Model files
Derived from NVIDIA Parakeet TDT-CTC 110M, developed by NVIDIA NeMo and Suno. The original model and these converted weights are licensed under CC BY 4.0. This adaptation is not endorsed by the original authors.
Modifications: export of the TDT branch to ONNX; unsigned INT8 quantization of encoder MatMul weights; signed INT8 quantization of decoder/joiner weights. Encoder convolutions retain floating-point precision. The vocabulary is unchanged. No retraining or fine-tuning was performed.
The conversion script adapts the k2-fsa sherpa-onnx export recipe and retains the Xiaomi/Fangjun Kuang copyright and Apache-2.0 license. See APACHE-2.0.txt.
SHA256SUMS.txt and model-files.json identify the four runtime files, totaling 177,063,730 bytes. The source checkpoint SHA-256 is 1e6b5c3bcaa390b479bd8112512ede515d2ef5b61495568f2c266d6d5e680c3a.
Reproducible export environment: Python 3.12, NeMo 2.4.0, PyTorch 2.6.0+cu124 on CPU, ONNX 1.17.0, and ONNX Runtime 1.19.2. Run python export-balanced-tdt.py /path/to/checkpoint.nemo /path/to/new-output.