Whisper-Small-24Lang — 24 languages fine-tune, standard architecture (scientific control)
A plain openai/whisper-small (unmodified architecture) fine-tuned on the 24 languages of the CC0
Whispered corpus. This is the
matched scientific control for burakaydinofficial/whisper-small-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-small-24lang") # no trust_remote_code
processor = AutoProcessor.from_pretrained("burakaydinofficial/whisper-small-24lang")
Reproduce the conversion cost
Evaluate this control and whisper-small-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)
| Lang | this control |
|---|---|
| en | 12.2 WER |
| de | 14.9 WER |
| es | 10.0 WER |
| fr | 19.1 WER |
| it | 15.4 WER |
| pt | 14.8 WER |
| ru | 14.0 WER |
| nl | 13.2 WER |
| pl | 17.3 WER |
| id | 19.5 WER |
| tr | 21.1 WER |
| hi | 23.2 WER |
| ms | 20.1 WER |
| sv-SE | 21.9 WER |
| th | 12.1 CER |
| zh-CN | 15.6 CER |
| cs | 27.3 WER |
| vi | 28.4 WER |
| fi | 28.5 WER |
| el | 30.5 WER |
| da | 32.3 WER |
| ja | 23.3 CER |
| nn-NO | 43.1 WER |
| ko | 44.2 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.