Process Whisper Output (mtb)

The description tells you the job in one line: "Process Whisper output into timestamped chunks." Whisper is OpenAI's open speech-to-text model, and a handful of ComfyUI packs (mtb's ecosystem included) run it to transcribe audio inside a graph - useful for lip-sync timing, subtitle generation, or driving any node that needs to know when words were spoken, not just what was said. Raw Whisper output isn't immediately friendly to work with downstream: this node takes that output and reshapes it into clean, timestamped segments you can actually wire into the rest of your graph.

This node consumes a WHISPER_OUTPUT type - meaning something earlier in your graph, a Whisper transcription node, has to produce that type first. This node is purely the processing step after transcription, not the transcription itself.

Inputs and outputs that matter

• whisper_output - the raw output from whatever upstream node ran the actual Whisper transcription.
• min_chunk_length (0–10, default 0) - a minimum duration, in seconds, for a chunk to be kept as-is. With the default of 0, you likely get every chunk Whisper produced, including very short ones (a single word, a fragment). Raising this filters out or merges chunks below that length - useful if raw Whisper segmentation is giving you a lot of tiny, choppy fragments that aren't useful for downstream timing.

Two outputs: a plain STRING (almost certainly the flattened transcript text) and WHISPER_CHUNKS - the structured, timestamped segments, which is the piece you'd feed into anything that needs to know when each bit of speech happens rather than just reading the words.

Installing it

• ComfyUI Manager - search "MTB Nodes", install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/melMass/comfy_mtb, restart. A manual clone needs pip install -r requirements.txt run inside the comfy_mtb folder first.

This node itself is just post-processing - the real weight (and the real dependency footprint) is in whatever node upstream actually runs Whisper. Whisper models range from tiny and fast to large and much more accurate; if you haven't set up transcription yet, that's the bigger install step to sort out first, separate from this node.

Common issues & troubleshooting

Output is fragmented into a lot of tiny chunks. Raise min_chunk_length above 0 to filter out or merge the shortest fragments - a single-word chunk is rarely useful on its own for timing purposes.

No whisper_output to connect. This node has nothing to process without an upstream transcription step already in your graph. Confirm you actually have a node producing WHISPER_OUTPUT before wiring this in - it's not a self-contained transcription node.

Timestamps feel slightly off from the actual audio. That's a property of whatever Whisper model and settings produced whisper_output in the first place, not something this processing step introduces or can correct - Whisper's own segmentation accuracy varies by model size and audio quality.

Some mtb nodes fail to load at startup. Standard for this pack - check the console for [comfy_mtb] STATUS and http://127.0.0.1:8188/mtb for what didn't load. Audio-adjacent nodes are a reasonable candidate for missing optional dependencies if your ComfyUI install skipped them.