☁️OpenAI ASR

The display name says "ASR" - automatic speech recognition, the general term - but underneath it's a direct wrapper around OpenAI's Whisper transcription endpoint. Point it at audio, get text back. It's the input-side counterpart to this pack's OpenAI TTS node, and together they're most of what the README means by a "one-stop LLM + TTS + ComfyUI workflow" for streaming and content-creation use cases - this is how spoken input gets into the LLM chain in the first place.

How it works

You give it audio either as a file path or as ComfyUI's native AUDIO type, and it calls the Whisper API to transcribe it into text.

The inputs and outputs that matter

• audio_path (default empty) - a path to an audio file on disk. Required by the schema, though in practice you'll typically use one input or the other, not both.
• audio (optional, AUDIO) - wire in audio produced by an upstream ComfyUI node directly, rather than pointing at a file path.
• base_url / api_key - optional, standard OpenAI-format credentials; leave blank to fall back on config.ini.
• is_enable - the pack's standard bypass toggle.

One output: text (STRING) - the transcription, ready to feed into any of this pack's LLM nodes, a Save Text node, or straight into a chat history.

How to install it

• ComfyUI Manager: search "comfyui_LLM_party", install, restart.
• Manual: cd ComfyUI/custom_nodes && git clone https://github.com/heshengtao/comfyui_LLM_party, then pip install -r requirements.txt from inside the pack folder using ComfyUI's own Python, then restart.

This is a thin API wrapper - no local Whisper model, no GPU cost of its own, just an API key and a working connection.

Common issues & troubleshooting

Calls fail with an auth error. A real OpenAI key with active billing is required - Whisper is a paid endpoint with no free tier. Check api_key on the node, or in config.ini if you left the node field blank.

Long recordings fail or time out. OpenAI's Whisper endpoint caps upload size - a very long or high-bitrate file may need to be split or compressed first rather than fed in whole. If a short clip transcribes fine but a long one fails, that's almost certainly the cause.

text comes back garbled or in the wrong language. Whisper auto-detects the spoken language reasonably well on clear audio, but background noise, overlapping speakers, or a strong accent it wasn't trained heavily on can degrade accuracy - that's a property of the model itself, not something this node's few configuration fields can tune around.

The pack won't load at all. Separate from this node, comfyui_LLM_party has a real, recurring install issue where the whole node set fails to import after a fresh Manager install, usually a dependency conflict in its large bundled requirements.txt. Check the ComfyUI console for an import error before assuming anything's specifically wrong with audio handling here.