---
license: cc-by-4.0
pipeline_tag: automatic-speech-recognition
language:
  - en
tags:
  - executorch
  - xnnpack
  - metal
  - asr
base_model: nvidia/parakeet-tdt-0.6b-v3
---

# Parakeet TDT (0.6B v3) — ExecuTorch Exports

This repository contains ExecuTorch `.pte` exports of the Hugging Face ASR model [`nvidia/parakeet-tdt-0.6b-v3`](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3) for:

- **XNNPACK** (CPU): **fp32** and **INT4 weights** (8da4w)
- **Metal** (macOS): **bf16**

Each export directory contains:

- `model.pte` — ExecuTorch program
- `tokenizer.model` — SentencePiece tokenizer (needed by the runner)

## Artifacts

| Variant | Backend | Precision / Quantization | Path |
|---|---|---|---|
| `xnnpack-fp32` | XNNPACK | fp32 | `xnnpack/fp32/model.pte` |
| `xnnpack-int4` | XNNPACK | **8da4w** (int8 dynamic activations + int4 weights), group size 32 | `xnnpack/int4/model.pte` |
| `metal-bf16` | Metal (MPS) | bf16 | `metal/bf16/model.pte` |

Tokenizers:

- `xnnpack/fp32/tokenizer.model`
- `xnnpack/int4/tokenizer.model`
- `metal/bf16/tokenizer.model`

## How to run (C++ runner)

Build the Parakeet runner from an ExecuTorch checkout:

```bash
./install_executorch.sh
make parakeet-cpu
make parakeet-metal  # macOS only
```

Then run from the ExecuTorch repo root:

### CPU / XNNPACK

```bash
./cmake-out/examples/models/parakeet/parakeet_runner \
  --model_path /path/to/model.pte \
  --audio_path /path/to/audio.wav \
  --tokenizer_path /path/to/tokenizer.model
```

### Metal (macOS)

```bash
DYLD_LIBRARY_PATH=/usr/lib ./cmake-out/examples/models/parakeet/parakeet_runner \
  --model_path /path/to/model.pte \
  --audio_path /path/to/audio.wav \
  --tokenizer_path /path/to/tokenizer.model
```

## Export recipe (ExecuTorch)

Exports were produced with `examples/models/parakeet/export_parakeet_tdt.py` from ExecuTorch commit `efe4f0cce36c4b95ac70eac969c0d55163f23c78` using:

- PyTorch `2.11.0.dev20251222`
- `TORCHAO_FORCE_SKIP_LOADING_SO_FILES=1` (export-only workaround for torchao C++ extension loading)

Commands:

```bash
# XNNPACK fp32
TORCHAO_FORCE_SKIP_LOADING_SO_FILES=1 python examples/models/parakeet/export_parakeet_tdt.py \
  --backend xnnpack \
  --output-dir ./parakeet_xnnpack_fp32

# XNNPACK int4 (8da4w)
TORCHAO_FORCE_SKIP_LOADING_SO_FILES=1 python examples/models/parakeet/export_parakeet_tdt.py \
  --backend xnnpack \
  --qlinear_encoder 8da4w --qlinear_encoder_group_size 32 \
  --qlinear 8da4w --qlinear_group_size 32 \
  --output-dir ./parakeet_xnnpack_int4

# Metal bf16
TORCHAO_FORCE_SKIP_LOADING_SO_FILES=1 python examples/models/parakeet/export_parakeet_tdt.py \
  --backend metal --dtype bf16 \
  --output-dir ./parakeet_metal_bf16
```

## Verification

All variants were verified to transcribe the same test audio correctly (and match eager PyTorch output for the XNNPACK exports).

- Audio sample: `hf-internal-testing/librispeech_asr_dummy` (`clean`, `validation[0]`)
- Reference transcript (dataset):
  - `MISTER QUILTER IS THE APOSTLE OF THE MIDDLE CLASSES AND WE ARE GLAD TO WELCOME HIS GOSPEL`
- Model output:
  - `mister Quilter is the apostle of the middle classes, and we are glad to welcome his gospel.`

## License / Attribution

- Upstream model license: **CC-BY-4.0** (see `nvidia/parakeet-tdt-0.6b-v3`).
- Please attribute NVIDIA when using these exported artifacts.