Voice Clone Prompt

VoiceBridge's whole pitch is "translate the speech, keep the speaker's voice." That trick lives here. This node takes a short sample of someone's voice and turns it into a voice clone prompt - an internal description of that voice that Qwen3-TTS then uses as the identity for synthesized speech. It's the difference between a dubbed video in a generic announcer voice and one that sounds like the original person speaking a different language.

How it works

Feed it a reference audio clip and it runs the loaded Qwen3-TTS model's create_voice_clone_prompt routine, which extracts the voice's characteristics (timbre, prosody, the whole fingerprint) into a compact prompt object. That object - the voice_clone_prompt output, a custom VOICE_CLONE_PROMPT type - is passed to SRT To Audio, which uses it for every sentence it synthesizes. You do the clone extraction once, up front, and every subtitle downstream shares the same voice.

Two inputs, and only two:

• model_key - from VoiceBridge TTS Loader. The TTS model must be loaded first; like all the inference nodes in this pack, this one looks the model up from the cache by key.
• ref_audio (AUDIO) - the voice sample, in ComfyUI audio format. Anything with a clean few seconds of the speaker works. This is where quality matters: no background music, no other voices, just the person talking.
• ref_text (STRING) - the transcript of what's in the reference audio. The tooltip and the README both push this: it's "highly recommended for better quality." Qwen3-TTS aligns the reference text to the audio to lock onto the voice more precisely. If you skip it, cloning still works but it's guessing - so if your clone sounds off, adding the transcript is the first fix to try.

Installing it

Part of the comfyui_voicebridge pack. ComfyUI Manager search "VoiceBridge", or:

cd ComfyUI/custom_nodes
git clone https://github.com/YanTianlong-01/comfyui_voicebridge.git
cd comfyui_voicebridge
pip install -r requirements.txt

Restart ComfyUI. The Qwen3-TTS weights (a few GB, auto-downloaded on first TTS load) are what this node actually needs to do its thing.

Honest expectations on cloning quality

Zero-shot voice cloning is "this sounds pretty close" territory, not "indistinguishable" - even the best open models (Chatterbox, and Qwen3-TTS in the same class) produce a credible same-family voice rather than a perfect copy, and the community's bar for open cloning is exactly that. This pack is also brand new with no community feedback yet, so treat the clone quality on your own reference audio as the thing to sanity-check before committing to a long render. If the clone comes out weak, your order of operations is: add ref_text, use a cleaner reference clip, and make sure the reference isn't too short - a few seconds beats one.