Lyrics JSON Parser

If you're making music videos with LTX-2.3, you've probably discovered the awkward middle step: you ask an LLM to write lyrics and describe the song, and it hands you a JSON blob. Now what? You could paste the whole thing into a CLIP text encode, but that's ugly and mixes data types. Lyrics JSON Parser is the glue: it takes a JSON string containing music metadata and splits it into six typed outputs your graph can actually wire around.

It's a narrow node with one job, and it does that job cleanly. The author built it for his own music-video production pipeline, where an LLM authors the track metadata and this node turns it into usable values.

How it works

The JSON input accepts either raw JSON or a markdown-fenced code block - so you can paste an LLM's entire ```json ... ``` response straight in without stripping the fences first. It decodes the payload and validates a fixed schema. Required fields: lyrics, bpm, key, caption, and duration_sec. One optional field, keyscale, defaults to the value of key when missing. Validation is loud: missing fields, non-object payloads, and wrong types all raise clear errors instead of silently producing garbage.

The input

One socket, JSON (multiline STRING). The default shows the exact expected shape:

{
  "lyrics": "",
  "bpm": 120,
  "key": "C major",
  "caption": "",
  "duration_sec": 180
}

The outputs

Six typed sockets, ready to plug in anywhere:

• lyrics (STRING) - the lyrics text. Into a CLIP text encode, or into a save/caption node for a sidecar file.
• bpm (INT) - beats per minute.
• key (STRING) - musical key, e.g. "C major".
• keyscale (STRING/COMBO) - key with scale; falls back to key if not provided.
• caption (STRING) - short description of the music. Handy as a filename fragment or a metadata tag.
• duration_sec (FLOAT) - track length in seconds. Useful for timing math against your video pipeline.

The value of the typed outputs shows up the moment you start building: bpm and duration_sec can feed timing calculations or audio nodes directly, lyrics goes to the prompt side, and caption rides along as your file label - all without a string-parsing detour.

Common issues

• "JSON input is empty" - nothing connected, or an empty string. This is an input node; it needs data.
• "missing required field(s)" - the LLM didn't emit one of the five required keys. The error names them; fix the prompt or the JSON.
• "must be an integer / must be numeric" - an LLM occasionally writes "bpm": "120" as a string. Fix the type in the JSON.
• Must decode to an object - a bare list or string won't parse. It has to be a {...} object.

Installation

Install via ComfyUI Manager (search "CoachBate") or:

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

Restart ComfyUI; it appears under CoachBate in Add Node. No Python dependencies, no model files. Alpha-stage pack, but this node is about as simple as custom nodes get - just JSON validation and six wires.