端点自由DP应用于语音识别中的时间轴校正

Free-regions DP used for time alignment on speech recognition

本文介绍了一种新的时间轴校正方法,克服了传统端点固定DP在语音起止区间存在的固有误差,有效的提高了语音识别率。

Opitmal methods to solve speech non-linear time lack fidelity is dynamic programming. By selecting an optimal time aligament function with some constrains, it makes the so-called goal function minium along with the optimal path. thatis, the distance between the temple parameter matrix and input speech parametermatrix is miniumlized. on this optimal path (and/or optimal time alignment function ) , optimal matching obtained. But fixed-regions DP makes the recognition rate decadent because of the detection error on start region and end region of speech. The propose of the free-regions DP could avoid the errors from the fixed-regions DP and detect the speech regions out by the DP matching results.