Kokoro vs XTTS vs Chatterbox: Best Local TTS in 2026?

For local TTS in 2026, pick by job: Kokoro-82M (Apache 2.0) is the best choice for fast, clean narration and is the only one of the three you can safely use commercially without caveats; Coqui XTTS v2 still produces the most convincing voice clone from a ~6-second sample but its CPML license is non-commercial only; and Resemble AI's Chatterbox (MIT) is the pick when you need emotional, expressive speech and a permissive license with cloning. There is no single winner — Kokoro can't clone voices at all, XTTS v2 can't be used commercially, and Chatterbox is heavier to run. The right answer depends entirely on whether you need narration, an exact voice clone, or expressive emotion, and whether the output is commercial.

If you only remember one thing: license and use case decide this, not raw audio quality. All three sound good. The difference that will actually bite you is whether you're allowed to ship the output and whether you need to clone a specific voice.

TL;DR — which local TTS should I use?

• Want clean narration / audiobook / app voice, and you might sell it? Use Kokoro-82M. Apache 2.0, tiny (82M params), runs faster than real time on a CPU, no GPU required.
• Need to clone a specific person's voice and it's a personal / non-commercial project? Use XTTS v2. Best zero-shot clone from ~6 seconds of audio, 17 languages — but the license forbids commercial use.
• Need emotional, expressive speech and a clean commercial license? Use Chatterbox (MIT). Zero-shot cloning from ~5 seconds plus a unique emotion-exaggeration dial.

What are Kokoro, XTTS v2, and Chatterbox?

These are the three most-recommended open-weight, run-it-yourself TTS models in mid-2026. They solve different problems, so comparing them on a single "best" axis is misleading. Here's the one-line identity of each:

• Kokoro-82M — an open-weight model from the developer "hexgrad," built on the StyleTTS 2 + iSTFTNet architecture. At just 82 million parameters it's tiny by TTS standards. Its v1.0 shipped on January 27, 2025 with 54 voices across 8 languages. It is trained on long-form narration and reading, has no voice-cloning ability, and is released under the permissive Apache 2.0 license.
• Coqui XTTS v2 — the multilingual voice-cloning model from Coqui. It clones a voice from roughly a 6-second reference clip across 17 languages. The catch: Coqui Inc. shut down in January 2024, and XTTS v2 ships under the Coqui Public Model License (CPML), which is non-commercial. With the company gone, there is effectively no one to sell you a commercial license. See our full XTTS v2 (Coqui TTS) setup and benchmark guide for install steps, VRAM/RTF numbers and quality tests.
• Chatterbox — Resemble AI's open-source TTS, released under the MIT license. It's built on a 0.5B-parameter Llama backbone, does zero-shot voice cloning from ~5 seconds of audio, and is notable for being one of the first open TTS models with an explicit emotion-exaggeration control. Resemble has reported it beating ElevenLabs in side-by-side listener preference tests.

If you want the wider field beyond these three, our roundup of the best local TTS models covers Piper, F5-TTS and others alongside these — and if fast English voice cloning is the goal, the F5-TTS setup guide shows how to run the strongest research-grade cloner locally.

Kokoro vs XTTS vs Chatterbox: full comparison table

Here is the head-to-head. Figures are taken from each model's official model card, repository, or vendor documentation; parameter counts and licenses are confirmed against the official sources.

A few honest caveats on that table. XTTS v2's exact public parameter count isn't headlined the way Kokoro's "82M" or Chatterbox's "0.5B" are, so we've left it as an approximate GPT-style scale rather than invent a precise number. Chatterbox's language count depends on which release you pull: the original English-first model versus the later Chatterbox Multilingual, which covers 23 languages. And "emotion control" for XTTS means you can nudge tone through reference-clip choice, not a dedicated dial like Chatterbox's.

Which has the best license? (This is the real decision)

If your output will ever be sold, monetized, or used in a product, license is the first filter — and it eliminates one model outright.

• Kokoro — Apache 2.0. The most permissive of the three. You can use it commercially, modify it, and embed it in closed-source products without paying or asking.
• Chatterbox — MIT. Also fully commercial-friendly. MIT and Apache 2.0 are both safe for shipping products; the practical difference is negligible for most users.
• XTTS v2 — CPML, non-commercial. The Coqui Public Model License restricts the model to non-commercial use. Because Coqui Inc. closed in January 2024, there's no remaining entity to grant you a paid commercial license. Treat XTTS v2 as personal / research / non-commercial only.

This single fact reshapes the whole comparison: XTTS v2 may produce the best clone, but you can't legally ship it in a paid product. For the full breakdown of what is and isn't allowed, see our deep dive on the XTTS / Coqui commercial license.

Which has the best voice cloning?

If your goal is "make it sound like this specific person," Kokoro is out — it has no cloning at all, only its 54 fixed voices. That leaves XTTS v2 vs Chatterbox.

