EditorNodesPricingBlog

ZONOS3-Mini & Tiny-Voca: Open Source Voice Cloning AI

November 18, 2026
ZONOS3-Mini & Tiny-Voca: Open Source Voice Cloning AI

Zyphra ZONOS3-Mini (arXiv:2611.16525, Apache 2.0) and n4ze3m Tiny-Voca (Apache 2.0)

Share this post:

ZONOS3-Mini & Tiny-Voca: Open Source Voice Cloning AI

ZONOS3-Mini (Zyphra/ZONOS3-Mini, arXiv:2611.16525) and Tiny-Voca (n4ze3m/tiny-voca) are two November 2026 Apache 2.0 open source text-to-speech releases that cover opposite ends of the filmmaking voice production spectrum. Released on November 9, 2026 and already downloaded more than 114,000 times on Hugging Face, Zyphra's ZONOS3-Mini delivers high fidelity zero shot voice cloning and emotional dialogue across 21 languages via SGLang-Omni and GGUF, while Tiny-Voca (released November 14, 2026) packs a complete English TTS and waveform generator into just 4.5M parameters running 70x faster than real time on 4 CPU threads.

Studio availability. ZONOS3-Mini and Tiny-Voca are not available in AI FILMS Studio. This article covers public open source releases from November 2026. To generate speech or clone voices in the Studio, open the voice workspace to run ElevenLabs Eleven V3 Text to Speech, MiniMax Speech 2.8 HD Text to Speech, or MiniMax Voice Clone, and animate dialogue on the Image to Video page.

ZONOS3-Mini Character Voice (Irish Male Acting Sample)

Zero-shot expressive character voice from ZONOS3-Mini GGUF preset bank (Apache 2.0).

Tiny-Voca 4.5M ONNX CPU Voice (70x Real-Time Female F2)

Generated on 4 CPU threads in ONNX with zero GPU memory required (Apache 2.0).

ZAC-1 Audio Codec Architecture vs 4.5M Parameter CPU ONNX

Previsualization editors and animation directors face two very different voice synthesis tasks during production. Table reads and animatics need instant scratch dialogue generated locally on a laptop without unloading a video diffusion model from GPU VRAM, while final localization and ADR require expressive voice cloning across multiple European and Slavic languages.

  1. ZAC-1 neural codec and 21-language autoregressive loop (ZONOS3-Mini, arXiv:2611.16525). Zyphra's ZONOS3-Mini encodes reference speaker timbre and acoustic prosody through the custom ZAC-1 neural audio codec, then predicts speech tokens in a low latency autoregressive loop. Supporting 21 languages (en, fr, es, de, it, pt, nl, pl, ru, sv, tr, uk, cs, da, fi, el, hu, no, ro, sk), it includes prebuilt character voices (such as Irish, British, American, whispered ASMR, and dramatic announcer presets) and runs natively in SGLang-Omni or quantized GGUF.
  2. 4.5M parameter end to end CPU synthesis (Tiny-Voca). Created by n4ze3m and paired with the custom dizione phonemizer (n4ze3m/dizione), Tiny-Voca trains a compact 4.5 million parameter acoustic and waveform model from scratch on 18 hours of clean synthetic speech (F2 female and M2 male voices). Exported to ONNX, it synthesizes a 10 second dialogue line in roughly 140 milliseconds on four standard CPU cores (70x faster than real time).
  3. Unrestricted Apache 2.0 commercial licensing. Both Zyphra/ZONOS3-Mini and n4ze3m/tiny-voca ship under the Apache 2.0 license, allowing indie filmmakers and game cutscene teams to ship generated dialogue commercially.
Zyphra ZONOS3-Mini ZAC-1 audio codec and autoregressive voice cloning architecture

ZONOS3-Mini architecture combining the ZAC-1 audio codec with autoregressive speech token modeling. Credit: Zyphra (Apache 2.0).

Decoupling acoustic tokenization inside `ZAC-1` lets `ZONOS3-Mini` clone a performer's vocal timbre from a brief 5 to 10 second reference clip while shifting delivery style from whispered tension to energetic projection.

When served through `SGLang-Omni`, the model exposes an OpenAI compatible `/v1/audio/speech` endpoint for automated script-to-animatic batch rendering.

Try ZONOS3-Mini and Tiny-Voca Studio Live in the Browser

Both official Hugging Face Spaces are active and embeddable below without login. Test ZONOS3-Mini zero shot voice cloning in the first Space, or test Tiny-Voca Studio instant CPU synthesis in the second Space.

