MGB-3 but system: Low-resource ASR on Egyptian YouTube data

This paper presents a series of experiments we performed duringour work on the MGB-3 evaluations. We both describethe submitted system, as well as the post-evaluation analysis.Our initial BLSTM-HMM system was trained on 250 hoursof MGB-2 data (Al-Jazeera), it was adapted with 5 hours ofEgyptian data (YouTube). We included such techniques asdiarization, n-gram language model adaptation, speed perturbationof the adaptation data, and the use of all 4 correctreferences. The 4 references were either used for supervisionwith a confusion network, or we included each sentence 4xwith the transcripts from all the annotators. Then, it was alsohelpful to blend the augmented MGB-3 adaptation data with15 hours of MGB-2 data. Although we did not rank with oursingle system among the best teams in the evaluations, we believethat our analysis will be highly interesting not only forthe other MGB-3 challenge participants.