The aim of this work is to improve speech recognition performance by forming "phonological concepts". In order to improve speech recognition performance, the phoneme models or phone models of the system need to satisfy following two properties: l)preciseness and 2)ro-bustness. These two properties usually trade off against each other in traditional stochastic models. In order to satisfy the both simultaneously, we propose to simulates situations of a human infant learning a language. We call this "phonological concept formation", which is a task of acquiring knowledge of phonological system from spoken word samples without using any transcriptions except for the identification of each word in a lexicon. This knowledge includes what is the set of the whole phonemes, how the acoustic phonetic features of each phoneme are described, and how they are appropriately discriminated. The basis of this idea is specifying essen-tial situations of speech communication instead of pro-viding all encompassing universal samples of a spoken language. Based on this framework, this paper discusses about the method for generating both precise and robust models.
|出版ステータス||Published - 1994|
|イベント||3rd International Conference on Spoken Language Processing, ICSLP 1994 - Yokohama, Japan|
継続期間: 1994 9月 18 → 1994 9月 22
|Conference||3rd International Conference on Spoken Language Processing, ICSLP 1994|
|Period||94/9/18 → 94/9/22|
ASJC Scopus subject areas