Breeze TTS 2 Speaker

If you've only seen the clone/design/direction trio, this node is why the pack is worth more than the sum of its parts. Speaker is a cast member: one named character, ready to be wired into the Multi-Speaker node for a full dialogue. The clever bit is that a single Speaker node is both a clone and a design at once - give it a reference clip and it clones, give it no clip and an instruction and it designs. Same slot, either kind of voice.

Think of it as the casting sheet before the read-through. You build up to 8 of these, name them, give each a voice, and Multi-Speaker handles the actual conversation.

How it works

The node does its work at prepare time, before any generation happens. It loads your reference clip from the ComfyUI input folder (or takes wired audio), and the key convenience is in the transcript handling: leave reference_text empty with auto_transcribe_if_empty on, and it runs the built-in Whisper helper (whisper-large-v3-turbo) on your clip, shows the transcript right on the node, and passes it along. That kills the single biggest voice-cloning failure - a hand-typed wrong transcript - with zero extra wiring.

Pick none in the audio dropdown and the instruction field takes over: that speaker is now a designed voice, anchored to a stable seed so it stays consistent across every turn.

The inputs that matter

• name - what the script uses to reach this character ("Alice: hello"). Case-insensitive, and matching is forgiving about spaces and punctuation, but keep names unique.
• audio - a dropdown of audio files in your ComfyUI input/ folder (click to browse or drag-drop a file onto the node), or none for a designed voice. There's also an optional reference_audio input on the right side - wire a Load Audio node in and it overrides the file dropdown.
• reference_text - exact transcript; leave empty for auto-transcription.
• instruction - required when there's no clip (it designs the voice); with a clip it steers tone/emotion like the Direction node.
• cfg_scale - 1.0 keeps a cloned voice as-is; go around 4.0 for designed voices or strong directions.

The outputs are the interesting part: speaker is the BREEZE_SPEAKER object the Multi-Speaker node consumes, audio is a preview you can listen to, and transcript is the (possibly auto-whispered) transcript as a string - wire it anywhere.

Installing and using it

Same pack, same install:

cd ComfyUI/custom_nodes
git clone https://github.com/saganaki22/ComfyUI-Breeze-TTS-2

or ComfyUI Manager → search "ComfyUI-Breeze-TTS-2", restart, let the model download. Then the flow is: Load Model → a handful of Speaker nodes (each named, each with a clip or instruction) → Multi-Speaker with the script.

Gotchas worth knowing: no audio and no instruction is an instant error - a speaker must be some kind of voice. A reference clip with no transcript (and auto-transcribe off) errors too. And reference clips are capped at 60 seconds, with ~20 s the practical sweet spot - longer clips burn the 2048-token context the model needs for actual speech. Check the console log after prepare: if the auto-transcript looks wrong, fix it before you generate, because a wrong transcript quietly sabotages the clone.