oddadmix/Whisperv3-tunisian-codeswitch

NADI 2026 Subtask 1.3 — Tunisian Code-Switched ASR (faruk-v7)

Whisper-large-v3 full fine-tune for Tunisian Arabic ↔ French/English code-switched ASR (NADI 2026 shared task, subtask 1.3).

• Recipe: fresh full-FT from tun-asr-aug-v4 on FARUKxAUTO/tunisian-asr-cleaned (46K, dense Tunisian↔French code-switch) + NADI TEDx train replay, 2 epochs, --spec_augment, paged_adamw_8bit, lr 5e-6 (a single longer cosine schedule — 2 epochs is the sweet spot).
• Results: validation WER 17.26% / CER 6.91%; blind test WER 15.22 (🥉).
• Decode greedy, language="ar", task="transcribe". Submit raw (scorer applies clean_transcription).

from transformers import WhisperForConditionalGeneration, WhisperProcessor
import torch
m = WhisperForConditionalGeneration.from_pretrained("oddadmix/nadi2026-subtask1.3-tunisian-codeswitch-asr-faruk-v7", torch_dtype=torch.bfloat16).cuda().eval()
p = WhisperProcessor.from_pretrained("oddadmix/nadi2026-subtask1.3-tunisian-codeswitch-asr-faruk-v7")
# feats = p(audio, sampling_rate=16000, return_tensors="pt").input_features.cuda().to(torch.bfloat16)
# ids = m.generate(feats, language="ar", task="transcribe", max_new_tokens=256)
# print(p.batch_decode(ids, skip_special_tokens=True))