MiniMax H3 Voice Profile / 音色资料 (EXP/T8)

MiniMax H3 generates video with native audio, and its dialogue actually listens to a reference voice - which means the first thing you need is a voice it can be told to use. That's this node. MiniMaxH3VoiceProfileT8 builds a reusable "voice profile" - a workflow-memory object that either describes a voice in text or wraps a real reference clip, so your prompts can consistently say "this character sounds like that."

H3's audio story is worth knowing before you touch this: unlike the bolted-on audio passes of earlier video models, H3 treats audio as part of the joint generation. Speech, ambience and room tone come out of the same Transformer. That makes reference audio genuinely useful for consistency - and it also means you should treat voice matching as a soft conditioning signal, not a phoneme-perfect guarantee. This node's job is to package the reference so the rest of the chain can use it consistently.

Two modes

voice_mode defaults to described_voice: you type a voice_description (the default is a warm, natural, conversational adult speaker) and H3 does its best to render it. The other option is reference_voice, where you connect reference_audio and the node prepares a 2–15 second, 32kHz stereo anchor clip. That anchor is what gets referenced during generation. Reference mode requires the rights_confirmed toggle - the author is explicit that you need consent or another lawful right to use the connected person's voice, and this node won't let you build a reference profile without ticking it. That's a sensible guardrail; don't fight it, and don't upload voices you don't own.

The prep knobs that matter

reference_start_seconds and reference_duration_seconds (0 = "up to the first available 15s after start") pick the slice. highpass_60hz (default on) cuts rumble, and peak_limit_minus_3_dbfs only attenuates peaks above −3 dBFS - the tooltip is careful to say it never boosts a quiet reference. speaker_id and language are for labeling and routing.

Outputs:

voice_profile (wire to a Voice Library Save or into the speech workflow), prepared_reference_audio (the cleaned anchor), and the pack's usual report_json.

Installing:

this is part of MiniMax H3 Audio T8 - ComfyUI Manager, search "MiniMax H3 Audio T8", or:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8

Restart after. No mandatory pip deps; H3 weights and VAE are yours to supply.

The honest limits

A reference profile makes the voice consistent, not identical - H3's joint generation can still drift, and the pack's own notes say exact lip-sync / voice fidelity for broadcast work needs dedicated lip/voice retargeting afterward. Also, the described-voice mode is surprisingly decent for a stock narration voice, so don't assume you need a recording. And keep the license reality in view: H3 local weights are restricted to users outside the US/EU/UK/Korea, which gates everything in this pack.