基於模型統計特性之詞語驗證臨界值估計方法

A Statistics Model Based Approach for Threshold Estimation of Utterance Verification

林政賢(Cheng-Hsien Lin) ； 邱祺添(Chi-Tien Chiu) ； 張森嘉(Sen-Chia Chang)

《電腦與通訊》 133期 (2010/06) Pp. 118-124

https://doi.org/10.29917/CCLTJ.201006.0020

摘要

本論文提出一種詞語驗證臨界值的估計方法。在一個包含詞語驗證功能的語音辨識系統裡，事先利用詞語驗證模型透過訓練語料將語音單元的詞語驗證分數統計特性儲存至語音單元分數統計資料庫。當辨識目標成立後，則利用該儲存的統計特性將辨識目標的詞語驗證分數分佈描繪出來，並預測合適的臨界值給詞語驗證模組使用。如此，不但可以免除使用者線上調整臨界值的麻煩；當某些裝置上無法提供臨界值調整時，也可省略不斷跟著辨識目標收集語音資料分析的工作。經由實驗發現，使用本方法所預測的詞語驗證臨界值與實際資料所分析所得相比較，在不同雜訊環境及不同錯誤拒絕率下均小於6%。

關鍵字

詞語驗證 ； 語音辨識 ； 錯誤拒絕率

並列摘要

A method for the estimation of threshold in utterance verification is presented here. In a speech recognition system with utterance verification module, the statistics of utterance verification scores are stored in advance before using the corresponding verification models and training corpus. The distribution of utterance verification scores with respect to recognition targets will be depicted using these statistics, and the threshold will be predicted for the verification module. By using this method, both the online adjustment of the threshold and the analysis of the collected data with respect to recognition targets are no longer needed. From the experimental results, the prediction error is always less than 6% under different noise environment and false rejection rate.

並列關鍵字

Utterance Verification ； Speech Recognition ； False Rejection Rate