Journal of the Acoustical Society of Japan (E) 15(2) 87-96 1994年
We describe a new real-time isolated word recognizer with improved user interface. The recognizer is designed for an Extension Number Guidance System which looks up and announces an extension number by telephone dialogue with users. To deal with telephone quality speech input which include noise and distortion during transmission over the telephone network, we developed feature extraction and a word detection algorithm. These techniques use wide band-pass filter outputs which are generally employed to decide whether speech is voiced or unvoiced. To achieve a friendly interface, the system can accept user input at any time by using an echo canceler and the new word detection algorithm. Finally, the recognizer is evaluated using a large telephone voice database consisting of more than 500 speakers.
「例文」と「例文の正しい統語解析木」から共起関係を蓄積し、解析への応用を試みている。今回、2つの手法で共起関係の蓄積・利用を行った。一つは、句の中心を成す語(ガバナ)に注目するもので、書換え規則の兄弟節点のガバナに共起関係があると仮定し、このガバナの並びを蓄積する。もう1つの方法は共起関係を人手によって書換え規則に記述する手法で、使用したい共起関係やその引数を自由に定義できる。どちらも正しい統語構造と解析結果を比較することで、出現した共起関係を正の事例・負の事例に分類する。そして、負の事例のみの蓄積例を解析のあい昧性解消に利用できる。蓄積例の増加に伴う解析性能の変化を両手法において測定した。蓄積した共起関係の利用により不適当な木の抑止や正しい木の選択に効果が見られた。This manuscript describes two methods which acquire lexical co-occurrences information and utilize it. These methods obtain co-occurrence relations from each example sentence and the corresponding right syntactic structure of the sentence. First of the two methods treats the governors appearing on sister nodes of the syntactic structure as a co-occurrence. On the other method, co-occurrence relationships are described manually in the rewrinting-rules. Both methods discriminate between the proper appearances of co-occurrence and wrong ones, using the right syntactic structure affixed to the sentence treated. The experiment is conducted for these methods, to observe the performances of the analysis using the stored co-occurrences data.