NeoRoth/parakeet-tdt-0.6b-v3-mlx

Parakeet TDT 0.6B v3 — MLX

MLX safetensors conversion of nvidia/parakeet-tdt-0.6b-v3 for Apple Silicon.

Architecture

• Encoder: Conformer (1024 hidden, pre-encoding convolutions + transformer layers)
• Decoder: TDT Transducer (predictor LSTM + joint network, 5 duration classes: 0-4)
• Vocabulary: 1025 tokens (SentencePiece)
• Parameters: ~0.6B
• Audio input: 16 kHz mono, 128 mel bins

Contents

File	Description
model.safetensors	All weights (encoder + predictor + joint), float32, ~2.3 GB
config.json	Full NeMo model configuration
tokenizer.model	SentencePiece tokenizer
tokenizer.vocab	Tokenizer vocabulary
vocab.txt	Text vocabulary

Notes

• Weights converted from NeMo PyTorch format to MLX safetensors
• Convolution weights use MLX layout (OHWI for 2D, OKI for 1D) — not directly compatible with PyTorch
• CTC head, preprocessor, spec augmentation, and loss weights are excluded (inference only)

License

CC-BY-4.0 — following the upstream nvidia/parakeet-tdt-0.6b-v3 license. Attribution to NVIDIA is required.

Source

Converted from nvidia/parakeet-tdt-0.6b-v3.