qenneth/parakeet-tdt-0.6b-v3-finetuned-for-ATC

Parakeet-TDT-0.6B-v3 Fine-Tuned on ATC-ASR Dataset

Overview

This repository contains a fine-tuned version of NVIDIA Parakeet-TDT-0.6B-v3, optimized for automatic speech recognition (ASR) in air traffic control (ATC) communications.
The model was fine-tuned using NVIDIA NeMo on the Jacktol ATC-ASR Dataset to improve recognition accuracy in noisy, domain-specific ATC environments.
Following fine-tuning, the model achieves a state-of-the-art word error rate (WER) of 0.0599 on the dataset’s official test split.

---

Results

Metric	Value
Validation Word Error Rate (WER)	0.0558
Test Word Error Rate (WER)	0.0599
Training Time	< 1 hour on NVIDIA H200
Framework	NVIDIA NeMo
Checkpoint Size	2.34 GB

---

Model Details

Attribute	Description
Base Model	nvidia/parakeet-tdt-0.6b-v3
Dataset	jacktol/ATC-ASR-Dataset
Framework	NVIDIA NeMo
Epochs	16
Batch Size	16
Learning Rate	1e-4
Optimizer	AdamW (weight decay 1e-3)
Scheduler	CosineAnnealing
Warmup Steps	5000
Min LR	5e-6
Precision	Mixed precision (FP16)
Tokenizer	Parakeet default subword tokenizer

---

Dataset

• Name: Jacktol ATC-ASR Dataset
• Domain: Air Traffic Control communications
• Language: English
• Sampling Rate: 16 kHz
• Format: WAV + JSON transcripts

---

Citation

If you use this model, please cite both the base model and dataset authors:

@misc{nvidia2024parakeet,
  title={Parakeet-TDT-0.6B-v3},
  author={NVIDIA},
  year={2024},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3}}
}

@dataset{jacktol_atc_asr,
  title={ATC-ASR Dataset},
  author={Jacktol},
  year={2023},
  howpublished={\url{https://huggingface.co/datasets/jacktol/ATC-ASR-Dataset}}
}