Spoken Document Retrieval: Sub-sequence DTW Framework and Variants

Abstract

We address the problem of spoken document retrieval (alternately termed content-based audio-search and retrieval), which involves searching a large spoken document or database for a specific spoken query. We formulate the search within the sub-sequence DTW (SS-DTW) framework proposed earlier in literature, adapted here to work on acoustic feature representation of the database and spoken query term. Further, we propose several variants within this framework, such as (i) path-length based score normalization, (ii) clustered quantization of acoustic feature vectors for fast search and retrieval with invariant performances and, (iii) phonetic representation of the database and spoken query term, derived from ground-truth annotation as well as HMM based continuous phoneme recognition. We characterize the performance of the proposed framework, algorithms and variants in terms of ROC curves, EER and time-complexity and present results using the TIMIT database with annotated spoken sentences from 400 speakers.

A. Seshasayee—Author carried out this work as Research Associate at PESIT-BSC, Bangalore.