RALI-OLST Weekly Talks

Investigative Audio Keyword Search System

Vishwa Gupta (Vishwa (point) Gupta <at> crim (point) ca)

CRIM

Wednesday 8 April 2009 at 11:30 AM

Salle 3195, Pavillon André-Aisenstadt


We report results on a spoken audio keyword search system for investigative purposes. Investigative keyword search requires high recall rate at zero false alarm rate. The requirement for the system we have installed at an energy utility is over 93% recall of relevant calls with no false alarms. This high level of recall is achieved by automatically generating keyword spots with high false alarm rates and then listening to the spots to eliminate false alarms. For high recall rate, we train the acoustic models from over 50 hours of audio, and adapt the acoustic models to each call. At 300% false alarm rate, the best recall rate is 73.8%. This corresponds to less than 2% miss rate for relevant conversations. It turns out that in general, relevant conversations are calls with 3 or more occurrences of keywords. The innovation that has a major impact on the recall rate is the boosting of the likelihoods of the keywords in the language model. This boosting improves the keyword recall rate from 54.0% to 73.8%.


Follow this link to subscribe to future RALI-OLST announcements.
http://rali.iro.umontreal.ca/rali/?q=fr/node/1631

See all the weekly talks for the year:

1991 1992 1993 1994 1995 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024