READ metric cuts ASR errors by 20% without reference transcriptions

A team of researchers from Shanghai Jiao Tong University (Zhihan Li, Hankun Wang, et al.) has introduced READ, a novel metric for evaluating automatic speech recognition (ASR) hypotheses without relying on reference transcriptions. Traditional ASR evaluation requires ground-truth text, but READ instead uses a pretrained auto-regressive text-to-speech (TTS) model to compute the conditional likelihood of speech tokens given a text hypothesis. This allows it to measure fine-grained acoustic discrepancy directly from the speech signal, making it a truly reference-free approach. The method requires no additional training and can be applied for hypothesis refinement.

In experiments submitted to Interspeech 2026, READ showed strong correlation with specific recognition errors and improved ASR outputs by up to 20% relative error rate reduction. The gains were particularly pronounced under noisy conditions, where traditional confidence-based methods often falter. By grounding evaluation in acoustic evidence rather than language model predictions, READ offers a more robust way to assess and improve ASR systems in real-world environments where clean references are unavailable.