Show HN: Voice cloning for a TTS whose style encoder was never released

Category: ai-ml

Tags: text-to-speech, voice-cloning, pytorch, inverse-optimization, wavlm

Score: 7.5/10 (Innovation: 8, Technical: 8, Documentation: 9, Utility: 5)

This project extracts voice style vectors from a frozen TTS model by inverse optimization, bypassing an unreleased style encoder. Using WavLM-based loss and rate matching, it recovers timbre and duration from a WAV file, enabling voice cloning without training. It's a clever research-driven approach with clear evaluation and a demo.

Target audience: AI researchers, ML engineers, TTS practitioners

Repository: https://github.com/kdrkdrkdr/supertonic.embed · Python · NOASSERTION · 54 stars

View on Hacker News