M. Akbacak, L. Burget, W. Wang, and J. van Hout, “Rich system combination for keyword spotting in noisy and acoustically heterogenous audio streams,” in Proc. 38th International Conference on Acoustics, Speech, and Signal Processing (ICASSP), 2013, pp. 8267–8271.
Abstract
We address the problem of retrieving spoken information from noisy and heterogeneous audio archives using a rich system combination with a diverse set of noise-robust modules and audio characterization. Audio search applications so far have focused on constrained domains or genres and not-so-noisy and heterogeneous acoustic or channel conditions. In this paper, our focus is to improve the accuracy of a keyword spotting spotting system in a highly degraded and diverse channel conditions by employing multiple recognition systems in parallel with different robust frontends and modeling choices, as well as different representations during audio indexing and search (words vs. subword units). Then, after aligning keyword hits from different systems, we employ system combination at the score level using a logistic-regression-based classifier. When available, side information (such as signal-to-noise ratio or the output of an acoustic condition identification module) is used to guide system combination that is trained on separate held-out data. Lattice-based indexing and search is used in all keyword spotting systems. We present improvements in probability-miss at a fixed probability-false-alarm by employing our proposed rich system combination approach on DARPA Robust Audio Transcription (RATS) Phase-I evaluation data that contains highly degraded channel recordings (SNR as low as 0 dB) and different channel characteristics.