Voice Cloning & Voice Conversion AI Tools

Reproducing a specific person's voice splits into two jobs that get conflated: synthesizing new speech in that timbre from text, and converting an existing recording so it keeps the original timing and performance but sounds like someone else. Localization and dubbing teams, game studios rerecording dialogue without recalling actors, and accessibility projects rebuilding a lost voice all end up here. The 2026 split is between zero-shot cloners conditioned on seconds of reference audio and per-speaker models trained on a dataset. OpenVoice V2, XTTS-v2, CosyVoice, and Fish Speech sit on the zero-shot side: setup takes minutes, cross-lingual transfer works, but prosody comes from the base model rather than the target speaker and drifts over long passages. The retrieval-based conversion family, RVC with its forks Applio and Mangio-RVC, plus FreeVC for one-shot conversion, trades GPU hours and a clean dataset for tighter timbre and a preserved source performance, which is why singing conversion and dub-to-picture work still runs there. W-Okada Voice Changer covers the realtime path. The reasonable entry points are Applio, the actively maintained RVC fork with a usable training interface, and OpenVoice V2 for text-driven cloning when there is no training budget. The item worth checking first is the weights license, which is often not the repository license. XTTS-v2 ships under Coqui's non-commercial model license, and Coqui is gone, so no relicensing is coming. Several Fish Speech releases carry non-commercial terms too. An Apache or MIT badge on the repo covers the inference code, not the checkpoint doing the work.