Escuela Politécnica Superior

Universidad Autónoma de Madrid

Google Research Award

Google Research Award to Joaquin Gonzalez-Rodriguez from ATVS-UAM.
His project proposal entitled "Exploiting prior knowledge for robust recognition and indexing of audio information sources" has been funded by Google Research with 40.000 US$ for a two year research period starting april 2009.

Title: Exploiting prior knowledge for robust recognition and indexing of audio information sources Research Abstract and Goals

Current audio indexing applications and speech, speaker and language recognition technologies rely mostly on the information available in the input signal. As a result, all those techniques are highly influenced by the unknown conditions of the test recording, as session mismatch (channel, type of speech, emotions ...), noisy and/or reverberant recordings, or data scarcity. However, recent developments on factor analysis techniques allow to introduce very strong priors, estimated from huge amounts of data, drastically reducing the amount of free parameters to be estimated to a very limited subspace. Prior subspaces of both speaker and session variability between different recordings have been successfully exploited in last NIST 2008 Speaker Recognition Evaluation, but other applications fully fitted to the Google environment, as speaker diarisation and audio indexing or adaptation to mismatched conditions can be extraordinarily benefited from different sources of prior knowledge. The goals of the research proposal are to obtain significant improvements relative to present performance introducing different sources of prior knowledge in two sample applications, namely multi-environment speaker diarisation (broadcast programs, conversational telephone speech, meetings transcription ...) and speaker recognition from severely mismatched channels with scarce test data. Extensions to other information sources in audio indexing, and the use of factor analysis in speech recognition will also be addressed.