Humans versus machines: the case of conversational speech recognition

Recent work at Microsoft and IBM has pushed the accuracy of automatic speech recognition systems on conversational speech to a level that is on par with humans, when measured on historical government evaluation data sets. I will review the history of this task and then review a recent experiment at Microsoft to benchmark state-of-the-art recognition technology against human transcription performance. The second half of the talk will look at the question of how human and machine transcription accuracy differ quantitatively and qualitatively.