Supertonic 3 Guide: 31-Language Browser TTS

Supertonic 3 is a 99M-parameter, 31-language text-to-speech model designed for local ONNX Runtime inference. The official project supports web, Python, Node.js, iOS, Android, and C++ runtimes. OfflineTTS integrates its browser path with 10 built-in voice presets, speed and step controls, waveform playback, MP3/WAV export, and optional synchronized subtitles.

Open Supertonic 3 in OfflineTTS to try it without an account or API key.

Quick Facts

The official Supertonic 3 site distinguishes the OpenRAIL-M model weights from MIT-licensed sample code. Review the model license before commercial deployment because OpenRAIL-M includes use restrictions and attribution requirements.

The 31 Supported Languages

OfflineTTS currently exposes:

Supertonic is especially useful when a Kokoro language would otherwise require remote text phonemization. After the Supertonic model files are available, its supported language synthesis remains local.

Ten OfflineTTS Voices vs the Wider Ecosystem

The number 10 describes the built-in presets currently packaged by OfflineTTS. It does not describe every voice available across Supertone products.

The official Supertonic site links to separate hosted products for preset voices and voice building. Those services can have different catalogs, licensing, accounts, and data flows. OfflineTTS does not silently send a recording to those services and does not present its 10 local presets as a complete Supertone catalog.

Choose a preset by listening to the same sample text in several voices. Voice fit is more important than a generic “best” label, especially across languages.

How Browser Inference Works

1. The page downloads the Supertonic model assets.
2. The browser caches eligible files for later sessions.
3. Text is normalized for the selected language.
4. ONNX Runtime Web executes the model locally.
5. The generated PCM audio is displayed in the waveform player.
6. The result can be downloaded as WAV or encoded as MP3.

WebGPU is preferred when available. A WASM path provides a CPU fallback. First load time and generation speed vary with model caching, browser memory, available GPU features, power mode, and input length.

Generate Supertonic Speech

1. Open the Supertonic TTS app.
2. Select one of the 31 languages.
3. Choose a built-in voice.
4. Enter a short test paragraph.
5. Set speed and generation steps.
6. Enable synchronized subtitles if you need SRT or highlighted playback.
7. Generate, review the waveform, and export the result.

Start with the default controls. Change one setting at a time so you can hear whether speed, voice, or steps caused the difference.

Speed and Steps

Speed changes speaking rate. Very high or low values can reduce intelligibility, so test numbers, names, and sentence endings.

Steps control the amount of iterative synthesis work. More steps can increase processing time and may change output quality; fewer steps are useful for drafts. The most useful value depends on the language, device, voice, and passage.

Do not transfer a speed benchmark from the official CPU test directly to a browser promise. The official 5.00x result used a 16-thread CPU and a defined 30-sample methodology. Browser WebGPU and WASM performance will differ.

Long Text and Audiobooks

Neural TTS models work better when long text is divided into coherent passages. For articles:

• preserve paragraph breaks;
• avoid feeding a whole book as one request;
• review abbreviations and headings;
• generate one section before committing to a voice.

For EPUB, TXT, or text-based PDF, use the ebook-to-audio workflow. It divides a book into chapters and paragraphs, caches each completed result, and supports pause, resume, export, and read-along.

Synchronized Subtitles

Supertonic does not need to return native word timestamps for OfflineTTS to create subtitles. When timing is enabled, OfflineTTS runs local Whisper alignment against the generated audio. If complete word timing is unavailable, it uses a labeled weighted estimate so playback and SRT remain available.

Read TTS Word Timestamps: Create SRT Subtitles for the alignment, cue grouping, karaoke highlighting, and review process.

Privacy and Caching

Supertonic synthesis runs locally after required files are downloaded. Generated audio is held by the browser and only leaves the device if you upload or share it.

Browser storage is not permanent archival storage. Cached models or results may be removed when:

• the user clears site data;
• private browsing ends;
• the browser reclaims storage;
• a device policy limits IndexedDB or cache usage.

Download important output and keep the original text elsewhere.

Supertonic vs Kokoro, Piper, and Kitten

For French, compare French text to speech with Supertonic. For a wider engine decision, use the browser TTS comparison.

Troubleshooting

Model Loading Stops

Check storage, reload the page, close memory-heavy tabs, and retry in a current browser. Private mode may provide less persistent storage.

Output Uses the Wrong Language

Verify the language selector rather than relying on automatic inference. Mixed-language text may need to be split into separate generations.

Speech Is Too Fast or Unclear

Return speed and steps to their defaults, shorten the test, and add punctuation. Then adjust one control at a time.

Subtitles Use an Estimate

Try the larger Whisper alignment option or a shorter passage. Estimated subtitles remain available for playback and SRT drafts, but should be reviewed before precise video publishing.

Start with a Real Workflow

• Generate 31-language speech with Supertonic
• Compare all browser TTS engines
• Convert an ebook into read-along audio
• Create synchronized SRT from TTS