Whisper large-v3-turbo — ATC fine-tune
Fine-tuned openai/whisper-large-v3-turbo for air traffic control (ATC) radio transcription.
Results
Held-out validation (2,024 samples, ATCO2 + ATCoSIM), normalized WER/CER (lowercase, punctuation- and article-stripped):
| Model | WER | CER |
|---|---|---|
| This model (large-v3-turbo) | 7.83% | 3.58% |
| whisper-small fine-tune (prior) | 12.82% | 9.24% |
≈39% relative WER and ≈61% relative CER reduction vs the small fine-tune — mostly from eliminating hallucinated insertions. Inference is ~2× the small model's latency (still faster than real-time on Apple Silicon / a modern GPU).
Training
| Base model | openai/whisper-large-v3-turbo (809M params) |
| Data | ATCO2 + ATCoSIM — 8,095 train / 2,024 val (80/20, seed 42), transcripts normalized |
| Recipe | lr 5e-6, effective batch 4, warmup 500, fp16, early-stopping (patience 2) on eval loss |
| Best checkpoint | step 1,100 — eval_loss 0.0975 |
| Hardware | single NVIDIA H100 |
Usage
import librosa, torch
from transformers import WhisperProcessor, WhisperForConditionalGeneration
repo = "SingularityUS/ATC-whisper-turbo-v1"
processor = WhisperProcessor.from_pretrained(repo, language="en", task="transcribe")
model = WhisperForConditionalGeneration.from_pretrained(repo).eval()
audio, _ = librosa.load("atc_clip.wav", sr=16000)
features = processor(audio, sampling_rate=16000, return_tensors="pt").input_features
with torch.no_grad():
ids = model.generate(features)
print(processor.batch_decode(ids, skip_special_tokens=True)[0])
Intended use & limitations
For English ATC VHF radio transcription. The validation set is predominantly clean studio-quality audio (ATCoSIM), so real-world noisy-VHF error rates will be higher. Research / non-safety-critical use only.
License & data
Base model is MIT-licensed (OpenAI Whisper). Training data (ATCO2, ATCoSIM) is subject to its respective licenses; this fine-tune is intended for research use — confirm the applicable terms before any commercial use.