askurios8/supertonic-3-int8

Supertonic 3 — int8 quantized variants

This is a modified copy of Supertone/supertonic-3. It is not an official Supertone release. No retraining was done; graph inputs and outputs are unchanged, so the official Supertonic inference code runs as is.

Variants

Variant	Path	Size	What was done	Speed
Quality (default)	onnx/	105 MB	Weight-only int8: Conv / MatMul / Gemm / Gather weights stored as per-channel symmetric int8 with a DequantizeLinear node; all activations stay fp32. duration_predictor is left fp32 (3.7 MB) so utterance length matches the original exactly.	same as fp32
Fast	variants/mixed-fast/onnx/	105 MB	text_encoder and vector_estimator: ONNX Runtime dynamic quantization (quantize_dynamic, QUInt8, per-channel). vocoder: weight-only int8 as above. duration_predictor: fp32.	about 2× faster on native multi-threaded CPU

Both variants share onnx/tts.json, onnx/unicode_indexer.json and voice_styles/*.json (unmodified copies of the original files).

Why the vocoder is never dynamically quantized

An earlier revision of this repository dynamically quantized all four graphs. Measured against the fp32 model with the same noise seed, dynamic quantization of the vocoder alone cut output level to 0.46× and gave the largest spectral error of any module; Korean speech became faint and hard to understand. Dynamic quantization of the other modules had a small effect. Weight-only int8 keeps the output level at 0.99× of fp32.

Configuration	Size	RMS vs fp32 (ko / en)	log-spectral distance (ko / en)
fp32 original	398 MB	1.00 / 1.00	0
all modules dynamic int8 (old revision)	102 MB	0.48 / 0.98	3.2 / 2.0
Quality (this repo, onnx/)	105 MB	0.99 / 0.99	0.64 / 0.80
Fast (variants/mixed-fast/)	105 MB	1.09 / 1.07	0.84 / 0.92

License

Distributed under the same BigScience Open RAIL-M license as the original model; see LICENSE. The use restrictions in Attachment A of that license apply to these copies and to anything you build with them. Speech generated with this model is AI-synthesised.

Original model © Supertone Inc.