On the quality versus speed Pareto frontier (`arXiv:2611.16525`), `ZONOS3-Mini` matches or beats significantly larger open weights TTS models on speaker similarity and word error rate while running at a fraction of their latency.

Because `ZONOS3-Mini` also ships a `gguf/` directory alongside WebGPU and `SGLang-Omni` runners, dialogue tracks can be generated on consumer GPUs or Apple Silicon laptops.

ZONOS3-Mini quality versus inference speed Pareto benchmark comparison

ZONOS3-Mini quality vs speed Pareto frontier across open source TTS models. Credit: Zyphra (arXiv:2611.16525) (Apache 2.0).

Model Specs, Languages & Hardware Matrix

Model Hugging Face ID Parameters & Format Languages & Voices Hardware & Speed License
ZONOS3-Mini (arXiv:2611.16525) Zyphra/ZONOS3-Mini Compact AR + ZAC-1 Codec (BF16 & GGUF) 21 languages + zero shot voice cloning 6 GB to 12 GB VRAM (SGLang-Omni / GGUF) Apache 2.0
Tiny-Voca n4ze3m/tiny-voca 4.5M parameters (ONNX + dizione) English (F2 female, M2 male) CPU only (4 threads, 70x real time) Apache 2.0
Fish Audio Drama-3 fishaudio/drama-3 Multi speaker dialogue TTS Multi turn emotional script dialogue 12 GB to 16 GB VRAM Open Weights
Kokoro-82M hexgrad/Kokoro-82M 82M StyleTTS2 ONNX/PyTorch Multi language fast voice presets CPU or 2 GB VRAM Apache 2.0

Running ZONOS3-Mini (SGLang-Omni) and Tiny-Voca Locally

To serve Zyphra/ZONOS3-Mini locally behind an OpenAI compatible endpoint on NVIDIA GPUs, install sglang-omni and launch the server.

# Install SGLang-Omni and download Zyphra/ZONOS3-Mini (Apache 2.0)
git clone https://github.com/sgl-project/sglang-omni.git
cd sglang-omni
pip install -e .

huggingface-cli download Zyphra/ZONOS3-Mini --local-dir ./models/ZONOS3-Mini

# Launch OpenAI-compatible ZONOS3-Mini voice cloning server
python -m sglang_omni.launch_server \
  --model-path ./models/ZONOS3-Mini \
  --port 8000

Query the local server or clone a character voice from one of the bundled gguf/voices/ reference .wav files.

curl http://localhost:8000/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Zyphra/ZONOS3-Mini",
    "input": "We only have thirty seconds before the blast doors seal the corridor.",
    "voice": "Cillian_en-Irish_M_1"
  }' \
  --output dialogue_line.wav

To run Tiny-Voca on a CPU only laptop at 70x real time for storyboard scratch tracks, clone n4ze3m/tiny-voca.

git clone https://github.com/n4ze3m/tiny-voca.git
cd tiny-voca
pip install -r requirements.txt

# Synthesize English scratch dialogue on 4 CPU threads via ONNX
python infer.py \
  --text "Camera two pushes in on the detective as the rain starts." \
  --voice F2 \
  --threads 4 \
  --output scratch_f2.wav

Production Playbook for Filmmakers

  1. Keep GPU VRAM free during animatics with Tiny-Voca. When your GPU is occupied running Wan 2.2, MiniMax-H3, or FLUX.2, run Tiny-Voca on 4 CPU threads to generate timing accurate scratch dialogue for every storyboard panel in milliseconds.
  2. Use ZONOS3-Mini for 21-language dubbing and lip sync drivers. Feed cloned ZONOS3-Mini dialogue .wav tracks directly into audio-driven lip sync models such as LatentSync 1.6 and MuseTalk 1.5, JoyVASA portrait animation, or Kandinsky S2V 1.0.
  3. Compare across open source voice models. Explore our related voice guides on Fish Audio Drama-3 multi-speaker dialogue TTS, Qwen3-TTS, F5-TTS, Chatterbox, and Kokoro voice cloning, and Bark TTS, Voicebox, and InSwapper.

To generate broadcast dialogue or clone voices in the cloud without local setup, open the voice workspace or wire voice nodes directly into lip sync and video nodes inside the Nodes Graph Editor.


Sources