Whisper Tiny Mla 24lang

Whisper-Tiny-MLA (24 languages) — on-device-tier MLA-Whisper, 62.5% smaller decode KV-cache

openai/whisper-tiny (39M) with decoder self-attention converted MHA→MLA (Whisper-MLA, arXiv:2603.00563), recovery-fine-tuned on a 24-language application set — the smallest, most deployable tier of the 24-language line.

from transformers import AutoModelForSpeechSeq2Seq
model = AutoModelForSpeechSeq2Seq.from_pretrained("burakaydinofficial/whisper-tiny-mla-24lang", trust_remote_code=True)  # transformers==4.46.x

Honest sizing note (read first)

Conversion cost GROWS as the model shrinks — measured across the 24-language family: small ≈+0.6 → base ≈+1.1 → tiny ≈+1.9 (approximate per-language medians). At the tiny tier you pay ≈+1.9 WER (median) for the 62.5% cache cut, and absolute quality is tiny-tier (much higher WER than small — that is the base model, not MLA). If quality is the priority prefer whisper-small-mla-24lang; this tier is for memory-constrained deployment where the cache cut matters most.

Results (CommonVoice-17 test — Malay from FLEURS, n≤1500/lang (ko 339, no 370, ms 749, vi 1274; rest 1500); cost = paired vs an identically-trained control; CER for th/zh/ja)

lang	WER%	CER%	conversion cost (WER pts)	sig
en	29.8	16.0	1.69	ns
de	44.1	17.0	1.97	sig
es	30.9	11.2	1.39	sig
fr	47.1	20.9	2.38	sig
it	45.1	13.8	1.32	sig
pt	44.7	17.3	1.65	ns
ru	43.9	14.6	3.28	sig
nl	40.5	16.1	1.76	sig
pl	49.6	16.0	2.33	sig
id	54.9	20.1	1.94	ns
tr	54.2	17.9	2.17	sig
hi	47.6	25.1	1.66	sig
ms	51.1	19.2	1.64	sig
sv-SE	60.5	24.9	2.57	sig
th	83.8	32.5	0.43	ns
zh-CN	55.6	34.5	-0.23	ns
cs	69.6	22.0	2.39	sig
vi	56.3	29.3	1.94	sig
fi	69.5	18.2	2.57	sig
el	69.0	28.2	5.37	sig
da	75.8	33.8	2.05	ns
ja	n/a	48.3	-0.50	ns
nn-NO	82.9	34.6	-10.16	ns
ko	78.0	44.3	6.61	sig

Thin-CommonVoice languages (Korean, Norwegian, Danish, Greek, Finnish, Czech, Vietnamese) are the weakest — a DATA limit, not an MLA one. At this tiny tier several are FLOOR cells (WER 69–83%: nn-NO 82.9, ko 78.0, da 75.8, cs 69.6, fi 69.5, el 69.0) — reported for transparency, NOT usable, exactly like Georgian at small sizes. At that error level the paired "conversion cost" is metric noise, not an MLA effect (e.g. nn-NO −10.16 just means one near-garbage arm hallucinated less on a 370-clip pool) — so the size-cost curve applies only to the languages where both arms are actually usable, not to these floor cells. Per-language cost also depends on the recovery mix: a language's cost here differs from the same language in whisper-tiny-mla-cv11 (the 11-language model) — expected (mix composition moves each language's low-rank fit), not irreproducibility. Malay was trained on FLEURS (no CV Malay exists; FLEURS is CC-BY-4.0; its eval-only license respected — audio used for training, never redistributed).

Matched control now published — verify the conversion cost yourself: burakaydinofficial/whisper-tiny-24lang (trained identically, minus the MHA→MLA conversion). Evaluate both with scripts/validate.py.

Limitations

• What the 62.5% is (cache scope): it is the decode self-attention KV-cache — the part that grows with output length and concurrency. The (larger, encoder-length ~1500-frame) cross-attention/encoder memory is NOT compressed, so single-stream total decode-memory savings are modest; the 62.5% cut compounds at output-length × batch concurrency, which is where it pays off.
• Runtime: requires trust_remote_code + transformers==4.46.x (no whisper.cpp/CT2/faster-whisper).
• Language coverage: covers these 24 languages; erodes Whisper's others.
• Decoding: greedy-decode evals (beam-5 adds ~1-2 WER on both arms; conversion cost unchanged — measured).
• Domain: consumer-mic read-speech.
• Training: 15k steps, warmup+cosine, encoder frozen both arms, dev-selected, bf16 (weights released as fp16).