以狄式分佈為基礎之多語聲學模型拆分及合併

Multilingual Acoustic Model Splitting and Merging by Latent Dirichlet Allocation

在整合型多語辨識環境下，如何避免不同語言間的音標混淆是重要的課題之一。本篇論文針對相異語言間的聲學模型混淆，提出以狄式分佈（LDA, Latent Dirichlet Allocation）為基礎的聲學模型混淆度偵測。以三連音素聲學模型為基礎將聲學模型分裂後再使用潛藏狄式分佈選擇合併的聲學模型組後並進行合併，以解決因為不同語言發音變異所產生的聲學模型混淆度。本篇論文分為三個部分，第一部分為介紹發音屬性和語音事件，其為從訊號面尋找各種特徵並與特定聲學模型之間的相關性。第二部分為介紹狄式分佈（LDA, Latent Dirichlet Allocation）以及模型間混淆度的偵測方法，狄式分佈是一個階層式的數學模型，早期是由David M. Blei等人提出用來做為文件分類及文件產生所使用，但其架構相當適合應用在語音辨識、自然語言處理等領域。最後部分則是對本論文所提出之方法進行實驗驗證並分析。

To avoid the confusion of phonetic acoustic models between different languages is one of the most challenges in multilingual speech recognition. We proposed the method based on Latent Dirichlet Allocation to avoid the confusion of phonetic acoustic models between different languages. We split phonetic acoustic models based on tri-phone. And merging the group that selected by Latent Dirichlet Allocation Detector to solve pronunciation variants problems between different languages. This paper has three parts. First part is introduced the Pronunciation Event and Articulatory Features. Second part is about Latent Dirichlet Allocation and the acoustic model selecting method using Latent Dirichlet Allocation. Latent Dirichlet Allocation is a Hierarchical math model that proposed by David M. Blei at 2003. It is often used on documents classification and document generation. The structure of LDA is also suitable for speech recognition and nature language processing. The final is experiment result and verification the method we proposed.