SupertonicTTS Voice Style Extractor
github.com/kdrkdrkdr/supertonic.embed- Category
- AI Tools
- Rank
- No. 1404Tools index
Previous survey · No. 1414 ·
- Pricing
- Open Source
- Type
- TOOL
- Use case
- Design & Media
- GitHub
- 61 stars
- Date
About
A tool that extracts a voice style embedding from any WAV recording for use with SupertonicTTS, producing a style JSON compatible with the model's shipped voice presets even though the original style encoder was never publicly released. It requires an NVIDIA GPU, needing about 10GB by default or 2.5GB with a reduced batch setting, and ships with a demo of five example voices synthesized in five languages.
At a glance
The SupertonicTTS Voice Style Extractor derives a style file from recorded speech for use with a frozen speech model. It is worth reading as a research prototype when you need to understand the trade-off between resemblance and intelligibility.
Good fit
Speech researchers evaluating recordings they own or have explicit permission to use.
Consider something else when
Use preset voices when you need predictable narration or lack permission to reproduce a voice.
Start here
Start by reading the published evaluation and comparing the reported held-out results with the preset baseline. If you later run a trial, use your own consented recording and check intelligibility on a new sentence.
Cost and requirements
MIT-licensed code. The documented workflow requires CUDA and separately licensed model weights.
Limits to know
The author reports occasional dropped and repeated words. Speaker similarity metrics do not establish naturalness or permission to distribute the generated voice.
Sources checked 2026-11-05. Source-based assessment.
Why it made the leaderboard
Unlocks custom voice cloning for SupertonicTTS users who were previously limited to its five shipped preset voices, filling a gap the original project left open.
Tags
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.