Supertonic
Supertonic's archived repository preserves an on-device text-to-speech system and its setup examples. The archive says development and support have ended, with no further bug fixes or security patches. It remains a local software option for people prepared to maintain their own integration, rather than a currently supported hosted service.
Download the matching code and model assets
The archive's default quick start uses Supertonic 3. It instructs users to download a pinned model snapshot containing ONNX files, configuration and preset voices, then run a Python example against those local assets. Dependencies and model files require an initial download; the documented inference example runs without a network connection afterward.
Older Python package releases may still try the original download namespace. The archive advises using a local model directory with automatic downloads disabled. Check the preserved setup guide rather than copying a historical command that points to an unavailable service.
Examples cover Python, Node.js, browser execution and other runtimes. Some require native dependencies, including the ONNX Runtime C library for Go or a JDK for Java. Match the example to the environment you intend to maintain.
Review voice quality on your own text
The preserved Supertonic 3 description lists thirty-one languages and 44.1 kHz, 16-bit WAV output. It also describes a roughly 99-million-parameter model and inline expression tags. Those features provide a starting specification; they do not establish pronunciation quality or response time on your hardware.
Test names, abbreviations and numeric expressions from the material you need to narrate. The repository contains benchmark results and demonstration clips, but some samples are historical and externally hosted. Franklin has not independently reproduced their speed or error rates.
For a browser-first creator workflow, AnySpeech's voice-selection interface instead describes pasting a script, previewing voices and downloading MP3 output. Supertonic requires local assets and runtime setup. That is a deployment and delivery-format distinction, not a comparison proving either system sounds better.
Separate the archive from retired services
The current archive says hosted demos and Voice Builder are not included. Historical release notes still mention managed services and voice creation, but those passages should not be read as current archive availability. The open-weight repository focuses on fixed-voice speech and does not provide an official voice-cloning pipeline.
The sample code uses the MIT license, while the accompanying model uses OpenRAIL-M. Review both before redistribution or commercial integration; open code does not make every model restriction disappear.
Because support has ended, assess dependency maintenance and security obligations before building a production feature around the archive. Keep a known working snapshot and test the generated audio after runtime changes.