FireRedAudio 多音频比较理解 · T8star-Aix

FireRedAudio isn't just a TTS model - it's an audio language model, which means it can listen as well as speak. This is the node that puts the listening to work: you feed it one to eight audio clips and a question, and it answers in text. That's "is this the same person in both recordings?", "which take has the more natural reading?", or "list the timestamps where speaker A and speaker B overlap." The single-audio version of this exists too (FireRedAudio 理解); this one is the comparison tool, and it's the one worth reaching for when you're judging takes, verifying reference clips, or trying to figure out what a messy multi-speaker recording actually contains.

How it works

The node auto-grows its audio inputs - connect one, and it stays one; connect more, and it adds them, up to eight (audios, with audio_1…audio_8 sockets). Each clip gets written to WAV and shipped to the pack's isolated worker, which runs the FireRedAudio model's understand task with your prompt and returns the answer. No API call, no key - this is your own GPU doing the whole thing.

The inputs that actually matter:

• model - wire the FireRedAudio 模型/隔离运行时 output in. Worth noting: understanding only needs the main model, so it works fine on the lite profile download.
• prompt - plain-language question. The default ("Compare these recordings and explain the relevant similarities and differences.") is a fine starting point, but you'll get sharper answers if you ask something specific like "which clip is spoken by the same person as clip 1?"
• enable_thinking - off by default. Flip it on when you want the model to show its reasoning before the answer; costs tokens and latency.
• max_new_tokens - default 1024. Bump it if your prompt is asking for a detailed comparison and the answer is getting cut off.

Outputs are 回答 (the answer string), 思考过程 (reasoning, empty when thinking is off), and 运行报告 (a JSON report with timing and worker info). Wire the answer into a text display or the pack's 保存字幕/文本 node and you have a written summary you can actually file.

Where it fits

Think of it as the analysis step before you commit to a take. Run SeedAudition or BatchDubbing, drop the candidates into this node with a "compare these and describe the differences in delivery" prompt, and get a text readout you can paste into a review note - then use the review board for the actual human listen. It's also handy for the reverse direction: transcribe a long recording with the pack's long-ASR nodes, then ask this node to summarize what happened per segment.

Installing

Same story as the whole pack: ComfyUI Manager search comfyui-fireredaudio-T8 (it's been submitted to the registry; if it hasn't appeared yet, use the manual route), or:

cd ComfyUI\custom_nodes
git clone https://github.com/T8mars/comfyui-fireredaudio-T8.git
cd comfyui-fireredaudio-T8
python scripts\setup_runtime.py

That setup_runtime.py step is not optional - it creates the isolated Python 3.10 / torch 2.8 / Transformers 5.8 worker the model runs in, and the pack deliberately never installs anything into ComfyUI's own environment. Then download the model (python scripts\download_models.py --target "ComfyUI\models\TTS\FireRedAudio" --profile lite is enough for ASR/understanding). If the node complains that no model was found, that's the download step you skipped, not a bug.

One honest caveat: with zero impressions and no community threads yet, you're an early adopter. The pack's own docs are thorough and the author (T8star-Aix, same person behind the IndexTTS 2.5 pack) clearly tests with real models - but expect the usual early-stage polish gaps.