使用在不連續語音辨認系統上的語音切割方法

Abstract

在大字彙的語音辦認系統中，如何克服龐大計算量且仍維持高辦認率是語音辦認系統實用化中一個主要的難題。針對這個問題，有人提出預先處理器的觀念，其主要目的是先去除大量與測試音不相似的字彙以減輕系統的運算負荷。本文即是提出一個新的預先處理器，它不但具有比現存架構更低的計算能力，而且辦認率很高。 The major difficulty in large size vocabulary speech recognition system is its high computation cost to achieve good recognition performance. Recently, Pan et al, have proposed a word-based preprocessor to screen out all word candidates which are unlikely to match the test token, this preprocessor greatly alleviates the computation load of the recognition system. In this paper, we proposed a new preprocessor which has lower computations cost than the conventional preprocessor, but still has very high recognitions rate.