and a basic garbage OOV word model.