Best Local TTS Engine for an NVIDIA GPU Setup (2026)

XTTS v2, Chatterbox, and Bark all run on CPU, but a GPU is what makes them practical for real-time or batch use — Kokoro is the exception, since its 82 million parameters make it fast on CPU alone. The three heavier engines share a common shape: they are cloning- or generation-quality-first models where the architecture trades compute for expressiveness, so CUDA acceleration removes the bottleneck rather than adding a marginal speedup.

Use a GPU-oriented engine (XTTS v2, Chatterbox, or Bark) if your use case is voice cloning, expressive non-speech audio, or any workload where generation speed matters — a batch narration job, an interactive voice application, or a pipeline generating many clips. Avoid spending GPU VRAM on TTS if your only need is straightforward narration of plain text with no cloning: Kokoro or Piper cover that case on CPU alone, freeing your VRAM for an LLM or other workload running alongside it.

XTTS v2, released by Coqui and run through the Coqui TTS toolkit, is the best local voice-cloning option for GPU owners who need the highest quality achievable from a short reference clip. It clones a voice from as little as 6 seconds of reference audio and speaks it back across 17 languages, including cross-lingual cloning — clone a voice from English audio, speak the result in Spanish or Japanese.

On VRAM, XTTS v2's model weights are roughly 2 GB, but 4 GB is the practical minimum to run it and 4-6 GB is recommended for real-time inference, per PromptQuorum's dedicated XTTS v2 review. The license is the deciding factor for most readers: the Coqui Public Model License (CPML) is explicitly non-commercial, and because Coqui AI, the company, shut down its paid services in December 2023, there is currently no confirmed active pathway to a commercial license.

Chatterbox, released by Resemble AI on GitHub, is the best pick for GPU owners who want real-time conversational voice cloning under a fully permissive, commercial-friendly license. It is licensed under the MIT License, which — unlike XTTS v2's non-commercial CPML — permits commercial use without a separate agreement.

Chatterbox ships in multiple sizes: a Nano variant (110 million parameters) that Resemble AI's own documentation says runs faster than real time on 8 CPU cores, a Turbo variant (350 million parameters) built for low latency, and a Multilingual variant (roughly 500 million parameters, on a backbone that credits Llama 3 in its architecture) supporting 20+ languages. Zero-shot voice cloning works from a reference audio clip — Resemble AI's own example uses a 10-second clip, though the README does not state an official minimum duration the way XTTS v2's does.

Resemble AI's public README does not publish an exact VRAM figure, so treat any specific GB number you see elsewhere as unverified until you test it on your own card. What is documented: the Nano variant runs acceptably on CPU alone, while the larger Turbo and Multilingual variants are built for GPU-accelerated, low-latency generation — CUDA and Apple Silicon (MPS) are both supported.

Bark, released by Suno on GitHub, is the right pick if you want more than speech — laughter, sighs, gasps, and simple music generated from text prompts alone — and it is the engine in this comparison that benefits most from a GPU, since its generative, token-by-token architecture is the slowest of the four without CUDA acceleration. It does not support custom voice cloning; per Suno's own documentation, it "does not currently support custom voice cloning."

On VRAM, PromptQuorum's dedicated Bark review documents the full model needing about 12 GB, with a small-model environment flag (SUNO_USE_SMALL_MODELS) reducing that to about 8 GB — meaningfully more than XTTS v2's 4-6 GB. It is licensed under the MIT License, made fully commercial-use-eligible on May 1, 2023, but its maintenance status is a real open question: the public GitHub repository shows no commits since April 5, 2024.

Kokoro is included here as the counterpoint: owning an NVIDIA GPU does not mean every TTS workload needs one, and Kokoro is proof. At 82 million parameters, it is dramatically smaller than XTTS v2, Chatterbox, or Bark, and its own Hugging Face model card documents real-time or faster synthesis on CPU alone, with GPU use adding headroom rather than being a requirement.

Kokoro is licensed under Apache 2.0, which — like Chatterbox's MIT license — permits commercial use without restriction. It does not support voice cloning, so it is not a substitute for XTTS v2 or Chatterbox if cloning is the requirement, but for straightforward narration, read-aloud text, or a voice layer in an app where cloning is not needed, it is a lighter and simpler choice.

This table scores the four engines specifically on GPU-fitness criteria — VRAM requirement, how much each one benefits from CUDA acceleration, voice-cloning capability, and license — not on raw audio quality.

A GPU with 6 GB of VRAM or more comfortably covers every engine in this comparison except Bark's full model, which needs about 12 GB (or about 8 GB with its small-model flag). Match your card to the engine you actually need rather than buying for the heaviest option by default.

For general guidance on choosing a GPU by VRAM tier for local AI workloads (not TTS-specific), see PromptQuorum's GPU buying guide for local LLMs — the same VRAM-first purchasing logic applies to TTS, and if you already run a local LLM alongside TTS, the two workloads compete for the same VRAM pool.

Cloning a real person's voice without their knowledge or explicit consent is a distinct problem from licensing, and it applies regardless of which engine, license, or commercial status you are operating under. XTTS v2 and Chatterbox are both capable, general-purpose voice-cloning tools; neither one verifies that the reference audio you provide belongs to someone who has agreed to be cloned.

Cloning a voice without consent can raise consent, right-of-publicity, and in some jurisdictions fraud or impersonation concerns, independent of whether your project is commercial or personal, and independent of the model's license terms. This is a factual note, not legal advice — consult a lawyer if you are building a product that clones real people's voices, and always obtain clear, documented consent from anyone whose voice you clone.

For readers who already have an NVIDIA GPU and want to use it for text-to-speech, the choice comes down to what the audio needs to do. XTTS v2 is the pick for the highest achievable voice-cloning quality in a non-commercial or research context, given its non-commercial CPML license. Chatterbox is the pick when the same cloning capability needs to ship in a commercial product, thanks to its MIT license and real-time-oriented smaller model sizes. Bark is the pick specifically when the project needs expressive, non-speech audio — laughter, sighs, simple music — and can spend 8-12 GB of VRAM to get it, with the caveat that its maintenance status is uncertain. Kokoro remains the right choice whenever cloning is not the requirement: it runs comfortably on CPU alone, so reserving GPU VRAM for it is rarely worth doing. If unsure, start with Kokoro for simple narration and move up to XTTS v2 or Chatterbox only once a real cloning requirement appears — that keeps your GPU's VRAM available for whatever else you are running alongside it, including a local LLM. For the licensing details behind every engine mentioned here, see PromptQuorum's local TTS and voice-cloning licensing guide.