Supertonic-3-ONNX-WebGPU

Supertone/supertonic-3 with its weights quantized for running in a web browser with ONNX Runtime Web's WebGPU backend, as used by Quick Inference.

Changes from the original

• The four graphs (onnx/duration_predictor.onnx, text_encoder.onnx, vector_estimator.onnx, vocoder.onnx) are quantized without calibration (round to nearest, symmetric, blocks of 32): matrix multiplications to 4-bit MatMulNBits, embedding tables to GatherBlockQuantized, and convolution weights to blocked int8 DequantizeLinear (opset 21). Activations stay float32. 398 MB → 110 MB.
• Convolutions are 8-bit rather than 4-bit because 4-bit convolutions made the vocoder audibly buzz.
• onnx/tts.json, onnx/unicode_indexer.json and the ten voices in voice_styles/ are the original files.

These are modified files: their outputs are close to, but not identical to, the original model's. On eight test sentences, Whisper small.en heard the quantized model with an 8.0% word error rate (the original: 11.4%), at the same speech length and without extra high-frequency noise.

The inference loop is unchanged: see the original repository and its supertonic Python package.

Licence and use restrictions

Licensed under the BigScience OpenRAIL-M License, as the original (LICENSE). Its use-based restrictions (paragraph 5 and Attachment A) apply to these files and to anything made from them: you may not use them for the uses listed in Attachment A, and if you distribute them or a derivative you must include those restrictions as an enforceable provision and give recipients a copy of the licence.