ASR/parakeet/README.md

---
language:
  - en
license: other
tags:
  - nemo
  - parakeet
  - tdt
  - automatic-speech-recognition
  - air-traffic-control
  - atc
  - singapore
  - military
base_model: nvidia/parakeet-tdt-0.6b-v2
pipeline_tag: automatic-speech-recognition
metrics:
  - wer
model-index:
  - name: parakeet-tdt-0.6b-v2-atc-singapore
    results:
      - task:
          type: automatic-speech-recognition
        metrics:
          - name: Validation WER
            type: wer
            value: 0.72
---

# Parakeet-TDT 0.6B v2 - Singapore Military ATC

Fine-tuned NVIDIA Parakeet-TDT 0.6B v2 for Singapore Air Force air traffic control speech recognition.

## Performance

| Checkpoint | Validation WER | Notes |
|------------|----------------|-------|
| `model.ckpt` | **0.72%** | Best checkpoint from epoch 76 |
| `epoch=100-val_wer=0.0073-last.ckpt` | 0.73% | Final checkpoint, not published here |

## Model Details

| Key | Value |
|-----|-------|
| Base model | `nvidia/parakeet-tdt-0.6b-v2` |
| Framework | NeMo |
| Model class | `EncDecRNNTBPEModel` |
| Format | Raw `.ckpt` checkpoint + tokenizer artifacts |
| Checkpoint size | 7.0 GB |
| Domain | Singapore military ATC (Tengah WSAT, Paya Lebar WSAP) |
| Training data | Clean originals plus mild ATC radio, speed, and stress augmentations |

## Included Files

- `model.ckpt` - best fine-tuned checkpoint
- `artifacts/705f11d22dc04b169effc35ce5cd1361_tokenizer.model`
- `artifacts/a4715c7f6b2d4c2bb709306073d0c0a4_tokenizer.vocab`
- `artifacts/4cf78c8ca4ca44fca36c3754478fb188_vocab.txt`

## Usage

This repo currently publishes the fine-tuned model as a raw NeMo checkpoint rather than a packaged `.nemo` archive. The tokenizer artifact paths therefore need to be pointed at the local `artifacts/` folder before restore.

```python
from pathlib import Path

import torch
from omegaconf import OmegaConf
from nemo.collections.asr.models import EncDecRNNTBPEModel

model_dir = Path("ASR/parakeet")
ckpt_path = model_dir / "model.ckpt"
artifacts_dir = model_dir / "artifacts"

bundle = torch.load(ckpt_path, map_location="cpu", weights_only=False)
cfg = bundle["hyper_parameters"]["cfg"]
cfg = OmegaConf.create(OmegaConf.to_container(cfg, resolve=False))
cfg.tokenizer.model_path = str(
    artifacts_dir / "705f11d22dc04b169effc35ce5cd1361_tokenizer.model"
)
cfg.tokenizer.vocab_path = str(
    artifacts_dir / "4cf78c8ca4ca44fca36c3754478fb188_vocab.txt"
)
cfg.tokenizer.spe_tokenizer_vocab = str(
    artifacts_dir / "a4715c7f6b2d4c2bb709306073d0c0a4_tokenizer.vocab"
)

model = EncDecRNNTBPEModel.load_from_checkpoint(str(ckpt_path), cfg=cfg)
model.eval().cuda()

hypotheses = model.transcribe(
    ["audio.wav"],
    return_hypotheses=True,
    timestamps=True,
)

if isinstance(hypotheses, tuple):
    hypotheses = hypotheses[0]

hyp = hypotheses[0]
print(hyp.text)
print(hyp.timestamp["word"])
```

Tested in ASTRA with `nemo_toolkit[asr]==2.7.3`.

## Output Format

The model outputs normalized spoken text intended for downstream ATC formatting:

| Input audio says | Model outputs |
|-----------------|---------------|
| "CAMEL climb flight level zero nine zero" | `camel climb flight level zero nine zero` |
| "Contact Tengah Approach one three zero decimal zero" | `contact tengah approach one three zero decimal zero` |
| "Squawk four five two one" | `squawk four five two one` |

ASTRA then applies a deterministic formatter to convert normalized speech into display text such as `CAMEL climb FL090`.