Parakeet TDT 0.6B v3 — ONNX INT8

File layout

Pipeline

Quantization strategy

• Encoder: static INT8 quantization including Conv, MatMul, and Gemm
• Decoder + joint: dynamic INT8 quantization applied before graph merging

Notes

• nemo128.onnx is the preprocessing graph.
• vocab.txt contains the SentencePiece/BPE vocabulary with <blk>.
• Total package size: approximately 673.6 MB.
• Designed for Handy / transcribe-rs-style Parakeet ONNX layouts.

Expected files

• encoder-model.int8.onnx — quantized encoder
• decoder_joint-model.int8.onnx — merged decoder + joint network
• nemo128.onnx — feature preprocessor
• vocab.txt — vocabulary

Validation checklist

1. ONNX Runtime can load both ONNX files.
2. nemo128.onnx, encoder, and decoder_joint run in your target runtime.
3. A short sample audio produces a non-empty transcript.