NTH

ZONOS3-Mini Technical Report

AuthorsGabriel Clark, Sofian Mejjoute, George Close, Beren Millidge

AffiliationsCorresponding authors

November 15, 2026 2 min read
Watch on YouTube
The one-line take

ZONOS3-Mini aims to make multilingual, voice-cloning text-to-speech fast enough to run in real time on everyday devices.

Key results

886M
Total backbone parameters

Total parameter count of ZONOS3-Mini.

100M
Active parameters per token

Parameters activated per token by the latent mixture-of-experts backbone.

15
ZAC-1 frame rate

Audio codec frames per second.

55
ZTTS2-Eval language sets

Clean and in-the-wild language sets used for evaluation.

6.71%
Macro WER

Speaker-embedding system result on ZTTS2-Eval.

81.2
Macro speaker similarity

Speaker-embedding system result on the benchmark's 100-point similarity scale.

What the paper found

ZONOS3-Mini is a compact, zero-shot voice-cloning text-to-speech model built around an 886M-parameter latent mixture-of-experts backbone, with 100M parameters active per token. Its central component, Zyphra Audio Codec 1, encodes speech at 15 Hz using semantic and residual codes, shortening audio sequences while supporting streaming synthesis. The backbone combines Gated DeltaNet layers with global attention, and uses SuperBPE text tokens; post-training with direct preference optimization and group relative policy optimization improves generation reliability. On ZTTS2-Eval, spanning 55 language sets, the speaker-embedding version achieves 6.71% macro word error rate and 81.2 speaker similarity on a 100-point scale. The report compares it with systems including Qwen3-TTS and Fish Audio Drama 3: the latter has lower reported macro WER, but uses a much larger model. ZONOS3-Mini is designed to run in real time on laptops and smartphones, and its model weights, inference software, and training code are released.

Original abstract

ZONOS3-Mini is an ultra-compact, ultra-fast, zero-shot voice-cloning TTS model. The model can run in real time on a laptop or smartphone. For its latency and size, it achieves state-of-the-art accuracy and cloning fidelity. To achieve this we propose a latent MoE backbone with only 886M parameters (100M active per token) and a new audio codec, Zyphra Audio Codec 1 (ZAC-1), which represents 44.1 kHz audio at a 15 Hz frame rate. On the new ZTTS2-Eval benchmark, ZONOS3-Mini reaches a 6.71% macro-average WER over 55 language sets (35 clean and 20 in the wild), which is competitive with larger and slower state-of-the-art models. We release the model weights, inference, and training code.

Read the original paper

More in Speech AI

Browse all 48 papers →
02Speech

Marimba: A System One Model for Speech

Solano Amadori

Marimba reuses a speech recognition model’s features to answer questions about words, speakers, language, and turn endings.

Read analysis
03Speech

VoxMel TTS Technical Report

Wenhao Guan, Jiaxin Ye, Haitao Li, Yifan Duan, Junxi Liu, Wangjin Zhou, Lin Li, Qingyang Hong, Xie Chen

VoxMel generates speech by autoregressively modeling transformed mel-spectrograms, avoiding a learned tokenizer while aiming for high fidelity and faster inference.

Read analysis