Load Voice (QwenTTS)

Load Voice (QwenTTS) is the simplest node in the ComfyUI-QwenTTS pack, and it does exactly one job well: hand you back a voice you already saved, so you never have to re-extract a speaker embedding from the reference audio again. It's the read half of the pack's voice library, paired with Create Voice (QwenTTS), which writes the files.

What it is

Think of it as a loader in the same spirit as a LoRA loader: it points at a saved file, loads it, and outputs a VOICE object - a serialized voice-clone prompt (speaker embedding plus, if you saved it that way, the reference speech-token "code" and transcript). That VOICE plugs straight into the voice input of Voice Clone (QwenTTS) or its Advanced variant. When a VOICE is provided there, you don't need reference_audio or reference_text at all - the clone just speaks in that saved voice.

The saved files are .pt (PyTorch) files. The node lists whatever .pt files live in the default voice library folder (ComfyUI/output/qwen3-tts_voices/) in the voice_name dropdown, so your saved voices show up as a simple pick-list.

The inputs

• voice_name - dropdown populated from the default voice library folder. Pick the saved voice you want.
• custom_path - optional full path to a .pt file anywhere on disk. If set, it overrides voice_name (the tooltip says so explicitly). This is your escape hatch for voices saved to a custom folder via Create Voice's save_path, or for sharing voice files between machines.

Output: one VOICE. With no voice selected and no path, it outputs an empty value - wire it up properly and it just works.

How to get voices into it

The round trip is: Create Voice (QwenTTS) takes your reference_audio + reference_text, builds the speaker prompt, and saves a voice_1.pt (or whatever voice_name you gave it) to ComfyUI/output/qwen3-tts_voices/. Restart or refresh, and it appears in this node's dropdown. Build a small library of characters once, then load them instantly for the life of your project - that's the whole value proposition, and it's a genuinely nice workflow for multi-character narration.

Install

Same pack, same story - ComfyUI Manager (search ComfyUI-QwenTTS) or:

cd ComfyUI/custom_nodes && git clone https://github.com/1038lab/ComfyUI-QwenTTS.git
python3 -m pip install -r ComfyUI/custom_nodes/ComfyUI-QwenTTS/requirements.txt

Restart, done. This node loads no models by itself (the heavy Qwen models only spin up when the clone node runs), so it's instant and VRAM-free. The only thing that trips people up is expecting the dropdown to refresh mid-session - if a freshly-saved voice doesn't show, the dropdown is populated at node-creation time, so reload the page or recreate the node.