Kokoro vs Dia vs XTTS — Open TTS
Self-hostable voice generation: three open TTS models that cost a fraction of ElevenLabs.
Verdict
Kokoro for speed and permissive licensing, Dia for English dialogue cloning, XTTS for multilingual cloning.
Side-by-Side Comparison
| # | Model | Naturalness Elo | TTFB | Cost / 1M Chars | Voices | License |
|---|---|---|---|---|---|---|
| 16 | Kokoro 82M Hexgrad | 1155 | 60ms | $3.00 | 54 / 9 langs | Apache-2.0 |
| 15 | Dia 1.6B Nari Labs | 1158 | 300ms | $8.00 | Voice clone / EN | Open Weights |
| 19 | XTTS v2 Coqui | 1140 | 260ms | $5.00 | Voice clone / 17 langs | Coqui CPML |
Model-by-Model
Kokoro 82M — Hexgrad
Kokoro 82M is shockingly capable for its size: 1155 Elo, 60ms TTFB, Apache-2.0 license, $3/1M chars — the open TTS default.
Dia 1.6B — Nari Labs
Dia 1.6B (1158 Elo) specializes in dialogue and voice cloning for English at $8/1M chars under open weights.
XTTS v2 — Coqui
XTTS v2 (1140 Elo) pioneered open voice cloning across 17 languages — older but still the most flexible multilingual cloner at $5/1M chars (CPML license).
FAQ
What is the best open-source TTS model?
Kokoro 82M — Apache-2.0 licensed, 60ms latency, 1155 Arena Elo, runnable on a laptop. For cloning specifically, Dia and XTTS are stronger.
Which open TTS is cheapest?
Kokoro at ~$3/1M characters via hosted APIs, or effectively free self-hosted.
Can these do voice cloning?
Dia and XTTS support zero-shot voice cloning; Kokoro does not clone — it ships 54 fixed voices.