BCN2BRNO: ASR System Fusion for Albayzin 2020 Speech to Text Challenge

This paper describes the joint effort of BUT and Telefónica Researchon the development of Automatic Speech Recognitionsystems for the Albayzin 2020 Challenge. We compare approachesbased on either hybrid or end-to-end models. In hybridmodelling, we explore the impact of a SpecAugment layeron performance. For end-to-end modelling, we used a convolutionalneural network with gated linear units (GLUs). Theperformance of such model is also evaluated with an additionaln-gram language model to improve word error rates. We furtherinspect source separation methods to extract speech fromnoisy environments (i.e. TV shows). More precisely, we assessthe effect of using a neural-based music separator named Demucs.A fusion of our best systems achieved 23.33% WER inofficial Albayzin 2020 evaluations. Aside from techniques usedin our final submitted systems, we also describe our efforts inretrieving high-quality transcripts for training.