souba67/whisper-tiny-sundanese-slr36

whisper-tiny Sundanese, SLR44 + SLR36

whisper-tiny fine-tuned for Sundanese ASR on OpenSLR SLR44 (4,213 studio utterances) plus a 20,000 utterance sample of SLR36 (crowdsourced phone-recorded speech, archives 0-3), 2,500 steps, lr 1e-4, fp16, single Kaggle T4. Both corpora are CC BY-SA 4.0. ONNX weights included for in-browser inference through transformers.js.

Accuracy

An ablation across three training mixes, all evaluated on the same frozen SLR44 test split (seed 42) and the same private 60-clip laptop-microphone set (531 reference words, .strip().lower() normalisation, word-level Levenshtein):

training mix	SLR44 clean test	laptop mic
SLR44 only (whisper-tiny-sundanese)	10.6%	69.68%
+ 1,500 TTS-synthetic clips	10.52%	70.06%
+ simulated noise/reverb/channel	12.61%	72.69%
+ 20K real SLR36 phone speech (this model)	8.43%	62.34%

Synthetic augmentation of either kind did not move microphone accuracy; real channel and speaker diversity moved both metrics at once. Known limits: the mic set is one speaker and one device, and 12 more SLR36 archives (~165K utterances) remain unused.

ONNX export fidelity

Sixty SLR44 clips, exact-match agreement of transcriptions against the PyTorch weights: fp32 100.0%, int8 96.7%, with int8 WER equal or better on the same clips. The WER numbers in that parity run (1.74%) are NOT an accuracy figure: those clips overlap this model's training data, and the run exists only to confirm the exports reproduce PyTorch. For accuracy, use the table above. Quantisation cuts the decoder from 118.4 MB to 30.1 MB and the encoder from 32.9 MB to 10.1 MB.

Usage

import { pipeline } from '@huggingface/transformers'
const asr = await pipeline('automatic-speech-recognition',
  'souba67/whisper-tiny-sundanese-slr36', { dtype: 'q8' })
const out = await asr(audio, { language: 'su', task: 'transcribe' })

Kaggle runs: training muhammadshobir/phase-g3-whisper-slr36 v2, export muhammadshobir/phase-b3-whisper-onnx-slr36 v1.