New-word addition and adaptation in a stochastic explicit-segment speech recognition system

Abstract

The authors extend on automatic procedure for the addition of new words to a speech recognition system to include alternative pronunciations for the new words. They investigate methods for adaptation to new words after these are added to the system. For adaptation, the goal was the improvement of the accuracy of the system on the new words, using only a limited amount of speech data. All the experiments are performed within the stochastic explicit-segment speech recognition system. The authors evaluated 25 isolated city names from a speech corpus, CITRON, collected from real users over the telephone network. For this task, improvement in accuracy is shown from a 34% error rate, when trained on the NTIMIT database alone, to 8% after adapting to 30 tokens, on average, from each new word.<>