Multi-stream ASR trained with heterogeneous reverberant environments

摘要: A common problem with automatic speech recognition (ASR) systems is that the performance degrades when it presented from a different acoustic environment than one used during training. An important cause feature distribution to which ASR system trained no longer matches of new environment. Reverberant environments can be especially harmful. We test multi-stream in constituent streams are each separate environments. When training modeling stages separately clean data and heavily reverberated data, we find combined improve unseen data.