ZONOS-Mini

by Zyphra · Model · Voice & audio

promising

Small, openly licensed multilingual speech model with decent output; one tester so far.

Open TTS with zero-shot voice cloning; 800M total, 100M active parameters; controls accent, age and emotion; runs on GPU, Apple Silicon or CPU; about 1.8 GB

AccessOpen weightsNeeds~25 GBDownloadabout 1.8 GBSize0.8B parametersLicenseApache 2.0

Why we say so

Fahd Mirza ran it locally and found English and French, Italian and Spanish speech good and fast, with dates and numbers handled well. He found English and Chinese voice clones somewhat plasticky and called it decent but not perfect. He measured about 25 GB of GPU memory, so a 24 GB card is borderline despite his remark. The Apache 2.0 license and 1.8 GB size come from Zyphra itself and are consistent across its statements.

Evidence

  1. Nov 13
    Open-source multilingual text-to-speech model under Apache 2.0, allowing commercial use
  2. Nov 13
    Mixture-of-experts TTS with 800M total parameters, about 100M active per token; supports 43 languages and voice cloning from a short clip

Cost to run

Alternatives

Made for ZONOS-Mini