The Effects of Automatic Speech Recognition Quality on Human Transcription Latency

Transcription makes speech accessible to deaf and hard of hearing people. This conversion of speech to text is still done manually by humans, despite high cost, because the quality of automated speech recognition (ASR) is still too low in real-world settings. Manual conversion can require more than 5 times the original audio time, which also introduces significant latency. Giving transcriptionists ASR output as a starting point seems like a reasonable approach to making humans more efficient and thereby reducing this cost, but the effectiveness of this approach is clearly related to the quality of the speech recognition output. At high error rates, fixing inaccurate speech recognition output may take longer than producing the transcription from scratch, and transcriptionists may not realize when transcription output is too inaccurate to be useful. In this paper, we empirically explore how the latency of transcriptions created by participants recruited on Amazon Mechanical Turk vary based on the accuracy of speech recognition output. We present results from 2 studies which indicate that starting with the ASR output is worse unless it is sufficiently accurate (Word Error Rate of under 30%).