Yiivgeny/parakeet-tdt-0.6b-v3-sherpa-onnx-fp32

Parakeet TDT 0.6B v3 sherpa-onnx FP32

This repository contains a sherpa-onnx-compatible FP32 export of `nvidia/parakeet-tdt-0.6b-v3`.

It is intended for runtimes that expect the sherpa-onnx transducer layout:

tokens.txt
bpe.vocab
encoder.onnx
encoder.weights
decoder.onnx
joiner.onnx

The encoder uses ONNX external data, so encoder.weights must stay next to encoder.onnx.

Downloads

The model files are available both directly in this repository and as a compressed archive:

sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-fp32.tar.bz2

Checksums are in:

SHA256SUMS

Verify after download:

sha256sum -c SHA256SUMS

On macOS, use:

shasum -a 256 -c SHA256SUMS

Usage With sherpa-onnx

Extract the archive:

tar -xjf sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-fp32.tar.bz2
cd sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-fp32

Start the offline websocket server:

sherpa-onnx-ws \
  --tokens=./tokens.txt \
  --encoder=./encoder.onnx \
  --decoder=./decoder.onnx \
  --joiner=./joiner.onnx \
  --port=6006 \
  --num-threads=4

For runtimes using ONNX Runtime external data, start the process with the model directory as the current working directory, or otherwise ensure encoder.weights is resolvable next to encoder.onnx.

Export

This export was produced from nvidia/parakeet-tdt-0.6b-v3 using the sherpa-onnx NeMo export script for Parakeet TDT 0.6B v3.

Expected FP32 output files:

encoder.onnx
encoder.weights
decoder.onnx
joiner.onnx
tokens.txt
bpe.vocab

The export was smoke-tested with sherpa-onnx by transcribing the sample en.wav from the k2-fsa sherpa-onnx ASR model assets.

Notes

This is the non-quantized FP32 export. It is larger and heavier than the published int8 sherpa-onnx package, but avoids int8 quantization. For smaller downloads, use the int8 release from k2-fsa/sherpa-onnx.

License

The base model is released by NVIDIA under the CC BY 4.0 license. This converted ONNX export keeps the same license.