Whisper-Base-CV11 — 11 languages fine-tune, standard architecture (scientific control)
A plain openai/whisper-base (unmodified architecture) fine-tuned on the 11 languages of the CC0
Whispered corpus. This is the
matched scientific control for burakaydinofficial/whisper-base-mla-cv11 —
trained identically, minus the MHA→MLA conversion — published so the MLA conversion cost is
independently reproducible. No custom code: loads directly in transformers, and — being a plain unmodified Whisper — is convertible for faster-whisper / CTranslate2 / whisper.cpp via their standard converters.
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor
model = AutoModelForSpeechSeq2Seq.from_pretrained("burakaydinofficial/whisper-base-cv11") # no trust_remote_code
processor = AutoProcessor.from_pretrained("burakaydinofficial/whisper-base-cv11")
Reproduce the conversion cost
Evaluate this control and whisper-base-mla-cv11 on CommonVoice-17 (scripts/validate.py in the
code repo); the per-language difference is the
conversion cost reported on the MLA card and in docs/results/.
Results (CommonVoice-17 test, greedy, Whisper normalization + Arabic folding; CER for th/zh/ja)
| Lang | this control |
|---|---|
| en | 20.1 WER |
| de | 29.3 WER |
| es | 18.1 WER |
| fr | 31.7 WER |
| ru | 27.4 WER |
| tr | 37.8 WER |
| cy | 77.8 WER |
| ar | 56.2 WER |
| ka | 99.1 WER |
| th | 22.1 CER |
Checkpoint note: this is the FINAL-step checkpoint; the dev-selected checkpoint that produced the reference numbers above was not retained, so an independent re-evaluation may differ by a few tenths of a WER point. The MLA card's reported cost used the dev-selected control.
Encoder frozen during fine-tuning; 15,000 steps, warmup+cosine, fp16. Read-speech domain (CommonVoice + FLEURS-validated). "Compression cost" does not apply to this unconverted control.