Skip to content

← The Lab

Two open TTS models worth knowing for voiceovers (ZONOS3-Mini and Phonon), plus Comfy's official skills plugin for coding agents

A curator list of open repos for a research-to-video-to-voice pipeline flagged two TTS models we hadn't listed. ZONOS3-Mini is Zyphra's Apache-2.0 multilingual voice-cloning model that runs on a GPU, a Mac or in the browser. Phonon is Gradium's tiny on-device model. Separately, Comfy's official skills repo turns Comfy Cloud into slash commands for Claude Code.

ZONOS3-Mini (Zyphra). Per Zyphra's blog, it's a sparse mixture-of-experts TTS model with 100M active and 800M total parameters, released under Apache-2.0, with high-fidelity zero-shot voice cloning and voice design. The repo ships one HTTP API across several backends: a CUDA server, an MLX server for Apple Silicon, a C++/ggml build that needs no Python, and an in-browser WebGPU version you can try with nothing installed. A local Voice Studio and a lightweight cloning page come with it. Zyphra also released ZTTS2-Eval (35 languages) and charts ZONOS3-Mini against other open models on latency and throughput. Those are vendor benchmarks.

Phonon (Gradium). It's built to ship inside apps: 31M and 40M models with a Rust runtime and no PyTorch, plus a CLI (uvx ptts --lang en "..." -o speech.wav), a Python package and an npm package for the browser. Gradium's own table claims 20x real time with a 7 ms time-to-first-byte on one MacBook Air core for the 40M model. Licensing matters here. The code is MIT or Apache-2.0, and the English 31M weights are Apache-2.0, but the 40M models (English, French, German, Spanish, Portuguese and multilingual) are CC BY-NC 4.0, so not for commercial work.

Use either for scratch narration, explainers and drafts where a cloud TTS bill adds up. As the curator's post says, check each licence before commercial use, and only clone a voice with the speaker's consent.

Comfy Skills (Comfy-Org). This is the official home of Comfy's agent plugins. The comfy-cloud plugin bundles the hosted Comfy Cloud MCP server and a set of skills: generate image, video, audio and 3D, clone a voice, remove backgrounds, upscale, keep characters consistent across shots, and search models, nodes and workflow templates. In Claude Code, add the Comfy-Org/comfy-skills marketplace, install comfy-cloud, then sign in with /mcp. Generations run on Comfy Cloud, so they use your Comfy Cloud plan. A separate comfy-sdk plugin is guidance-only, for people building apps on the Comfy SDK and Router.

Signals: tracked curators only (@RoundtableSpace's 11-repo list at 11:45 PM MST Nov 16, and @DanKornas on Comfy Skills at 9:26 PM MST Nov 16), so not HOT by our rule.

Skills & repos in our directory

More from the Lab

Topics: ComfyUI · Audio, voice & music · Video MCP servers