Zyphra has launched ZONOS3-Mini, an open-source text-to-speech model featuring 100 million active parameters. The company trained the model using 112 NVIDIA H100 GPUs over a period of 2.5 days. Designed for on-device use, the model operates in real time and succeeds the previous ZONOS2 version, which utilized 8 billion parameters.
The model includes identity controls for adjusting age, gender, accent, and voice texture via embedding vectors. Delivery controls manage pace, emotion, and projection through activation steering. Additionally, a separate edit model based on a VoiceCraft recipe allows for sentence modification after generation.
Zyphra tested the model using its ZTTS2-Eval benchmark, which covers 35 languages of clean speech and 20 languages of in-the-wild recordings. In speaker-embedding mode, ZONOS3-Mini achieved a score of 66.2 on the benchmark's Capability Index, surpassing Fish Audio S2 Pro at 61.7 and MOSS-TTS v1.5 at 60.2.
Performance metrics for the clean speech set include a word error rate of 7.79 and a speaker similarity score of 83.4. However, the model's reference-clip mode scored 44.7 on the index. The MSR-UTMOS quality score of 3.07 for ZONOS3-Mini remains lower than the 3.16 recorded by Fish S2 Pro and 3.54 by Qwen3-TTS 1.7B.
