A transcript can get most of a sentence right and still miss its most important word. A drug name or an account number may be only a few tokens long. Our post-training work starts by aligning a sampled transcript against its reference. Insertions, substitutions and deletions then identify which words deserve credit.
We apply group-relative reinforcement learning to those aligned samples. Keeping the master updates in float32 allows us to accumulate small changes before exporting inference weights. The released v2 lineage has its own weights, decoder settings and evaluation on its model card.
The second intervention acts on the encoder. Preserved-key projection takes a donor update and measures where its input directions overlap the target domain and a set of clean examples. A regularized projection limits how much of the donor update reaches the protected directions. Clean-speech regression is checked empirically after each edit.
Campaign 4 targets 4.95% word error rate on the full Open ASR board. The benchmark page separates internal estimates from published submissions. For deployment, compare the decoder, language mix and audio conditions against your own workload.
Sources and experiments
Continue exploring
Hear the difference on your own audio.
Try an English recording without an account, inspect the transcript and vocal-expression annotations, and use the quickstart to bring the result into your application.