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

Code: https://github.com//token-merging-multilingual-asr
Paper: TBD

Usage

python

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")
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	83.2	52.2	−31.0
Javanese	60.7	44.7	−16.0
Hausa	66.5	53.0	−13.5
Maltese	51.9	46.0	−5.9
Vietnamese	15.5	15.4	−0.1
Tamil	25.1	25.0	−0.1

Token merging on top of this adapter costs at most +0.59 pp WER at TRR = 0.40 (Lingala). See paper §5.3.

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 §5.3 ¶3.

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.
Wall-clock latency at batch=1 does NOT improve with encoder-side merging on modern GPUs — the merge op overhead dominates the compute saved. See paper Limitations.
Trained only on FLEURS narrow-band read speech. Domain shift to noisy / conversational speech not evaluated.

Citation

bibtex

@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},
  address   = {University of California, Los Angeles}
}

License

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