ThaiVanPhat95/wav2vec2-robust-uwb-supcon-hybrid

Wav2Vec2 Robust UWB SupCon Hybrid

Associated paper: Contrastive Regularization for Accent-Robust ASR.

This model performs English air-traffic-control automatic speech recognition. It was trained using CTC and supervised contrastive learning with a hybrid training schedule that combines:

• —Similar-transcript SupCon batches
• —Original and synthetic transcript pairs
• —CTC-only batches

The included processor supports greedy CTC decoding.

Intended Use

This model is intended for research on robust English ATC speech recognition. It may be used to transcribe English ATC recordings with acoustic and speaker characteristics similar to its training data.

Training Data

The model was trained using UWB-ATCC speech and synthetic speech derived from UWB-ATCC transcripts.

• —UWB-ATCC: https://huggingface.co/datasets/Jzuluaga/uwb_atcc
• —Training code: https://github.com/thaivanphat95/robust-atc-asr

Users should review and comply with the original dataset terms before using or redistributing this model.

Usage

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

model_id = "thaivanphat95/wav2vec2-robust-uwb-supcon-hybrid"

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

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

Audio should be mono. Resample audio to 16 kHz before inference when necessary.

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

• —The model is specialized for English ATC speech.
• —Performance may degrade for other domains, languages, recording conditions, or terminology not represented in the training data.
• —Greedy decoding does not use an external language model.
• —Model predictions should be reviewed before use in safety-critical settings.

License

The model weights are not covered by the Apache License 2.0 used for the training code. Their use and redistribution may also be affected by the licenses and terms of the pretrained model, UWB-ATCC, source transcripts, and synthetic speech-generation systems.