Higgs v3 Voice Clone

This is the node people actually come to the Higgs v3 pack for. Higgs v3 Voice Clone takes one reference clip and synthesizes new speech in that voice, zero-shot - no training, no fine-tuning, just a AUDIO input and a prompt. In the wider ComfyUI ecosystem, which has historically been image-first and is only now getting serious local TTS options, this is the moment the graph starts replacing your cloud TTS. If you've been fighting Qwen3-TTS clones that drift or flatten, Higgs v3 is the current community recommendation, and this node is how you run it inside ComfyUI.

Here's the honest part: the whole game is the reference. The node internally trims the clip (silence threshold −42 dB, max 100 seconds), then conditions the model on the cleaned audio. But the single biggest lever is reference_text - an exact transcript of the reference clip. The README says it flat out: a correct transcript materially improves cloning. Feed it a clip with no transcript and you'll get a voice that's recognizably the same but weaker and less stable. Feed it an exact transcript and the clone snaps into focus.

The inputs that matter

• reference_audio (required) - clean speech, little music or noise. Same clip reused for every longform chunk, so the voice stays consistent across a whole narration.
• reference_text - the transcript, and the field that separates a good clone from a mediocre one. This is where the pack's Whisper node feeds in (see Higgs v3 Whisper Transcribe), though a manually corrected transcript beats raw Whisper output.
• text - what you want spoken in that voice. Inline tags work here too: <|emotion:amusement|>, <|prosody:pause|>, <|sfx:sigh|>Ahh....
• Everything else is the shared generation set - max_new_tokens (2048 ≈ 25–30s), temperature (0.8–1.1 natural), top_p, top_k, seed, longform_chunking, words_per_chunk, pause_between_chunks. Same defaults, same behavior as Generate.

Output is audio (AUDIO), straight into Preview Audio or Save Audio.

The emotion-tag trap

Strong emotions at the start of the prompt can overpower the speaker conditioning and make the clone drift - the README specifically flags <|emotion:sadness|> first-token as a culprit, while milder tags like amusement usually survive. This is a known model limitation, not the node picking a random speaker. The workaround: let Higgs establish the cloned voice with at least one word first, then tag.

This <|emotion:sadness|>is a short test sentence to test the text to speech.

And put the next word directly against the tag with no space: <|emotion:sadness|>is, not <|emotion:sadness|> is.

Troubleshooting, grounded

Weak clone? Clean up the reference (cut music/noise), then fix the transcript. Whisper helps, but a corrected one is better - this is the fix people skip. Long script cutting off mid-sentence? Keep longform_chunking=True and words_per_chunk around 35–55. And if you're comparing takes, use a fixed seed and the same settings - the README gives a concrete recipe (fixed seed 12345, temp 0.8, top_p 1.0, top_k 50, chunking off) for verifying delivery controls like speed.

Two footnotes before you go clone your favorite podcaster: Boson releases this for research and non-commercial use, and the license is explicit that you don't clone voices without consent. This is a genuinely impressive model - people call it "incredible" when it lands - but it's also the kind of power you keep on the right side of.