Qwen3-TTS in ComfyUI: setup and multi-speaker dialogue

This tutorial walks through Qwen3-TTS in ComfyUI, with installation, voice design and a multi-speaker dialogue workflow. The presenter clones the custom-node repository, installs its Python requirements in the active environment and downloads model files before running the workflow to reduce the risk of incomplete automatic downloads.

The setup covers a Windows SoX detection problem: installing Python requirements may leave the audio executable unavailable inside the ComfyUI environment. The presenter also encounters missing components on the first run, when the node downloads the separate 12 Hz tokenizer and Voice Design weights.

For this local AI workflow, the speaker gives approximate VRAM requirements of 2 GB for the 0.6B model and 8 GB for 1.7B. These are stated requirements, not hardware benchmarks demonstrated in the tutorial. The interface includes CUDA, CPU and MPS device choices, and the presenter identifies Voice Design as a 1.7B option.

Examples use text descriptions to control age, accent and delivery, followed by a passage mixing four languages. The final workflow assigns three voices to a role bank, processes dialogue line by line and joins the audio. Saved generated voices can also become reference audio for later scenes. Voice design and reference-based cloning have separate roles in the demonstrated setup.