nadare/supertonic-3-onnx-q4

Supertonic 3 ONNX Q4

This is an Unofficial quantized derivative of Supertone/supertonic-3. It is not affiliated with or endorsed by Supertone Inc.

The source model is fixed to revision 724fb5abbf5502583fb520898d45929e62f02c0b.

Distributed components

Component	Distribution form
duration_predictor.onnx	Unchanged upstream FP32
text_encoder.onnx	Unchanged upstream FP32
vector_estimator.onnx	Asymmetric Q4 MatMulNBits, block size 16; final projection and depthwise convolutions remain FP32
vocoder.onnx	Asymmetric Q4 MatMulNBits, block size 16; the five boundary layers listed in MODIFICATIONS.md remain FP32

The original tts.json, unicode_indexer.json, and the ten official preset voice styles are included without modification. Custom or purchased voice styles are not included.

Reproduction and verification

The repository includes the conversion, staging, and verification tools under release-tools/. The distribution manifest and checksum files cover every runtime and publication file. The modified ONNX files also contain metadata identifying the source revision and modification.

Usage

The directory layout and ONNX input/output contracts match the upstream ONNX package. A CPU build of ONNX Runtime with support for com.microsoft::MatMulNBits is required.

Refer to the upstream model card for synthesis examples, supported languages, model limitations, and responsible-use guidance.

License

The model and this derivative are distributed under the included BigScience Open RAIL-M license. Its use-based restrictions, including Attachment A, continue to apply. See THIRD_PARTY_NOTICES.md and MODIFICATIONS.md for provenance and modification details.