Which local TTS model sounds the most natural?

By TTS Arena's crowd-blind vote, the largest independent sample of the three, Kokoro holds its own against far bigger models. For cloned voices specifically, most hands-on reviewers still rate XTTS v2's raw clone as the most convincing, with Chatterbox close behind.

Can I use XTTS commercially?

Not under its default Coqui Public Model License, which is non-commercial only. Chatterbox (MIT) and Kokoro (Apache 2.0) both allow commercial use without a separate agreement.

Is Chatterbox actually better than ElevenLabs?

In Resemble AI's own Podonos-run blind test, 63.75% of evaluators preferred Chatterbox over ElevenLabs. That's a real, methodologically documented result, and also a company grading its own homework, so treat the number as a strong signal, not neutral proof.

Which of the three needs the least hardware?

Kokoro, by a wide margin. It runs faster than real time on CPU alone; Chatterbox and XTTS both want a real GPU to feel responsive.

Kokoro vs Chatterbox vs XTTS: Best Local TTS in 2026

There's no universal winner here, and anyone who hands you one skipped reading a license. Kokoro can't clone a voice, at all. XTTS clones one beautifully and won't let you ship the result commercially. Chatterbox does both, cloning and a commercial license, but wants real GPU time Kokoro never asks for. Pick by whichever of those three sentences eliminates your project first.

By the numbers

Kokoro: 82M parameters, Apache 2.0, no cloning, roughly Elo 1056-1057 with a 54.4% win rate across 5,368 TTS Arena match-ups (TTS Arena leaderboard)
Chatterbox: roughly 0.5B parameters on a Llama backbone (MIT), clones from about 5-10 seconds of reference audio; in Resemble's own Podonos-run blind test against ElevenLabs, 63.75% of evaluators preferred Chatterbox (27.5% preferred ElevenLabs, 8.75% had no preference) (Resemble AI)
XTTS v2: Coqui Public Model License (non-commercial only), clones from about 6 seconds of reference audio, 17 languages, actively maintained today as the idiap/coqui-ai-TTS fork after the original company shut down in January 2024
Kokoro is the only one of the three you can drop into a commercial product without a second read of the license

	Kokoro-82M	Chatterbox	XTTS v2
Parameters	82M	~0.5B (Llama backbone)	Hundreds of millions (undisclosed)
License	Apache 2.0	MIT	Coqui Public Model License (non-commercial)
Voice cloning	No (54 fixed presets)	Yes, ~5-10s reference clip	Yes, ~6s reference clip
Languages	8	English-first; 23+ in the multilingual variant	17
Typical footprint	Under 1GB, CPU-friendly	Wants a real GPU for responsive use	4-6GB VRAM, wants a GPU
Best for	Fast, legally clean narration	Commercial cloning with emotional control	Best raw clone quality, non-commercial only

Voice cloning: only two of these even compete

Kokoro sits this one out entirely, so the real contest is Chatterbox against XTTS v2. Both clone from a few seconds of reference audio with no fine-tuning step. XTTS still gets named most often as the more convincing raw clone by hands-on reviewers, likely because Coqui's team spent years on nothing else. Chatterbox closes most of that gap and adds something XTTS doesn't have: an emotion-exaggeration dial you can turn from flat to theatrical, plus newer Turbo and Nano variants built for lower latency and lighter hardware. If cloning is the actual reason you're reading this, the dedicated XTTS setup guide goes deep on the part this comparison only sketches.

TTS Arena runs a public, crowd-sourced blind A/B vote and turns the results into an Elo-style score — the closest thing this category has to a neutral referee. Kokoro sharing that board at Elo 1056-1057 against systems many times its size, across more than five thousand match-ups, isn't a fluke result from a lucky clip. It briefly held #1 outright in January 2026. Chatterbox's 63.75% number is a different kind of evidence. It's real, it's methodologically documented down to the reference-clip length, and it's still Resemble grading its own homework rather than a blind crowd panel. Both numbers are worth knowing. Neither should be read as more than what it actually measured.

License is the axe that decides this, not audio quality

Apache 2.0 and MIT both clear you for commercial use with zero extra steps: that's Kokoro and Chatterbox. Coqui's CPML explicitly forbids commercial use of XTTS without a separate agreement, which rules it out the moment a project has a business model attached, no matter how good the clone sounds next to the other two. I'd love to tell you audio quality settles this comparison. It doesn't. The license does, for most people, before quality ever gets a vote.

Hardware reality check, on my own rig

Kokoro runs fine on CPU with nothing plugged into the GPU slot. Chatterbox and XTTS both want real GPU time to feel responsive rather than laggy. On my 16GB RTX 4080 both fit with room to spare, but if your only machine is a laptop with no discrete GPU, that's the filter that actually matters, not the Elo score two paragraphs up.

Our pick

Shipping commercially and don't need a specific cloned voice: Kokoro. Need to clone a specific voice and still ship commercially: Chatterbox. Need the single most convincing clone available and the project stays personal or research-only: XTTS v2. That's the whole decision tree. Audio quality is the tiebreaker, not the first filter.

How I picked these numbers

License terms, parameter counts, and cloning requirements are pulled directly from each project's own GitHub or Hugging Face page. The blind-naturalness data leans on TTS Arena's public leaderboard for Kokoro, since a five-figure crowd sample beats any single listening session I could run myself, and on Resemble's own published Podonos results for the Chatterbox figure, flagged clearly above as a vendor-run test, not an independent one.

Whichever one you pick, the CPU-only default is Kokoro, which is also why it gets a full setup guide of its own. The other two deserve real project time before you commit to either, starting with XTTS's actual cloning workflow.