TurkSpeak lv3-MAST — trajectory-averaged Turkish Whisper large-v3
Recovery model from "Winning the Leaderboard, Losing the Distribution": a whisper-large-v3
LoRA fine-tune on an 875.2 h provenance-verified Turkish mixture, followed by uniform
trajectory averaging (MAST) over the run's saved checkpoints. The averaging stage costs
nothing at inference and dominates single-checkpoint selection.
Measured (corpus-WER %, faster-whisper CT2 fp16, beam 5, trnorm both sides, full sets)
| Model | FLEURS-TR (715) | medv3 (300) | tsc_heldout (4181) | tvoice_heldout (10058) |
|---|---|---|---|---|
| raw whisper-large-v3 | 8.26 | 4.68 | 20.28 | 6.32 |
| best single checkpoint | 8.34 | 3.88 | 17.06 | 4.81 |
| this model (MAST-avg) | 8.16 | 3.71 | 17.01 | 4.66 |
Beats the unmodified base on all four held-out axes simultaneously.
Intended use
Turkish ASR, including spontaneous/broadcast/medical speech. Trained with a shingle-based contamination firewall against all evaluation references.
Robustness leaderboard
https://huggingface.co/spaces/RsGoksel/turkish-asr-benchmark-stt