🗣️ VRGDG_LoadAudioSplit_HUMO_TranscribeV2

VRGDG_LoadAudioSplit_HUMO_TranscribeV2 is the "one scene at a time" version of the transcription splitter, and it's built to ride the chunked queue system that keeps long video runs from blowing up your VRAM. Instead of splitting the whole track and transcribing everything at once, it processes one set - a batch of scenes - per queue step, keyed by set_index. That's the design that lets a 20-scene music video run as a series of small, resumable jobs rather than one monster.

It also adds two things the first version lacked: context prompts and scene overlap. Both exist because in practice, HUMO and LTX scene generation want more than just "here's audio" - they want direction about what's happening in the scene, and they want a little bleed between clips so the video doesn't hard-cut.

What it does

Inputs that matter:

• audio (AUDIO) - the loaded track.
• set_index (INT, default 0) - which batch of scenes this queue step handles. This is the node's hook into the chunked pipeline: bump it per run to process the next set.
• language (enum, default english) - Whisper's transcription language.
• enable_lyrics (BOOLEAN, default true) - transcription on by default here (note: V1 defaults it to off).
• overlap_lyric_seconds (FLOAT, default 0) - how much the scene clips overlap. A small overlap keeps motion continuous across scene boundaries - audio-driven video benefits from not hard-cutting at a lyric line.
• fallback_words (STRING, default thinking,walking,sitting) - comma-separated captions used when transcription misses or a scene has no clear vocal content.
• context_1 .. context_16 (STRING, optional, multiline) - per-scene context for the LLM. These let you tell the generator what's happening in each scene ("she's standing at a window, rain outside") so the video matches your vision rather than the model's guess.

Outputs: meta (DICT), total_duration (FLOAT), lyrics_string (STRING), and audio_1 .. audio_16 (AUDIO) - note the cap is 16 clips per set here, tighter than the 50 on the batch splitters, because each set is meant to be a manageable chunk.

How to drive it

Pair it with a loop that advances set_index. Run the queue with set_index 0, it processes the first batch of scenes and hands back their audio and lyrics. Bump to 1, next batch. The context_* inputs are where you spend most of your actual effort as a "director" - a well-written context for each scene does more for the result than any of the numeric knobs. And if scenes are cutting mid-word, raise overlap_lyric_seconds a bit.

Installing it

Part of the comfyui-vrgamedevgirl pack. ComfyUI Manager: search "vrgamedev", install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/vrgamegirl19/comfyui-vrgamedevgirl

then install the README's requirements:

pip install -r custom_nodes/comfyui-vrgamedevgirl/requirements.txt

If it's not working

• Only the first batch ever runs - you're not advancing set_index. Wire it to the pipeline's set counter, or bump it manually per run.
• Scenes hard-cut and look jumpy - raise overlap_lyric_seconds from 0.
• Silent or mumble scenes get nonsense captions - that's what fallback_words is for; put sensible default actions in there.
• Transcription errors - same story as V1: Whisper needs to be present in your environment, and the README's short requirements list doesn't mention it explicitly.