A novel system-independent method for estimating the word error rate (WER) of automatic speech recognition (ASR) transcripts, which outperforms previous ASR system-dependent approaches on out-of-domain data.
