Skip to content
ModelListedExternal
Download

Supertonic 3 ONNX INT8

Dynamic int8 quantization of Supertone/supertonic-3 for low-memory on-device inference. Same 4-graph non-autoregressive flow-matching pipeline (durationpredictor → textencoder → vectorestimator ×N → vocoder), 31...

Sourcehuggingface.co/soniqo/Supertonic-3-ONNX-INT8/tree/11f5965fd0bc7dfb191a16d83772fc658a3c03d8
KindOtherBase modelSupertone/supertonic-3VersionRevision 11f5965Bgrade

Model source

Kind
Other
File size
1.22–62.3 MB
Base model
Supertone/supertonic-3
Version
vRevision 11f5965
Updated
Sep 25, 2026
Downloads
27
Source
Hugging Face

Source description

SupertonicTTS-3 — int8 ONNX (on-device)

Dynamic int8 quantization of Supertone/supertonic-3 for low-memory on-device inference. Same 4-graph non-autoregressive flow-matching pipeline (duration_predictor → text_encoder → vector_estimator ×N → vocoder), 31 languages, 44.1 kHz, ONNX Runtime — just smaller and lighter.

What changed vs the base

  • Weights quantized to int8 via onnxruntime.quantization.quantize_dynamic (QInt8); activations fp32.
  • ONNX weights: 398 MB → 102 MB. Inference peak RSS: ~1026 MB → ~327 MB (3.1×) (measured, Apple Silicon, ORT CPU).
  • RTF ≈ neutral on Apple Silicon (~0.20 @ 8 steps); the int8 speed win lands on weaker ARM CPUs / NPUs.
  • Roundtrip verified for English, German, Korean.

Contents

onnx/{duration_predictor,text_encoder,vector_estimator,vocoder}.onnx (int8) · onnx/tts.json · onnx/unicode_indexer.json · voice_styles/*.json (10 voices) · config.json · LICENSE.

Usage

Drop-in for the supertonic package via model_dir, or run the 4 graphs directly with ONNX Runtime. The text front-end is G2P-free (NFKD + unicode_indexer.json lookup — no espeak/phonemizer).

Attribution & license

Derivative of Supertone/supertonic-3 (commit 3cadd1ee6394adea1bd021217a0e650ede09a323) by Supertone, Inc. (paper arXiv:2503.23108). Licensed under BigScience OpenRAIL-M — the upstream use-based restrictions carry over (no non-consensual impersonation/deepfakes, no undisclosed machine-generated content, etc.) and must pass through to downstream users. This card marks it a modified (quantized) artifact per the license. The original LICENSE is included.

Other Supertonic-3 formats

Ecosystem

  • soniqo.audio — website / use-case explorer (transcription, voice cloning, live ASR, voice agents).
  • speech-core — C++ orchestration library; Supertonic plugs in as a TTSInterface ONNX model.
  • speech-swift — Apple Silicon MLX + CoreML runtime.
  • speech-android — Android SDK consuming on-device LiteRT bundles.

Other ONNX models in this collection

Sources

1 source
Hugging FacePrimaryVersion 11f5965fd0bc7dfb191a16d83772fc658a3c03d8
huggingface.co/soniqo/Supertonic-3-ONNX-INT8/tree/11f5965fd0bc7dfb191a16d83772fc658a3c03d8

Verified Sep 6

Model artifacts

4 artifacts

onnx/duration_predictor.onnx

onnx · 1.22 MB · SHA-256 e83ca3cadfcd…fe9e · Hugging Face

Download

onnx/text_encoder.onnx

onnx · 9.35 MB · SHA-256 40cec03cc994…595d · Hugging Face

Download

onnx/vector_estimator.onnx

onnx · 62.3 MB · SHA-256 23a2100f2984…6846 · Hugging Face

Download

onnx/vocoder.onnx

onnx · 24.5 MB · SHA-256 33ff71746dc1…cb82 · Hugging Face

Download

Source excerpts

2 excerpts
ReadmeSep 6huggingface.co/soniqo/Supertonic-3-ONNX-INT8

--- license: openrail base_model: Supertone/supertonic-3 language: - en - ko - ja - ar - bg - cs - da - de - el - es - et - fi - fr - hi - hr - hu - id - it - lt - lv - nl - pl - pt - ro - ru - sk - sl - sv - tr - uk - vi tags: - text-to-speech - tts - onnx - int8 - quantized - on-device - supertonic - flow-matching library_name: onnx pipeline_tag: text-to-speech --- # SupertonicTTS-3 — int8 ONNX (on-device) **Dynamic int8 quantization** of [`Supertone/supertonic-3`](https://huggingface.co/Supertone/supertonic-3) for low-memory on-device inference. Same 4-graph non-autoregressive flow-matching pipeline (`duration_predictor → text_encoder → vector_estimator ×N → vocoder`), **31 languages**, 44.1 kHz, ONNX Runtime — just smaller and lighter. ## What changed vs the base - Weights quantized to **int8** via `onnxruntime.quantization.quantize_dynamic` (QInt8); activations fp32. - **ONNX weights: 398 MB → 102 MB.** Inference **peak RSS: ~1026 MB → ~327 MB (3.1×)** (measured, Apple Silicon, ORT CPU). - RTF ≈ neutral on Apple Silicon (~0.20 @ 8 steps); the int8 speed win lands on weaker ARM CPUs / NPUs. - Roundtrip verified for English, German, Korean. ## Contents `onnx/{duration_predictor,text_encoder,vector_estimator,vocoder}.onnx` (int8) · `onnx/tts.json` · `onnx/unicode_indexer.json` · `voice_styles/*.json` (10 voices) · `config.json` · `LICENSE`. ## Usage Drop-in for the [`supertonic`](https://pypi.org/project/supertonic/) package via `model_dir`, or run the 4 graphs directly with ONNX Runtime. The text front-end is **G2P-free** (NFKD + `unicode_indexer.json` lookup — no espeak/phonemizer). ## Attribution & license Derivative of **Supertone/supertonic-3** (commit `3cadd1ee6394adea1bd021217a0e650ede09a323`) by **Supertone, Inc.** (paper [arXiv:2503.23108](https://arxiv.org/abs/2503.23108)). Licensed under **BigScience OpenRAIL-M** — the upstream **use-based restrictions carry over** (no non-consensual impersonation/deepfakes, no undisclosed machine-generated content, etc.) and must pass through to downstream users. This card marks it a modified (quantized) artifact per the license. The original `LICENSE` is included. ## Other Supertonic-3 formats - [**Supertonic-3 — LiteRT**](https://huggingface.co/soniqo/Supertonic-3-LiteRT) — Android / Qualcomm NPU (.tflite). - [**Supertonic-3 — CoreML**](https://huggingface.co/aufklarer/Supertonic-3-CoreML) — iOS /...

NoteSep 6huggingface.co/soniqo/Supertonic-3-ONNX-INT8

Source context: 27 downloads · 2 likes · Pipeline text-to-speech · Library onnx · Repo soniqo/Supertonic-3-ONNX-INT8