本論文提出以強健式演算法(論文中簡稱REST)訓練隱藏式條件隨機域(Hidden Conditional Random Fields,簡稱HCRF)華語/英語聲學模型,嘗試解決(1)混 雜語音語音辨認之抗雜訊問題和(2)混雜語音語音辨認之跨語系辨認錯誤問題。REST演算法可以提高HCRF模型對雜訊環境的辨識效能,接著透過鑑別式法則訓練HCRF模型,提升HCRF模型對語音模型的鑑別能力,並且大幅降低跨語言語音辨認之錯誤。根據一連串之實驗證明,基於HCRF語音模型之錯誤率平均值比傳統HMM降低約16.41%(Rover_2雜訊),並且跨語言語音辨認之錯誤大幅降低。
This thesis presents the robust training techniques for hidden conditional random fiels (HCRF)-based acoustic modeling of Mandarin/English mixed-lingual speech recognition. Two issues were dealt with: (1) mixed-lingual speech recognition against with noise effects and (2) cross-lingual errors in mixed-lingual speech recognition. We solved first issue with the REST algorithm and reduce the errors in second issue with a discriminative training algorithm combined by the REST algorithm(D-REST). The experimental results indicate that 16.4% averaged error rate reduction by the HCRF-based framework is achieved under ROVER_2 noise environment compared with the result by the traditional HMM approach. In additional, the cross-lingual error is improved significantly with the HCRF-framework in mixed-lingual speech recognition.