不特定話者連続音声認識のための規則による音素識別

概要

論文の詳細を見る
The authors propose a method of phoneme identification by using the rules which are constructed based on the knowledge of speech. In the proposed phoneme identification system, the following are considered in order to identify the phonemes in continuously spoken speech by several speakers. (1) Feature parameters are extracted by using some kind of windows with several time period. (2) Many kinds of parameters, formant frequencies extracted from LPC spectra, number of zero-crossings, vocal tract areas estimated with PARCOR coefficients and so on, are used for phoneme identification. (3) Component ratios and time variations of feature parameters are maimly used for phoneme identification. (4) Some suitable feature parameters are used to identify a phoneme. Our phoneme identification system can identify correctly 77% of phonemes in sentences read aloud, and 70% of them in spoken dialogue.
山梨大学の論文