Speech Retrieval under Limited Resources and Open Domain Conditions

Speech Retrieval focuses on retrieving a segment of speech from a speech corpus correspond to a given query. A standard Speech Retrieval system usually composed by two systems, the Automatic Speech Recognition (ASR) system and the Information Retrieval (IR) system. The ASR system transcribes the speech and represents the transcript in different formats. The transcript is then indexed and searched by the IR system. As a result, Speech Retrieval is sensitive to the ASR, since IR system is depend on the transcript generated by ASR. The current challenge in Speech Retrieval is the limitation of ASR performance under certain conditions. Two such conditions are Limited Resources and Open Domain. Under Limited resources condition, the training data is not sufficient for creating a robust ASR system. A good example for this condition is to perform Speech Retrieval on limited resources languages such as Tagalog or Assamese. On the other hand, under Open Domain condition, the recorded speech varies in many perspectives. The high diversity of recorded speech limits the performance of a single ASR system. A good example for this condition is to perform Speech Retrieval on YouTube videos, such as online lectures. We believe Speech Retrieval under these conditions can be significantly improved from different approaches. The first one is to apply extra information, such as contexts from conversation. A context includes the other words in the same utterance or conversation. The second approach is to refine the existing IR system, by using better IR search strategy for Speech Retrieval. We analysis existing IR system and present a better search strategy, which is based on the diversity of current approaches. We have investigated how to integrate these two approaches to Speech Retrieval, and determined that the approaches can achieve improvement on Spoken Term Detection (STD) under the limited resources condition. The resulting system has been shown to be effective on multiple languages, implying that the improvement is language independent. Based on our positive result regarding the limited resources condition, we propose to extend existing approaches and develop new techniques for better Speech Retrieval under the open domain condition. We propose a new Speech Retrieval task called Spoken Snippet Retrieval (SSR), which retrieve a moderate size of speech from the speech collection with just enough context. The retrieved snippet is easier for user to listen through compare to the spoken document retrieved by Spoken Document Retrieval (SDR) systems, which has average length of 3 minutes. The snippet is more comprehensible compare to the term location detected by STD systems, since the context are given. The main contribution for the thesis is to complete SSR on the open domain data, which we believe is doing the adequate retrieval on the appropriate data.