ThaiVanPhat95/wav2vec2-large-l2-arctic-supcon-repeated-arabic-holdout

Wav2Vec2 Large L2-ARCTIC SupCon Repeated: Arabic L1 Holdout

Associated paper: Contrastive Regularization for Accent-Robust ASR.

English accented-speech recognition model trained with CTC and supervised contrastive learning. Repeated L2-ARCTIC prompts provide positive SupCon pairs. Arabic-L1 speakers are held out for testing.

The included processor supports greedy CTC decoding.

Data and Code

• —L2-ARCTIC: https://psi.engr.tamu.edu/l2-arctic-corpus/
• —Training code and split metadata: https://github.com/thaivanphat95/robust-atc-asr

Usage

import torch
import soundfile as sf
from transformers import AutoModelForCTC, AutoProcessor

model_id = "thaivanphat95/wav2vec2-large-l2-arctic-supcon-repeated-arabic-holdout"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForCTC.from_pretrained(model_id).eval()

audio, sample_rate = sf.read("audio.wav")
inputs = processor(audio, sampling_rate=sample_rate, return_tensors="pt")

with torch.no_grad():
    logits = model(**inputs).logits

print(processor.batch_decode(torch.argmax(logits, dim=-1))[0])

Audio should be mono and sampled at 16 kHz.

Citation

@article{thai2026contrastive,
  title={Contrastive Regularization for Accent-Robust ASR},
  author={Thai, Van-Phat and Dhruv, Aradhya and Pham, Duc-Thinh and Alam, Sameer},
  journal={arXiv preprint arXiv:2605.03297},
  year={2026},
  doi={10.48550/arXiv.2605.03297}
}

Limitations and License

The model is specialized for L2-ARCTIC-style English read speech. The Arabic holdout describes the evaluation split, not exclusive specialization for Arabic-accented English. Model-weight use and redistribution may be affected by the licenses of the pretrained model and L2-ARCTIC. The training code's Apache-2.0 license does not cover these weights.