csukuangfj/Inflect-Nano-v2-ONNX

** Forked from https://huggingface.co/owensong/Inflect-Nano-v2-ONNX **

Inflect-Nano-v2 ONNX

Official, verified FP32 ONNX Runtime export of Inflect-Nano-v2. This repository is a format export only: no training, pruning, or quantization was applied.

• Complete neural weights: approximately 16.21 MB
• Output: 24 kHz mono waveform
• Providers: CPU, CUDA, or DirectML through ONNX Runtime
• Frontend: Python and eSpeak-ng
• License: Apache 2.0

The learned model is split into:

Graph	Purpose	Size
onnx/duration.onnx	Tokens to aligned acoustic distribution	3.64 MB
onnx/decode.onnx	Acoustic distribution and seeded noise to waveform	12.57 MB

Quick start

python -m pip install -r onnx/requirements.txt
python onnx/inference_onnx.py \
  --text "A small voice can still have something meaningful to say." \
  --output sample.wav \
  --provider cpu

See the complete ONNX guide for the Python API, provider selection, long-text behavior, parity measurements, checksums, and reproducible export instructions.

Verification

The graphs pass ONNX checker validation, dynamic-length inference, short- and long-form synthesis, seed determinism, and direct numerical parity testing against the canonical PyTorch checkpoint. See onnx/parity_report.json and onnx/SOURCE.json.

The canonical architecture, evaluation, audio samples, PyTorch checkpoint, and project documentation remain at owensong/Inflect-Nano-v2.