Seed-VC
Free, open-source voice conversion model that changes a narration's voice identity while keeping its original timing, emotion, and delivery.
Pros
- + Completely free and self-hosted — no per-character or per-request billing at all
- + Preserves the expressive performance of the source audio (pauses, emphasis, pacing) instead of flattening it
- + Works well as a second pass on top of a TTS engine's output rather than replacing TTS entirely
Cons
- − Requires a GPU and some comfort with running Python projects locally
- − Voice cloning quality depends heavily on using a clean, dry reference clip — a processed or MP3-derived reference can introduce artifacts
- − No hosted API — this is a run-it-yourself tool, not a plug-and-play SaaS
Seed-VC solves a specific problem: you already like the performance a TTS engine gives you — its pacing, its emotional beats, its question/pause delivery — but you want a different voice identity on top of it. Instead of generating narration and voice in one step, you generate the performance first, then convert the voice.
How it works
You feed it a source narration file plus a short reference clip of the target voice (10-20 seconds of clean, dry audio works best — avoid MP3-derived or already-processed references, which can introduce metallic artifacts) and it outputs the same performance in a new voice. Typical settings run around 50 diffusion steps with a CFG around 0.45 for a clean, non-distorted result.
Who it’s for
High-volume channels running their own automation pipeline who want expressive delivery without paying per-character for it repeatedly — generate once with a TTS engine, convert cheaply and indefinitely after that. Pairs naturally with Gemini TTS as the first-pass performance layer.