Generating Automatic Keywords for Conversational Speech ASR Transcripts

While a plethora of conversational speech has been recorded and archived for over a century, it has not been easily accessible due to many technical challenges vs. text and rehearsed speech to be addressed before conversational archives can be effectively searched and used. In this paper, we describe two language modeling methods for automatically assigning keywords to automatic speech recognition (ASR) transcripts, to benefit search and browsing of conversational speech archives. Experiments performed with the English CLEF CL-SR MALACH collection of oral history interviews. In comparison to a prior baseline generating 20 keywords per conversation segment, we use 1/20th the training data yet improve Recall@20 in matching manual keywords. However, while indexing of manual keywords yields improved search accuracy, indexing automatic keywords (ours or the baseline) fails to improve search accuracy, evidencing the need for additional research.