Local Open TTS Models: The 2026 Landscape and Picks

The most important fact about local TTS in 2026 is that the era of "one model does it all" is over: the dominant community pattern is stacking, not picking. Use Kokoro for fast pipeline narration, Chatterbox for premium cloning, and another model for multilingual expressiveness, layered together. And a more important truth: license matters more than demo quality - the best-sounding XTTS v2, F5-TTS, and some Fish Speech variants are non-commercial (research), so you can't ship them in a paid product. In short: stack by task, but for commercial check the license first.

In plain terms: picking a local TTS is like picking kitchen knives. A mandoline (Kokoro = fast narration) and a sashimi knife (Chatterbox = precise cloning) are separate - trying to do everything with one all-purpose knife is the mistake. And some premium knives are stamped "home use only" (non-commercial license), so you can't use them in a restaurant (commercial).

What do you use for fast, lightweight narration?

Kokoro-82M. At 82M parameters and Apache 2.0, it runs in 2-3GB of VRAM, even on CPU, and its speed is overwhelming - about 36x real-time on a Colab T4, 5x on a 32-core CPU, and up to 210x real-time on a 4090. It offers 8 languages and 54 built-in voices. But it can't clone voices - with fixed voices it's ideal for narration, voice agents, and edge deployment. So if your goal is fast, stable read-aloud rather than "clone a specific voice," Kokoro is the default.

What do you use for voice cloning?

Chatterbox, Fish, or Qwen3-TTS for commercial; XTTS v2 for personal. Chatterbox (Resemble AI, MIT, 0.5B) does zero-shot cloning from 5-10s of reference, features emotion-exaggeration control, and fits in ~6GB - but it's English-only and its output is watermarked (PerTh). XTTS v2 is the zero-shot standard, cloning 17 languages from 6s of reference, but its CPML license is non-commercial, so it's for personal and demo use. For commercial multilingual, Fish Speech (verify the variant's license) or Qwen3-TTS (Apache 2.0, clones from 3s) are the practical picks.

Can you trust the "beat ElevenLabs" numbers?

No - most are benchmarks the makers ran themselves. Chatterbox 65%, Voxtral 63%, Fish 66% "preferred over ElevenLabs" are all the model maker's own blind tests. Even the independent TTS Arena V2 ELO moves week to week, and the top six sit within about 13 ELO - effectively a tie. So don't be dazzled by demo reels or self-reported percentages; A/B test on your own text, language, and voice - Korean especially differs from an English demo.

So what's the 2026 local TTS recommendation?

The key is stack by task, and for commercial verify the license first.

• Fast narration and edge: Kokoro-82M (2-3GB, CPU, ultra-fast, no cloning, Apache 2.0).
• Commercial cloning: Chatterbox (English, MIT, watermark) or Fish/Qwen3-TTS (multilingual, commercial - verify the variant license).
• Personal multilingual cloning: XTTS v2 (17 languages, non-commercial). Benchmarks are often self-run, so A/B in your own language.

Related reading: 오픈 음성 클로닝, 우리는 이렇게 운영한다 — 파이프라인 회고, 오픈 음성 클로닝 파이프라인: 직접 써보고 느낀 점과 한계

Related reading: 로컬 음성합성(TTS) 오픈모델, 직접 돌려본 속도·품질·라이선스, 음성 클로닝 오픈모델 VRAM·RAM 요구량 실측

Reference links

• Kokoro-82M (lightweight TTS, Apache)
• Chatterbox (Resemble AI, MIT cloning)
• Coqui XTTS (multilingual zero-shot)
• Fish Speech/OpenAudio (multilingual)
• F5-TTS (research-grade cloning)

Note: figures like speed (36-210x real-time), clone reference length (3-6s), ELO (~1339), and preference rates (65%) are 2026 public and maker-run benchmarks that vary by hardware, text, and language (not permanent; many are self-run). Licenses differ by variant (e.g., Fish Speech variants; Piper moved to a GPL fork in Oct 2025), so check the model card before commercial use. TTS models and licenses move fast, so this is reviewed quarterly.