任意語彙の追加登録可能な単語音声認識システム

抄録

In this paper, we describe a new word registration method on a speech recognition system which can treat user's specific words for a speech input based application. Since the speech application can be used on various equipments or environments, it is necessary for a speech recognition system on some applications or portable equipments to have a capability that a user can add a new word only by speaking its word, that is, without teaching its spelling or pronunciation to the system.
We use a subword HMM-based speaker-independent continuous speech recognizer for estimating pronunciations for new words. The estimated pronunciations are also used for an unsupervised speaker adaptation process in order to improve the recognition accuracy. The effectiveness of this approach is first investigated by performing the experiment of isolated word speech recognition both on existing lexicon and on estimated lexicon, along with the speaker adaptation processing and conventional DP-based approach. The effect of adding partial vocabulary words by speaking is also investigated. We confirmed that the method works well and this approach is suitable for keyboard-less speech recognizers.