Local TTS has gotten shockingly good. You can generate natural narration for free on any PC, or clone a voice from a 6-second clip — all offline. Here's what to use.
Yes, and it's genuinely good now. For natural narration, Kokoro generates high-quality speech in just 2-3GB of VRAM — even on CPU — with 54 built-in voices, free and offline. For voice cloning, XTTS v2 can clone a voice from a 6-second clip across 17 languages, and the newer Chatterbox rivals paid services like ElevenLabs. It all runs on your own machine, privately, for free. The main things to get right are picking the model for your use case (narration vs cloning) and checking the license. Here's the guide.
Narration vs voice cloning — pick your tool
There are two jobs people want from local TTS, and different models are best at each. For narration — reading text in a good preset voice (audiobooks, video voiceover, accessibility) — Kokoro-82M is the standout: it's tiny (runs in 2-3GB, even on CPU-only machines), ships 54 fixed voices across 8 languages, sounds natural, and is Apache 2.0 licensed so you can use it commercially. Piper is another excellent lightweight choice for a simple first setup. Neither clones voices — they use preset voices, which is exactly what most narration needs. For voice cloning — reproducing a specific voice from a sample — XTTS v2 is the long-standing open favorite, cloning from a ~6-second clip in 17 languages (note: its license is non-commercial). The newer Chatterbox Multilingual is the exciting 2026 entrant, covering 23 languages and, in its maker's blind test, being preferred over ElevenLabs by most listeners. F5-TTS is another strong English cloner. The catch with all cloning models: check the license (several are non-commercial) and, ethically and legally, only clone a voice you have consent to use.
Best local TTS models by use case (2026)
Narration (lightweight)
Use case
Kokoro-82M
Model
2-3GB, CPU-ok, Apache 2.0
Narration (simple setup)
Use case
Piper
Model
Great first setup
Voice cloning (open)
Use case
XTTS v2
Model
6-sec clip, 17 langs (non-commercial)
Voice cloning (new)
Use case
Chatterbox
Model
23 langs, beats ElevenLabs in blind test
Use case
Model
Notes
Narration (lightweight)
Kokoro-82M
2-3GB, CPU-ok, Apache 2.0
Narration (simple setup)
Piper
Great first setup
Voice cloning (open)
XTTS v2
6-sec clip, 17 langs (non-commercial)
Voice cloning (new)
Chatterbox
23 langs, beats ElevenLabs in blind test
Kokoro narrates in 2-3GB (even on CPU); XTTS and Chatterbox clone a voice from a short clip — all offline. · Unsplash
How to get started
The easiest on-ramp depends on your comfort level. For narration, Kokoro and Piper both have simple installers and lightweight requirements — you can run them on almost any machine, including one without a dedicated GPU, and they're a quick pip install or a downloadable app. For voice cloning with XTTS v2 or Chatterbox, you'll typically use a Python setup or one of the community GUIs that wrap these models; they want a bit more hardware (a GPU with several GB of VRAM helps, though modest cards work), and you feed them a clean voice sample. Because these are all local and offline, your text and any voice samples never leave your computer — which matters a lot for cloning, where you're handling someone's voice. Pair local TTS with local speech-to-text (Whisper) and you have a full private voice pipeline: transcribe with Whisper, generate speech with Kokoro, all on your own hardware for free.
Quick answers
What is the best free local text-to-speech AI?
For natural narration, Kokoro-82M is the best lightweight choice — it runs in just 2-3GB of VRAM (even on CPU), ships 54 voices across 8 languages, sounds natural, and is Apache 2.0 licensed for commercial use. Piper is another great simple option. For voice cloning, XTTS v2 is the leading open model (cloning from a 6-second clip in 17 languages, though non-commercial), and the newer Chatterbox rivals ElevenLabs. All run locally and free. Pick Kokoro/Piper for preset-voice narration, XTTS/Chatterbox for cloning.
Can I clone a voice locally with AI?
Yes. XTTS v2 can clone a voice from about a 6-second audio clip across 17 languages, running entirely on your own machine, and the newer Chatterbox model covers 23 languages and performed well against ElevenLabs in blind tests. The process runs offline, so the voice sample never leaves your computer. Two important caveats: check the license (several cloning models, including XTTS v2 and F5-TTS, are non-commercial), and ethically and legally, only clone a voice you have the person's consent to use.
Do I need a powerful GPU for local text-to-speech?
No, not for narration. Lightweight models like Kokoro-82M run in just 2-3GB of VRAM and even work on CPU-only machines, so almost any computer can do local TTS narration. Voice cloning models like XTTS v2 or Chatterbox want a bit more — a GPU with several GB of VRAM makes them faster — but they still run on modest hardware. Unlike large language models, TTS is relatively light, so you don't need an expensive GPU to generate good speech locally.
Local TTS is free, private, and genuinely good — Kokoro for narration, XTTS/Chatterbox for cloning (mind the license and consent). Combine it with local Whisper transcription for a full offline voice pipeline, and see the best free local models for the text side. Source: BentoML.