• XTTS v2 is still widely regarded as the most faithful zero-shot cloner from a short (~6-second) sample, particularly for matching timbre and accent across its 17 languages — that is the consensus across the open-TTS community rather than a number anyone can hand you, because zero-shot clone fidelity has no agreed benchmark. The blocker is legal, not technical.
• Chatterbox clones from ~5 seconds and is very close on fidelity, while adding expressiveness XTTS lacks. For any commercial cloning work, Chatterbox is the answer because XTTS can't be used commercially.

So the practical rule: non-commercial clone → XTTS v2; commercial clone → Chatterbox. Our step-by-step Chatterbox TTS setup guide walks through installing it and dialing in a cloned voice.

Which is best for emotion and expressiveness?

This is Chatterbox's standout. It exposes an exaggeration parameter that scales emotional intensity:

• 0.0 — flat, monotone
• 0.5 — natural, conversational (the default)
• 1.0+ — dramatic, theatrical

Paired with a CFG/pacing control, this lets you tune delivery without re-recording, which is genuinely useful for character voices, ads, or expressive narration. Resemble AI has reported Chatterbox being preferred over ElevenLabs in listener A/B tests — take vendor benchmarks with a grain of salt, but it signals the model is competitive at the top end.

Kokoro and XTTS, by contrast, give you the emotional tone baked into their voices or reference clips; there's no equivalent intensity dial.

How fast are they, and what hardware do you need?

Footprint is arithmetic, so you can work it out rather than trust anyone's benchmark. Model weights in memory are roughly parameters x bytes-per-weight — 2 bytes at FP16, 4 at FP32 — and then you add headroom for the vocoder, activations and audio buffers on top.

That 165 MB figure is the entire reason Kokoro behaves differently from the other two. It is roughly six times smaller than Chatterbox, which puts it in a different class of hardware requirement altogether: it does not need to fit in VRAM, because it fits comfortably in ordinary system RAM alongside everything else you are running.

The practical consequence: Kokoro is the only one of the three designed to be genuinely happy on CPU, which is what the Kokoro-82M model card and the project's own positioning emphasise — a small StyleTTS 2-based model built for cheap, fast inference rather than for cloning. XTTS v2 and Chatterbox are both autoregressive models several times larger, and both want a GPU to feel responsive.

Do not take our word for the throughput on your machine. Real-time factor for TTS depends on your CPU, the sentence length, and the phonemizer, and none of those transfer between machines. Time it yourself: generate one paragraph, divide the resulting audio duration by the wall-clock time it took, and you have your real-time factor. Anything above 1 is faster than real time. The Kokoro TTS local setup guide walks through the install if you want to try it on your own hardware.

Decision tree: pick your TTS by use case

Walk down this list and stop at the first match:

1. Are you selling or monetizing the output (commercial use)?
   • Yes → XTTS v2 is out (non-commercial). Continue to step 2.
   • No → all three are on the table; use the cloning/narration questions below.
2. Do you need to clone a specific voice?
   • Yes, commercial → Chatterbox (MIT, clones from ~5s, plus emotion).
   • Yes, non-commercial → XTTS v2 (best clone fidelity, 17 languages).
   • No, any prebuilt voice is fine → go to step 3.
3. Do you need expressive / emotional delivery?
   • Yes → Chatterbox (exaggeration dial).
   • No, just clean narration → Kokoro-82M (fast, light, Apache 2.0, runs on CPU).
4. Tight on hardware (no GPU)?
   • Pick Kokoro regardless — it's the only one that's genuinely happy on CPU.

In short: Kokoro for narration, XTTS for non-commercial clones, Chatterbox for commercial clones and emotion.

Key Takeaways

1. There is no universal winner. Kokoro, XTTS v2, and Chatterbox each win a different job — narration, best clone, and emotion/commercial clone respectively.
2. License is the decisive filter. Kokoro (Apache 2.0) and Chatterbox (MIT) are commercial-safe; XTTS v2 (CPML) is non-commercial only, and Coqui's shutdown means no commercial license is available.
3. Kokoro can't clone voices. It ships 54 fixed voices across 8 languages and is built for narration, not impersonation — but it's tiny (82M) and runs faster than real time on CPU.
4. XTTS v2 still clones best from a ~6-second sample across 17 languages, which is why it survives despite the dead company and restrictive license — for personal projects only.
5. Chatterbox is the expressive, commercial-friendly clone. MIT-licensed, ~5-second cloning, and a unique emotion-exaggeration dial (0.0–1.0+).

Next Steps

• Want the full field, not just these three? Read Best Local TTS Models for Piper, F5-TTS and more, with install notes.
• Ready to set up the most flexible commercial option? Follow the Chatterbox TTS setup guide step by step.
• Need to be sure you're allowed to ship XTTS output? Read the XTTS / Coqui commercial license breakdown before you build anything on it.
• Verify the specs yourself on the official model pages: the Kokoro-82M model card and the Chatterbox GitHub repo.