MiniMax H3 Speech Decode / 仅解码语音 (EXP/T8)

If you're building a speech-only workflow with the T8mars H3 pack, you don't want to decode a full joint AV latent just to hear the line. MiniMaxH3SpeechDecodeT8 is the slim node for exactly that: it decodes only the H3 audio latent through the audio VAE and hands you an AUDIO object. No picture, no GetVideoComponents overhead, no waste.

It slots into the speech pipeline right after the sampler: plan → conditioning → sampler → decode (audio only) → assemble → release. For a voiceover or ADR session where the picture is irrelevant, this is the decode you want.

The inputs

• av_latent - the sampled latent from your H3 sampler.
• audio_vae - the H3 audio VAE. If you mixed up video_vae and audio_vae here, you'll get garbage or an error; the audio VAE is the one shaped for the [B,32,2,T] audio latent.
• trim_mode - none by default; conservative_energy is the optional trim.
• energy_threshold_dbfs (-50) and trim_padding_seconds (0.1) - only used when trimming, and only worth touching if your decoder is leaving silence around the line.

The output is the decoded audio plus a report_json.

The honest caveat

The tooltip spells it out: a conservative energy trim is not a substitute for ASR or speaker verification. Trimming by energy can cut a soft consonant or leave noise that "sounds like" silence. If you need to know the line was actually spoken correctly, run it through a transcription check (the pack's speech tools use a fixed Faster Whisper for exactly this) rather than trusting the trim.

Where it sits

This is one of the simpler nodes in the speech family, and its value is mostly architectural: it keeps the speech path audio-only so you can chain many segments cheaply, and it feeds MiniMaxH3SpeechAssembleT8 (which stitches turns on exact sample boundaries) or a save node. The reference chain is in examples/workflows/05-speech-dialogue/ - e.g. 2026-08-10_H3_Speech_Described_Stock20_EXP.json.

Install

Pack standard - ComfyUI Manager search "MiniMax H3 Audio T8", or

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

then restart ComfyUI. No extra pip packages, no model downloads; you need the audio VAE weights in your model folder as usual.

When to use something else

If you're making a video and want the joint AV decode with synchronized picture, use the pack's full AV decode node (MiniMaxH3AVDecodeT8) instead - this one deliberately throws the picture away. And if you want the decoded line glued back onto a video timeline, decode with this node for the speech work, but keep the video path separate until final assembly. For pure speech production - voiceover, dubbing, ADR takes - this is the right tool and it's about as simple as H3 speech gets.