whisper-medium-dora-mix6

DoRA-adapted openai/whisper-medium on 6 low-resource FLEURS languages: Vietnamese, Hausa, Lingala, Tamil, Maltese, Javanese. Decoder-only adaptation — encoder is unchanged from base, so encoder-side interventions like token merging behave identically.

Backs §5.2 of "Token Merging for Multilingual Speech Recognition: A Systematic Study Across Model Scale and Fine-Tuning" (Dylan Luke Holyoak — ICNLSP 2026 oral).

• Code: https://github.com/chironk/token-merging-multilingual-asr

Usage

from transformers import WhisperForConditionalGeneration, WhisperProcessor
from peft import PeftModel

processor = WhisperProcessor.from_pretrained(
    "openai/whisper-medium",
    revision="abdf7c39ab9d0397620ccaea8974cc764cd0953e",
)
base = WhisperForConditionalGeneration.from_pretrained(
    "openai/whisper-medium",
    revision="abdf7c39ab9d0397620ccaea8974cc764cd0953e",
    torch_dtype="float32",
)
model = PeftModel.from_pretrained(
    base,
    "dylan01163104/whisper-medium-dora-mix6",
    revision="ad9144916cf661ea2ef462ad273077343c3d803d",
)
model.eval()
# ...standard Whisper inference from here

Training

• Base model: openai/whisper-medium revision abdf7c39ab9d0397620ccaea8974cc764cd0953e
• Method: DoRA (Weight-Decomposed LoRA), peft==0.19.1
• Adapter target: decoder self_attn + encoder_attn q_proj, k_proj, v_proj, out_proj (all 24 layers). Encoder frozen.
• Rank: 32
• Alpha: 64
• Data: FLEURS revision d7c758a6dceecd54a98cac43404d3d576e721f07, 6 target languages + 10% English anchor. Temperature-sampling T=0.5.
• Steps: 2000 · LR: 1e-5 · Warmup: 200 steps · Batch: 8 × 2 grad_accum × 2 GPUs = effective 32
• Hardware: 2× consumer GPU, DDP via torchrun
• Seed: 42

Full training script: tmm_asr/train/dora.py.

Results (WER, mean of 264 FLEURS test clips)

Language	Base whisper-medium	+ DoRA-mix6	Δ
Lingala	83.23	52.17	−31.05
Javanese	67.23	51.24	−15.99
Hausa	88.75	75.31	−13.44
Maltese	82.19	75.92	−6.27
Vietnamese	14.29	14.10	−0.19
Tamil	35.67	36.00	+0.33

These are project benchmark results, not Hugging Face benchmark results. Both columns use the same first 264 examples from each pinned FLEURS test split and the repository's shared decoding and text-normalization pipeline. The base column is evaluated without the adapter; the DoRA column loads adapter checkpoint step 2000 at the pinned revision shown above.

Evaluation code and frozen result tables:

• Evaluation protocol and reproduction commands
• Base-model results
• DoRA results

Token merging on top of this adapter costs at most +0.59 pp WER at TRR = 0.40 (Lingala). See paper §5.2.

Generalisation to held-out languages

Applied to 10 languages the adapter was NOT trained on (English, French, German, Spanish, Thai, Swahili, Afrikaans, Icelandic, Welsh, Kazakh):

• Anchors preserved (+0.02 pp mean drift from base)
• Untrained mid/low-res drift +1.77 pp mean, max +3.78 pp (Kazakh)
• Merging still cheap: mean +0.23 pp, max +0.97 pp (Icelandic)

Full held-out table in paper Appendix Table 5; see §5.2 for the discussion.

Limitations

• Encoder unchanged, so any encoder-language-code mismatch (e.g. Whisper's jw for Javanese) still needs to be handled at the processor level.
• Encoder-only latency improves by 1.27× at TRR = 0.40 on the reported RTX 3080 Ti benchmark. See paper §5.4 and tmm_asr.eval.wallclock for the measurement protocol.
• Trained only on FLEURS narrow-band read speech. Domain shift to noisy / conversational speech not evaluated.

Citation

@inproceedings{holyoak2026tokenmerging,
  title     = {Token Merging for Multilingual Speech Recognition:
               A Systematic Study Across Model Scale and Fine-Tuning},
  author    = {Holyoak, Dylan Luke},
  booktitle = {Proceedings of ICNLSP 2026},
  year      = {2026}
}

License

Apache-2.0. Base Whisper checkpoint carries its own upstream MIT license.