Qwen TTS Multi Speaker Node

If you've been piecing together multi-character TTS by generating each line separately and stitching audio files in an editor, this node is the thing that replaces that whole chore. You paste a script with [speaker] tags, connect a dict of voices, and it renders the entire dialogue as one audio file - with per-line gaps, per-line volume, and optional background music with ducking. It's a tiny "audiobook mixer" hiding inside a node pack.

The workflow to get here: build voices with the Voice Clone Prompt node, save them, load them back through a Speaker Bank (which outputs a DICT mapping names → voice prompts), then connect that dict here. The multi-speaker node then walks your script line by line.

The dialogue format

dialogue_text is a multiline script. Every block starts with a tag line, then the spoken lines follow:

[Narrator]It was a cold morning in the valley.
[Alice,400,1.0,overlap]I hate the cold.
[Bob,600,0.8]Well, at least it's quiet.

The comma parts after the name are optional and in this order:

1. gap in ms - silence before this character speaks (second value).
2. volume - 0.8 is quieter, 1.5 louder (third value).
3. mode - overlap (default) lets the previous line's tail bleed into this one instead of hard-cutting. There's a crossfade path in the code for smooth transitions between speakers.

Speaker names must match the keys in the speaker_prompts dict. A typo raises a clear "Speaker 'x' not found" error rather than silently dropping the line.

The inputs that matter

• model - the base_model from QwenTTSModelDownloader (this is the plain TTS model, not the clone or design one).
• speaker_prompts - the DICT from a Speaker Bank.
• dialogue_text - your script, above.
• language - per-model language choice; Auto usually does fine for mixed content.

The audio-post section is where this node punches above its weight:

• background_music - optional AUDIO that plays under the dialogue.
• music_volume, music_loop, music_trim - master music level, loop the track to cover long dialogue, and trim it to the final speech length.
• ducking - the music automatically dips while someone talks and swells between lines. ducking_strength (0.05–0.9) is how much it attenuates during speech - the tooltip says it's relative to the current music volume. ducking_fade_ms smooths the transitions so it doesn't pump.
• soft_limiter - a gentle tanh limiter that catches clipping when several speakers and music stack. If your mix sounds like it's peaking, flip it on instead of hand-fixing levels.
• The usual sampling dials (temperature, top_p, top_k, repetition_penalty, max_new_tokens_per_line, seed) apply per line.

Output is a single AUDIO object. Save it with any audio saver and your voiceover is done.

Install

Part of Vantage-Nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/vantagewithai/Vantage-Nodes.git
pip install -r requirements.txt

or ComfyUI Manager → search "Vantage-Nodes" → Install → restart. Download the Qwen3 TTS models with QwenTTSModelDownloader first.

Common issues

• "Speaker 'x' not found" - the dict key and script tag don't match, or the voice was never saved/loaded. Case-insensitive matching helps, but exact names are safer.
• Music swallows the dialogue - you left ducking off and music_volume at 0.4. Turn on ducking, or drop music_volume.
• Long scripts run very long - this generates and mixes line by line. Budget real time for a 2-minute script. It does use a progress bar, so you're not staring at a frozen screen.