🎙️ AIIA ASR (Word Timestamps)

Every serious audio pipeline needs a transcription step, and this is the AIIA pack's: speech-to-text built on FunASR, Alibaba's ASR toolkit, with word-level timestamps as the headline feature. The timestamps are the point, not the transcript - this node feeds the pack's subtitle and diarization machinery, so "what was said" matters less than "when, precisely, was it said." And unlike a lot of the pack's models, this one isn't a download-and-forget: the model files have to be placed by you, and the node is refreshingly blunt about it.

How it works

It loads a FunASR AutoModel from ComfyUI/models/funasr/ and runs recognition with word timestamps. The stock model in the pack's own testing is paraformer-zh - a Chinese-optimized Paraformer - and it also supports the SenseVoice family (which needs trust_remote_code; the node handles that itself). Transcription is batched dynamically by seconds, which is the batch_size_s knob: bigger batches are faster but eat VRAM.

The inputs

• audio - the track to transcribe.
• model - a dropdown of whatever models you've dropped into ComfyUI/models/funasr/. Here's the trap: if that folder is empty, the dropdown literally reads NO_MODELS_FOUND and there's nothing else to select. This is a manual-download node.
• device - cuda default, cpu for when you're out of VRAM.
• batch_size_s - dynamic batch size in seconds, default 300. Faster = thirstier.
• hotword - a list of words, one per line, that get boosted. Great for names and jargon the model keeps mangling; the tooltip is the doc: it improves accuracy on these terms.

Outputs

• asr_result (ASR_RESULT) - the structured result with word-level timestamps; this is what feeds subtitle alignment and diarization downstream.
• text (STRING) - the plain transcript, for when you just want the words in a text box.

Install

The pack install is standard, but the model is the extra step most people miss:

cd ComfyUI/custom_nodes
git clone https://github.com/havvk/ComfyUI_AIIA.git

Then drop a FunASR model folder into ComfyUI/models/funasr/ - for example a paraformer-zh directory with the model files. First run will also want the funasr Python package installed. If the node reports NO_MODELS_FOUND, that's the whole story: no model is installed, and no amount of restarting will fix it.

Gotchas

• NO_MODELS_FOUND is a state, not an error. The node is telling you to place a model. Until one appears in the folder, this node is a dead end.
• It's Chinese-first. Paraformer-zh will transcribe English acceptably but is tuned for Mandarin. For heavy English use, look at the SenseVoice variants or a Whisper-based node elsewhere.
• Hotword wins are real. The single highest-leverage setting if your content is full of proper nouns.