使用元學習技術於語碼轉換語音辨識之初步研究

A Preliminary Study on Leveraging Meta Learning Technique for Code-switching Speech Recognition

"語碼轉換（Code-switching）的語音辨識近年來愈來愈受到研究學者的重視，雖然人們在日常生活中使用語碼轉換的情形逐漸增加，但是可以用來訓練語碼轉換之語音辨識器的語音資料，相較於主流的單語言（如：英語或中文）語料，更是少之又少，這是因為語音語料的標註（Labeling）相當耗時費工。為了提升語碼轉換之語音辨識器的效能，在本研究中，我們提出利用近期逐漸興起的元學習（Meta Learning）方法，希望可以利用資料量較多的單語言語料，提升語碼轉換的語音辨識任務之成效。實驗中，我們採用SEAME（South East Asia Mandarin-English）資料集，透過元學習中的與模型無關之元學習法（Model-Agnostic Meta-Learning, MAML）進行語音辨識器訓練，實驗結果證明，模型能夠快速適應於最終的語碼轉換語音辨識任務。"

In recent years, code-switching speech recognition has become an important research topic. Code-switching in conversation speech is gradually increasing in our daily lives. However, compared with monolingual languages (e.g., English or Chinese), only a few resources can be obtained for training a code-switch speech recognizer. To mitigate the deficiency, in this paper, we propose a meta-learning approach for code-switching speech recognition. In other words, following the model-agnostic meta-learning (MAML) procedure, we first train the speech recognizer by using monolingual corpora, and then a fine-tune stage is performed to obtain the final code-switching speech recognizer by using code-switching data. We evaluate the proposed method on the SEAME (South East Asia Mandarin-English) dataset. A series of experiments show that the meta-learning method can improve the performance of the low-resource code-switching speech recognition task.