Short answer: for a local voice clone you can legally ship, install Chatterbox (pip install chatterbox-tts), feed it a clean ~10-second reference clip, and generate. It is MIT licensed, which almost nothing else at its quality tier is. If your project is personal, research, or a demo, IndexTTS2 and Higgs Audio v3 sound better - but their licenses will stop you the moment money is involved.
That is the whole decision, and it is the part every "best local TTS" listicle skips.
The real constraint is the license, not the audio
In 2026 the open-weights speech field has largely solved naturalness. Half a dozen models will clone a voice from a few seconds of audio and produce something a casual listener will not flag as synthetic. Audio quality is no longer the thing that separates them.
Licensing is. And the pattern is consistent: the labs producing the most expressive models are the ones training on scraped in-the-wild speech, and they hedge that exposure with restrictive licenses. F5-TTS says so directly in its README - the code is MIT, but "the pre-trained models are licensed under the CC-BY-NC license due to the training data Emilia, which is an in-the-wild dataset."
So before you evaluate a single sample, check what you are allowed to do with the output.
| Model | Params | License | Commercial use? |
|---|---|---|---|
| Chatterbox | 110M / 350M / 500M | MIT | Yes |
| Kokoro-82M | 82M | Apache 2.0 | Yes (no cloning) |
| F5-TTS | - | Code MIT, weights CC-BY-NC | No |
| IndexTTS2 | - | bilibili Model Use License | Restricted |
| Higgs Audio v3 | 4B | Research & Non-Commercial | Separate license required |
| Fish Speech | - | Fish Audio Research License | No |
Higgs Audio v3's README is explicit: "Production / hosted / revenue-generating use requires a separate commercial license." Fish Speech goes further and states it "will take action against any violation of the license." These are not ambiguities you get to interpret optimistically.
Chatterbox is the outlier. Resemble AI shipped it under a plain MIT license, and that single fact is why it is the default recommendation here rather than the model with the prettiest demo page.
The setup, start to finish
1. Pick your model size
Chatterbox is a family, not one model, and picking correctly matters more than tuning later:
- Chatterbox-Nano (110M) - English only. Runs on CPU at 3x faster than realtime on 8 cores, per the README. This is the one for on-device work or if you have no GPU at all.
- Chatterbox-Turbo (350M) - English only, built for low-latency voice agents. Its speech-token-to-mel decoder was distilled from 10 steps down to one, so it is markedly cheaper to run than the multilingual model. Supports paralinguistic tags like
[laugh],[cough], and[chuckle]natively. - Chatterbox Multilingual V3 (500M) - 23+ languages. Use it only if you need non-English; the English-specific models are leaner.
There is also a Single Language Pack of six dedicated finetunes (Mandarin, Latam Spanish, Spain Spanish, Brazilian and European Portuguese, Hindi) for when a regional dialect has to be right.
2. Install
pip install chatterbox-tts
That is genuinely it. No Docker, no separate weights download step - the model pulls from Hugging Face on first load.
3. Record the reference clip
This is where most people's output quality is actually decided, and no parameter will rescue a bad sample. Chatterbox's own example uses a 10-second reference:
import torchaudio as ta
from chatterbox.tts_turbo import ChatterboxTurboTTS
model = ChatterboxTurboTTS.from_pretrained(device="cuda")
wav = model.generate(
"Text you want spoken in the cloned voice.",
audio_prompt_path="your_10s_ref_clip.wav",
)
ta.save("output.wav", wav, model.sr)
What to feed it: one speaker, no background music, no room echo, no compression artifacts, and speech delivered in the register you want back. The model copies the style of the reference, not just the timbre - record the clip flat and monotone and you will get flat, monotone output for the next thousand generations.
4. Tune the two knobs that matter
Chatterbox exposes exaggeration and cfg_weight, both defaulting to 0.5, which the README notes "work well for most prompts across all languages." Two documented adjustments are worth knowing:
- If your reference speaker talks fast, drop
cfg_weightto around 0.3 to fix pacing. - For cross-language generation, set
cfg_weightto 0 to stop the reference clip's native accent bleeding into the target language.
Start at the defaults. Change one value at a time.
5. If you do not need cloning at all
A lot of people reach for a voice-cloning model when they only need a good narrator. If any voice will do, Kokoro-82M is Apache 2.0, 82 million parameters, and ships fixed preset voices instead of zero-shot cloning. Piper (pip install piper-tts) is similar in spirit - a fast local engine with pretrained voices that embeds espeak-ng for phonemization.
Both are dramatically lighter than any cloning model and sidestep the consent question entirely. If your use case is an audiobook, a screen reader, or narration for podcast production, start here and only escalate to cloning if a specific voice is the requirement.
The lesson nobody advertises: local does not mean permanent
On 2025-09-05, Microsoft removed the VibeVoice-TTS code from its own repository. The stated reason, still in the README: "After release, we discovered instances where the tool was used in ways inconsistent with the stated intent. Since responsible use of AI is one of Microsoft's guiding principles, we have removed the VibeVoice-TTS code from this repository."
