TL;DR: Open-source AI voice models now sound near-human. Kokoro, Chatterbox, Orpheus, Dia, and Fish Speech lead in 2026. Run them free, clone voices, and keep your audio private.
The top five open-source Text-to-Speech LLMs in 2026 are Kokoro, Chatterbox from Resemble AI, Orpheus from Canopy Labs, Dia from Nari Labs, and Fish Speech. All five are text-to-speech models with open weights you can run yourself.
Kokoro is only 82 million parameters and runs on a CPU. Chatterbox clones a voice from about five seconds of audio. The best open models now score around 4.7 MOS, close to paid tools like ElevenLabs.
Our picks use 2026 model cards from each lab, the public TTS Arena leaderboard, and reported benchmarks. For each tool we cover what it does best, its size, the license, and how to get started.
We focus on open-weight speech models, not closed APIs like ElevenLabs or OpenAI voices.
Key takeaways
- Kokoro is just 82M parameters and runs on a CPU or a 3GB GPU.
- Chatterbox clones a voice from about 5 seconds of audio and adds emotion control across 23 languages.
- Orpheus is built on Llama and ships in sizes from 150M to 3B for expressive speech.
- Dia generates realistic multi-speaker dialogue with laughs and other nonverbal sounds.
- The best open model scores about 4.7 MOS, within a hair of ElevenLabs.













