Loading the leaderboard…
Loading the leaderboard…
Model profile · Multilingual open
Best published zero-shot CS-WER
WER family in %, lower is better · MTR higher is better · RTFx = audio s ÷ wall s, higher is faster · normalizer v0.1
The Hard split holds 1.38 h / 658 utterances of rare and unseen code-switched terms — the generalization leaderboard. Published results cover Test only; the platform pilot re-runs every model on Test + Hard, and this gap closes when its run lands.
Per-topic WER is unavailable for this row — paper imports report corpus-level scores only. Platform-verified runs break WER down across the five ViMedCSS domains:
WER is comparable only within a normalizer version · bootstrap 95% CIs on headline metrics.