cjamcu/whisper-atc-co-medium

Whisper ATC-CO Medium

PEFT/LoRA fine-tuned Whisper-medium for Colombian Air Traffic Control (ATC) radio transcription.

Model Details

• Base model: openai/whisper-medium
• Method: PEFT (Parameter-Efficient Fine-Tuning) with LoRA
• Dataset: cjamcu/atc-co-curated (approved transcripts only)

Training Parameters

Parameter	Value
Epochs	15
Learning rate	1e-3
Batch size	4
Gradient accumulation	4
Warmup steps	25
LoRA rank (r)	32
LoRA alpha	64
LoRA dropout	0.05
Target modules	q_proj, v_proj
Train samples	122
Validation samples	15

Evaluation (test split, 16 samples)

Metric	Baseline (whisper-medium)	Finetuned	Delta
WER raw	0.96	0.73	-0.23
CER raw	0.83	0.57	-0.26
WER norm	0.92	0.65	-0.27
CER norm	0.81	0.53	-0.28

Usage

from transformers import WhisperForConditionalGeneration, WhisperProcessor
from peft import PeftModel

# Load base model + adapter
base = WhisperForConditionalGeneration.from_pretrained("openai/whisper-medium")
model = PeftModel.from_pretrained(base, "cjamcu/whisper-atc-co-medium", subfolder="adapter_model")
processor = WhisperProcessor.from_pretrained("cjamcu/whisper-atc-co-medium")

# Transcribe
import torch
audio, sr = ...  # load 16kHz mono audio
inputs = processor(audio, sampling_rate=sr, return_tensors="pt")
with torch.no_grad():
    generated = model.generate(inputs.input_features)
transcription = processor.batch_decode(generated, skip_special_tokens=True)[0]

Limitations

• Trained on 122 samples; performance may vary on out-of-distribution ATC audio
• Test set has only 16 samples; metrics have high variance
• Spanish/English mixed ATC communications

Citation

@misc{whisper-atc-co-medium,
  author = {ATColombia},
  title = {Whisper ATC-CO Medium: Colombian ATC transcription},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/cjamcu/whisper-atc-co-medium}
}