Whisper-Tiny-24Lang — 24 languages fine-tune, standard architecture (scientific control)

A plain openai/whisper-tiny (unmodified architecture) fine-tuned on the 24 languages of the CC0 Whispered corpus. This is the matched scientific control for burakaydinofficial/whisper-tiny-mla-24lang — 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-tiny-24lang")   # no trust_remote_code
processor = AutoProcessor.from_pretrained("burakaydinofficial/whisper-tiny-24lang")

Reproduce the conversion cost

Evaluate this control and whisper-tiny-mla-24lang 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)

Reconstructed numbers. This control's per-language WER/CER in the table are reconstructed as (the MLA model's absolute score − the measured paired conversion cost), not from a separate evaluation of this checkpoint; an independent re-eval may differ by a few tenths. The paired conversion cost is the directly measured quantity.

Lang	this control
en	28.1 WER
de	42.1 WER
es	29.5 WER
fr	44.8 WER
it	43.8 WER
pt	43.1 WER
ru	40.6 WER
nl	38.8 WER
pl	47.3 WER
id	53.0 WER
tr	52.1 WER
hi	45.9 WER
ms	49.5 WER
sv-SE	57.9 WER
th	32.0 CER
zh-CN	34.7 CER
cs	67.2 WER
vi	54.4 WER
fi	66.9 WER
el	63.7 WER
da	73.7 WER
ja	48.8 CER
nn-NO	93.1 WER
ko	71.3 WER

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.