---
license: cc-by-4.0
base_model: nvidia/parakeet-tdt-0.6b-v3
tags:
  - coreml
  - parakeet
  - transducer
  - tdt
  - nemo
  - apple-silicon
  - speech-to-text
  - conformer
language:
  - en
pipeline_tag: automatic-speech-recognition
---

# Parakeet TDT 0.6B v3 — CoreML

CoreML conversion of [nvidia/parakeet-tdt-0.6b-v3](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3) for Apple Silicon (ANE + GPU).

## Architecture

Split architecture optimized for Apple Neural Engine:

- **Encoder** (`encoder.mlmodelc`): Conformer encoder compiled for ANE/GPU (~1.18 GB)
- **Predictor + Joint** (`predictor_joint.safetensors`): LSTM predictor + Joint network as float32 safetensors (~69 MB)

The encoder runs on ANE/GPU via CoreML, while the predictor and joint networks run on CPU via Accelerate for optimal performance.

### TDT (Token-and-Duration Transducer)

- Conformer encoder: 24 layers, 512 hidden, 8 attention heads
- LSTM predictor: bidirectional, 640 hidden
- Joint network: 640 hidden, 5 duration classes (0–4)
- Vocabulary: 8192 SentencePiece tokens

## Audio Specifications

| Parameter | Value |
|-----------|-------|
| Sample rate | 16,000 Hz |
| FFT size | 512 |
| Hop length | 160 (10 ms) |
| Mel bins | 128 |
| Max frequency | 8,000 Hz |
| Window | Hann |

## EnumeratedShapes (Encoder Buckets)

The encoder supports 4 input duration buckets for optimized ANE scheduling:

| Bucket | Duration |
|--------|----------|
| 1 | 5 seconds |
| 2 | 10 seconds |
| 3 | 15 seconds |
| 4 | 30 seconds |

## Size

~1.2 GB total (vs ~2.3 GB MLX float32)

## Requirements

- macOS 15+ (Sequoia)
- Apple Silicon (M1+, ANE recommended)
- CoreML framework

## License

This model is licensed under **CC-BY-4.0**. Original model by NVIDIA — attribution required.

See [nvidia/parakeet-tdt-0.6b-v3](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3) for the original model.

## Source

Converted from [nvidia/parakeet-tdt-0.6b-v3](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3) via MLX intermediate format using [oriloq-mlx](https://github.com/oriloq-s/oriloq-mlx).

Conversion chain: NeMo (.nemo) → MLX (safetensors) → CoreML (.mlmodelc)