Synthetic-data fine-tuning is the other credible answer to domain vocabulary. Curious whether you re-benchmark the fine-tune when new base models ship?

Which open source STT models have you had success with for fine tuning?