ISCA Archive ICSLP 1992
ISCA Archive ICSLP 1992

Detection of unknown words and automatic estimation of their transcriptions in continuous speech recognition

Itou Katunobu, Hayamizu Satoru, Tanaka Hozumi

Current continuous speech recognition systems are designed to recognize words within a vocabulary. In order to make speech recognition systems more flexible, convenient and robust, they should be able to process unknown words. This paper introduces a new technique for processing unknown words in a continuous speech recognition system. Two types of processing, one with stochastic language models without any other linguistic knowledge and the other with a dictionary and a grammar, are dynamically controlled to detect and transcribe the unknown words automatically. We tested this method by speaker independent continuous speech recognition experiments using a task with 113 word vocabulary, with bunsetu perplexity 8.2. Preliminary results showed a detection rate for 75% of the unknown words, with a false alarm rate of 11% and a phone recognition rate of 51% for the unknown words detected.