VITS EVOlution
An open-source text-to-speech engine that clones voices from a reference clip and runs 5.6x faster than real-time on a plain CPU, no GPU needed.
🔗 Visit VITS EVOlution
Description
Real-time voice cloning is usually associated with GPU servers and cloud APIs, which makes it hard to use in offline or resource-constrained settings like embedded devices or older laptops. VITS EVOlution is built specifically for that gap: it clones a voice from a reference clip and generates speech in it fast enough to run live, using nothing but a regular CPU.
The project exports both its speaker encoder and TTS model to ONNX format, reaching about 5.6x real-time inference speed on a Xeon CPU — no GPU acceleration required. It supports zero-shot voice cloning from reference audio and "voice blending," averaging multiple speaker embeddings to create hybrid voices, and ships a Gradio web interface alongside its Python library. It's Apache-2.0 licensed, uses DeepPhonemizer for text processing, and at 9 GitHub stars with 4 commits is a very early-stage, lightly-documented project — a solid proof of concept more than a mature tool.
💬 Our review
The short version: VITS EVOlution is a promising, CPU-first take on real-time voice cloning with a genuinely useful trick (voice blending), but at 9 stars and 4 commits it's closer to a research prototype than something to build a product on top of today.
Against KokoClone, a similar open-source voice-cloning project in the same discovery batch, VITS EVOlution's specific pitch is raw CPU speed (5.6x real-time on a Xeon) rather than multilingual breadth — KokoClone supports 8 languages, while VITS EVOlution's language support isn't documented as clearly. Against a commercial API like ElevenLabs, it's free and edge-deployable but comes with none of the polish, support, or guaranteed voice quality a paid service provides. Worth exploring if you specifically need CPU-only, real-time voice cloning for an edge device or offline app and are comfortable with an early-stage codebase; not something to depend on without testing thoroughly first.
💰 Pricing
📊 Global score
🤖 AI-enriched data
Gratuit, licence Apache 2.0
Pros
Inférence CPU à 5.6x le temps réel, aucun GPU nécessaire
Clonage vocal zero-shot et mélange de voix par moyenne d'embeddings
Déployable sur appareils embarqués/edge grâce au format ONNX
Licence permissive Apache 2.0, interface Gradio incluse
Cons
Projet très jeune : 9 étoiles GitHub, seulement 4 commits
Documentation limitée par rapport à des alternatives plus établies
Segments audio plafonnés (~limite technique des embeddings positionnels)