This was not a small model. VibeVoice-TTS synthesized up to 90 minutes of multi-speaker audio with up to 4 distinct speakers in a single pass, and the paper was accepted as an Oral at ICLR 2026. It was pulled anyway. (Microsoft's VibeVoice repo still hosts VibeVoice-Realtime-0.5B, released 2025-12-03, with roughly 300ms first-audible-latency, and VibeVoice-ASR.)
The practical takeaway for anyone building on local models: archive the weights and the inference code the day you commit to a model. "It's open source, I can always get it later" is now demonstrably false. A repo you depend on can be emptied by its maintainer with no notice and no obligation to you.
Consent, and the watermark you cannot turn off
Cloning a voice that is not yours, without that person's explicit permission, is the fastest way to turn a fun weekend project into a legal problem. The reference clip requirement is around ten seconds, which means essentially anyone who has ever appeared on a podcast is clonable. Treat that capability accordingly.
The model authors clearly agree. Every file Chatterbox generates is watermarked with Resemble AI's Perth (Perceptual Threshold) Watermarker - imperceptible neural watermarks the README states "survive MP3 compression, audio editing, and common manipulations while maintaining nearly 100% detection accuracy." It is on by default, and the repo ships a script to extract it.
Read that as a feature, not an obstacle. If you are doing legitimate work - your own voice, a consenting client, a licensed narrator - a durable provenance signal on your audio protects you rather than constrains you.
So which setup should you actually run?
- Shipping a product, need cloning: Chatterbox-Turbo, MIT, English. Multilingual V3 if you need other languages.
- No GPU: Chatterbox-Nano on CPU.
- Narration only, no specific voice: Kokoro-82M or Piper. Lighter, Apache/permissive, no consent question.
- Personal or research project, want the absolute best expressiveness: IndexTTS2 for emotion and duration control, or Higgs Audio v3 for its 100+ language coverage. Both fine for non-commercial use. Neither is shippable without a license conversation.
If you would rather not run any of this yourself, the hosted market solves a different problem at a different price - we compared two of the main options in ElevenLabs vs Murf AI, and the broader category in the best AI audio tools. And if this is going into a client-facing product rather than a side project, Kodeit builds and ships that kind of integration work.
FAQ
Is local AI voice cloning free? The software is. Chatterbox, Kokoro, F5-TTS, and Piper all cost nothing to download and run, and there are no per-character fees like hosted APIs charge. The cost is hardware and setup time. The catch is licensing: free to download does not mean free to use commercially, and F5-TTS, Fish Speech, IndexTTS2, and Higgs Audio v3 all restrict commercial use.
How much audio do I need to clone a voice? Modern zero-shot models need very little. Chatterbox's documented example uses a roughly 10-second reference clip. Quality of that clip matters far more than quantity - a clean 10 seconds beats a noisy 10 minutes.
Can I run voice cloning without a GPU? Yes. Chatterbox-Nano is a 110M-parameter model that shares Turbo's architecture and, per Resemble AI, runs 3x faster than realtime on 8 CPU cores. Kokoro-82M and Piper also run comfortably on CPU, though neither does zero-shot cloning.
Which local voice cloning model sounds best? By their own published benchmarks, IndexTTS2 claims it "outperforms state-of-the-art zero-shot TTS models in terms of word error rate, speaker similarity, and emotional fidelity" (arXiv:2506.21619). Treat any vendor's self-reported benchmark as a starting point, not a verdict - and note that IndexTTS2's license makes the question academic for commercial projects.
Is it legal to clone someone's voice? Cloning your own voice, or a voice you have documented permission to use, is the safe path. Cloning a real person without consent exposes you to publicity-rights, impersonation, and fraud liability that varies by jurisdiction - and platforms increasingly detect synthetic audio. This is not legal advice; if a project depends on the answer, get a lawyer rather than a blog post.
Will these models disappear? Some already have. Microsoft removed the VibeVoice-TTS code from its repository in September 2025 after documented misuse. Archive weights and inference code for anything you depend on.
Sources
This guide is a primary-source review, not a subjective listening test - we did not run a blind audio quality shootout, and we do not claim to have ranked these models by ear. Every factual claim above (parameter counts, license terms, install commands, parameter defaults, release and removal dates) was read directly from each project's own repository, model card, LICENSE file, or paper on the dates linked below. Where a claim is a vendor's own performance assertion, it is attributed as such rather than presented as an independent finding.
Sources consulted:
- resemble-ai/chatterbox - README and LICENSE (MIT)
- SWivid/F5-TTS - README license section
- index-tts/index-tts - README, LICENSE, and arXiv:2506.21619
- boson-ai/higgs-audio - Higgs Audio v3 release notice and license terms
- fishaudio/fish-speech - README license notice
- hexgrad/Kokoro-82M - model card, Apache 2.0
- microsoft/VibeVoice - README changelog, 2025-08-25 through 2026-07-23
- OHF-Voice/piper1-gpl - README
- resemble-ai/perth - watermarking