Improving Searchability of Automatically Transcribed Lectures Through Dynamic Language Modelling
Year: 2012
Type: dissertation
Abstract: Recording university lectures through lecture capture systems is increasingly common. \nHowever, a single continuous audio recording is often unhelpful for users, who may wish \nto navigate quickly to a particular part of a lecture, or locate a specific lecture within a set \nof recordings. \nA transcript of the recording can enable faster navigation and searching. Automatic speech \nrecognition (ASR) technologies may be used to create automated transcripts, to avoid the \nsignificant time and cost involved in manual transcription. \nLow accuracy of ASR-generated transcripts may however limit their usefulness. In \nparticular, ASR systems optimized for general speech recognition may not recognize the \nmany technical or discipline-specific words occurring in university lectures. To improve \nthe usefulness of ASR transcripts for the purposes of information retrieval (search) and \nnavigating within recordings, the lexicon and language model used by the ASR engine may \nbe dynamically adapted for the topic of each lecture. \nA prototype is presented which uses the English Wikipedia as a semantically dense, large \nlanguage corpus to generate a custom lexicon and language model for each lecture from a \nsmall set of keywords. Two strategies for extracting a topic-specific subset of Wikipedia \narticles are investigated: a naïve crawler which follows all article links from a set of seed \narticles produced by a Wikipedia search from the initial keywords, and a refinement which \nfollows only links to articles sufficiently similar to the parent article. Pair-wise article \nsimilarity is computed from a pre-computed vector space model of Wikipedia article term \nscores generated using latent semantic indexing. \nThe CMU Sphinx4 ASR engine is used to generate transcripts from thirteen recorded \nlectures from Open Yale Courses, using the English HUB4 language model as a reference \nand the two topic-specific language models generated for each lecture from Wikipedia.
Author: Stephen Marquard
Institution: University of Cape Town
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Keywords: lecture transcripts, speech recognition
Open Access status: green
OpenAlex ID: W2507356423
