ThoxVoice-Out.

On-device speech synthesis. No audio sent to a server. Also the listening-test tool for picking the THOX default voice.

local-first probing… Open RAIL-M no default picked

1 · Load

WebGPU: …
Model: onnx-community/Supertonic-TTS-ONNX · fp32 only
First load: 262.9 MB (one time, then cached)
Output: 44.1 kHz mono · English only

steps speed

Not loaded. Nothing is downloaded until you press Load.

2 · Text

–RTF ↓

–synth ms

–audio s

–backend

3 · Listening test — pick the THOX default

All 10 shipped voices. Click one to hear it, then Set default. This exists because no naturalness claim may be made until a human listens — open item #9 in the review.

Licence obligations (BigScience Open RAIL-M). This audio is machine-generated and any THOX surface presenting it must say so (Attachment A(e)). The use restrictions must be passed to end users (§5) and are viral — a rebrand cannot strip them. No voice cloning: only these 10 stock voices ship, and no speaker-embedding upload is exposed (Attachment A(g)).

ThoxVoice-Out — speech-out half of the THOX voice stack. Runtime transformers.js 3.8 (Apache-2.0) on ONNX Runtime Web (MIT) via the stock pipeline("text-to-speech") API. No custom WebGPU kernels. Weights: onnx-community/Supertonic-TTS-ONNX from Supertone/supertonic under BigScience Open RAIL-M. THOX trained nothing here — this is an integration.

Work product of Thox.ai LLC, a Texas limited liability company — CTO Tommy Xaypanya, CEO Craig Ross. Not affiliated with, endorsed by, or sponsored by Supertone, Inc., HYBE, or Hugging Face; Open RAIL-M §8 grants no trademark rights. Provided AS-IS, without warranty of any kind. No network calls after the model download. No telemetry.