CoachBate Load H3 Audio Path

A loader you can drive from text

ComfyUI's stock audio loaders want you to click a file. That's fine when you're making one thing and useless when you're making forty of them from a shotlist, because the path is data - it lives in a JSON file, it changes per shot, and no dialog is going to appear at 3am with the right answer.

CoachBate Load H3 Audio Path takes an audio path as a string, with an optional in/out range in seconds, and returns an AUDIO you can feed to H3's reference-audio slots. That's why it exists: the shotlist carries reference_audio_1-style paths, and this is the node that turns one into something a model can hear.

Path resolution, since that's the whole node

audio_path is the only thing that matters and the rules are simple once you know them:

• Blank returns no reference audio at all - the node outputs None and doesn't try to decode, which is what you want for the shots that don't use a reference.
• A relative path resolves under ComfyUI's output directory. You can write [output] in front to be explicit about it, but it's the default.
• An absolute path is taken as-is, and surrounding quotes are stripped, so you can paste a Windows path straight out of Explorer without it breaking.

start_seconds (default 0) and end_seconds (default -1) slice the source. -1 means "to the end of the file", which is the honest default: most references are short clips used whole.

Under the hood it decodes with PyAV - the same av package ComfyUI itself ships, so there's nothing extra to install - and normalises integer PCM to float. It also hashes the file's mtime into its change detection, so replacing the file on disk re-triggers the node instead of handing you a cached stale waveform. That sounds like a small thing until you've spent twenty minutes wondering why your new VO take isn't being used.

Inputs and outputs

Required, all three: audio_path (STRING), start_seconds (FLOAT, ≥0), end_seconds (FLOAT, ≥-1). Output: audio (AUDIO).

The tooltip is worth reading once, since it's the author's own framing: this is H3 reference audio. If you're after a whole-song master for a music-video pipeline, that's what CoachBatePrepareShotContinuity's conditioning_audio input is for.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/CoachBate/ComfyUI-CoachBate.git

Or ComfyUI Manager → search "CoachBate" → restart ComfyUI. Nodes appear under CoachBate / H3. The pack declares no pip dependencies of its own; this node uses PyAV, which ComfyUI already requires.

Where people get burned

• "H3 reference audio not found: …" - the resolved path doesn't exist. Nine times out of ten it's a relative path that landed in output/ when you meant it to be absolute.
• "starts at X s, after the 4.2s file ends" - your start_seconds is past the end of the file. Usually a leftover from the previous shot's longer reference.
• "end (…) must be after start (…)" - end_seconds below start_seconds. Note that -1 is special-cased and is never the problem.
• The waveform is far too short for the shot. You sliced a range; the model then gets a reference it can't use for the whole duration. Check the slice bounds before blaming lipsync.