This paper describes the ASR system submitted by FBK to the CHiME-4 challenge for the single channel track. The proposed solution employs multiple subsystems, whose DNNs are trained with different training criteria and strategies (i.e. diverse training material, with and without batch normalization). A ``self" adaptation of acoustic models is applied to each subsystem, relying on a blind estimate of the accuracy of automatic transcriptions. This adaptation, performed in a batch fashion over the entire evaluation set, significantly improves the performance of each subsystem. The final output is obtained by combining the multiple transcriptions through ROVER, which provides a further improvement, reducing the average WER on the evaluation set from 22.3% to 16.5%.
The FBK system for the CHiME-4 challenge
Matassoni, Marco;Ravanelli, Mirco;Jalalvand, Shahab;Brutti, Alessio;Falavigna, Giuseppe Daniele
2016-01-01
Abstract
This paper describes the ASR system submitted by FBK to the CHiME-4 challenge for the single channel track. The proposed solution employs multiple subsystems, whose DNNs are trained with different training criteria and strategies (i.e. diverse training material, with and without batch normalization). A ``self" adaptation of acoustic models is applied to each subsystem, relying on a blind estimate of the accuracy of automatic transcriptions. This adaptation, performed in a batch fashion over the entire evaluation set, significantly improves the performance of each subsystem. The final output is obtained by combining the multiple transcriptions through ROVER, which provides a further improvement, reducing the average WER on the evaluation set from 22.3% to 16.5%.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
