canary-qwen-2.5b-atc-v1

Canary-Qwen-2.5B Fine-Tuned for ATC ASR (LoRA baseline, v1)

Fine-tuned nvidia/canary-qwen-2.5b on the UWB-ATCC corpus for Air Traffic Control speech recognition using plain LoRA adaptation, no additional regularization. This is the v1 run in the Pilot-to-ATC Research v1/v2/v3 comparison — see suideepmax/canary-qwen-2.5b-atc-lora for the regularized v3 model, which outperforms this one.

Trained fresh on 2026-09-07/08 to independently re-verify the historically-cited 23.32% WER figure (which it reproduced almost exactly) as part of resolving a reproducibility investigation — see the linked GitHub repo's research_log/VALIDATION.md (VAL-010, VAL-012) for the full account.

Results

Training

• Dataset: UWB-ATCC (Prague Airport ATC, 11,543 train / 2,886 test utterances)
• Steps: 10,000 | LR: 5e-4 | Warmup: 1,000
• LoRA: r=128, alpha=256, dropout=0.01, targets=[q_proj, v_proj]
• Regularization: none (no SpecAugment, weight_decay=1e-3)
• Strategy: FSDP across 4x RTX 2080 Ti | Precision: fp16-true (eps=1e-4)
• Framework: NVIDIA NeMo 2.8.0rc0
• Verified training time: ~21 hours (measured directly on this run — a previously-documented ~5.3h figure for this configuration was found to be incorrect)

Usage

from nemo.collections.speechlm2.models import SALM
import torch

model = SALM.from_pretrained('nvidia/canary-qwen-2.5b')
state = torch.load('consolidated_model.pt', map_location='cpu')
model.load_state_dict(state, strict=False)
model.cuda().eval()

answer_ids = model.generate(
    prompts=[[{
        'role': 'user',
        'content': f'Transcribe the following: {model.audio_locator_tag}',
        'audio': ['atc_audio.wav']
    }]],
    max_new_tokens=128,
)
print(model.tokenizer.ids_to_text(answer_ids[0].cpu()))

Part of

Pilot-to-ATC Research — Comparative evaluation of W2V2 vs Canary-Qwen for ATC domain ASR.