✨📝 Groq ALM API - Transcribe

Worth clearing up first, because people genuinely mix these up (even a dev announcing a similar ComfyUI LLM node once typo'd it in his own post): this is Groq, the fast-inference chip company, not Grok, Elon Musk's chatbot. Groq's whole pitch is serving open models - including Whisper - at speeds regular GPU inference can't touch, with a genuinely usable free tier. This node calls Groq's hosted Whisper models to transcribe an audio file into text, so you get Whisper-quality transcription in your graph without loading a Whisper checkpoint or spending your own GPU time on it.

How it works

Point it at an audio file, pick which Whisper variant to hit, and it sends the request off to Groq's API and hands the text back. You get a success flag and status_code alongside the transcription itself, so you can branch a workflow on whether the call actually worked instead of just hoping.

The inputs and outputs that matter

• model - whisper-large-v3-turbo, distil-whisper-large-v3-en, or whisper-large-v3. Turbo is the fast default; the distilled English-only variant is faster still if you know your audio is English; the full whisper-large-v3 is the most accurate of the three.
• file_path - a path to the audio file, as a plain string. Note this is a path, not a wired AUDIO tensor - if you're generating audio earlier in the graph, you need to save it to disk first and pass that path in here.
• preset - a transcription prompt preset (plain transcription, song lyrics, meeting notes, or your own custom presets added to UserPrompts.json).
• language - an ISO 639-1 code (en, ja, fr, and so on); the tooltip lists the full supported set.
• response_format - text, text_with_linebreaks, text_with_timestamps, json, or verbose_json, depending on how much structure you need downstream.
• temperature and max_retries - leave at defaults unless you're chasing more deterministic output or hitting transient failures.

Outputs: transcription_result (the text), success (BOOLEAN), status_code (STRING).

How to install it

Via ComfyUI Manager: search ComfyUI-mnemic-nodes, install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/MNeMoNiCuZ/ComfyUI-mnemic-nodes

Then the extra step this node actually needs - the README calls it out under "Configuration (only needed for Groq nodes)": copy .env.example in the pack's root folder to .env, and paste your Groq API key into it. This is the only setup step that's specific to this node (and its LLM/VLM siblings) versus the rest of the pack, which needs nothing beyond the clone.

Common issues & troubleshooting

Nothing transcribes, but ComfyUI doesn't crash. Check success and status_code before assuming the node is broken - a missing or invalid API key typically comes back as a clean failure in those two outputs rather than an exception, which is actually a nicer failure mode than most API nodes give you. Fix the .env key and re-run.

"File not found"-type failures. file_path has to be a real path the ComfyUI process can read - a relative path that made sense on your machine but not in the process's working directory is a common trip-up. Use an absolute path if you're not sure.

Rate limiting on a big batch. Groq's free tier is fast but not unlimited; hammering it with a large Load Text-Image Pairs (List)-style batch of audio files back to back is the kind of thing that gets you throttled. max_retries gives you some cushion, but for a genuinely large batch, throttle your own request rate rather than relying on retries to smooth it over.