Spoken content retrieval (SCR) is a process to provide a user with spoken documents in which the user is potentially interested. Unlike textual documents, searching through speech is not trivial due to its representation. Generally, automatic speech recognition (ASR) is used to transcribe spoken content such as user-generated videos and podcast episodes into transcripts before search operations are performed. Despite recent improvements in ASR, transcription errors can still be present in automatic transcripts. This is in particular when ASR is applied to out-of-domain data or speech with background noise.
This thesis explores improvement of ASR systems and search models for enhanced SCR on user-generated spoken content. There are three topics explored in this thesis. Firstly, the use of multimodal signals for ASR is investigated. This is motivated to integrate background contexts of spoken content into ASR. Integration of visual signals and document metadata into ASR is hypothesised to produce transcripts more aligned to background contexts of speech. Secondly, the use of semi-supervised training and content genre information from metadata are exploited for ASR. This approach is motivated to mitigate the transcription errors caused by recognition of out-of-domain speech. Thirdly, the use of neural models and the model extension using N-best ASR transcripts are investigated. Using ASR N-best transcripts instead of 1-best for search models is motivated because "key terms" missed in 1-best can be present in the N-best transcripts.
A series of experiments are conducted to examine those approaches to improvement of ASR systems and search models. The findings suggest that semi-supervised training bring practical improvement of ASR systems for SCR and the use of neural ranking models in particular with N-best transcripts improve the result of known-item search over the baseline BM25 model.