Database adaptation for ASR in cross-environmental conditions in the SPEECON project

As part of the SPEECON corpora collection project, a software toolbox for transforming speech recordings made in a quiet environment with a close-talk microphone into far-talk noisy recordings has been developed. The toolbox allows speech recognizers to be trained for new acoustic environments without requiring an extensive data collection effort. This communication complements a previous article in which the adaptation toolbox was described in details and preliminary experimental results were presented. Detailed experimental results on a database specifically collected for testing purposes show the performance improvements that can be obtained with the database adaptation toolbox in various far-talk and noisy conditions. The Hebrew corpus collected for SPEECON is also used to assess how close a recognizer trained on simulated data can get to a recognizer trained on real far-talk noisy data.