FireRedAudio 音频理解 · T8star-Aix

Feed it a clip, type a question, get a written answer. That's the whole node: FireRedAudio 音频理解 turns audio into text you can actually search, summarize and check - without listening to a second of it.

Why do you want that? Three workflows, in order of how often I reach for it:

1. Verify TTS output. Generate a line with the TTS node, then ask this node "what did the speaker say?" and compare against the script. Catches garbled clones and ASR-mangled references before they hit a render.
2. Summarize a reference clip. Instead of scrubbing through a minute of someone mumbling, just ask it to summarize or transcribe.
3. Compare two recordings. The optional second audio input turns this into a side-by-side comparison - "what's the difference between take A and take B?" - which is genuinely useful when you're picking between cloned takes.

It's the cheapest node in the pack to run, too. Unlike the TTS and voice-design nodes, understanding only needs the main model - the lite profile covers it - so you skip the RedAE decoder entirely.

How it works

FireRedAudio is an audio language model, so this is the same machinery as its TTS: the clip is encoded into audio tokens and fed to the Qwen backbone alongside your prompt, and the model generates a text answer. Turn on enable_thinking and you get the model's chain-of-thought reasoning in a separate output - slower, but great when the answer seems wrong and you want to see why.

Inputs and outputs

• model - from the "FireRedAudio 模型/隔离运行时" loader.
• audio - clip one.
• prompt - the question. Default is "请总结音频内容。" (summarize this audio); it's a plain text field, so write it in whatever language you want answers in.
• enable_thinking - off by default.
• max_new_tokens - 1024 default, up to 10240. If you're summarizing something long, bump it.
• audio_2 - optional. Wire a second clip in and the node answers about both.

Outputs are 回答 (the answer, a STRING), 思考过程 (reasoning, empty when thinking is off), and 运行报告 with the inference stats. Wire 回答 into any text display or straight into a workflow as a condition.

Installing it

Same pack as everything else here - it's one of the 45 nodes in comfyui-fireredaudio-T8:

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

Then download the model. --profile lite is enough for understanding, but if you'll ever run the TTS or voice-design nodes, just take full now and save yourself a second download:

python scripts\download_models.py --target "D:\ComfyUI\models\TTS\FireRedAudio" --profile full --variant int8-wo-safe-v1

The node runs in an isolated Python 3.10 / Transformers 5.8 worker precisely so it can't break your other custom nodes' Transformers - that's the setup script's whole job, don't skip it.

Gotchas worth knowing

The answer is only as good as what the model heard. Heavily accented or noisy audio will produce confident-but-wrong text, so treat the output as a strong draft, not gospel - that's also why the "check TTS output with this" loop works so well. Long clips give better answers if you ask short questions about segments rather than one huge "summarize everything" prompt. And if the default Chinese prompt throws you, just type your own; the model handles both languages fine. First run pays a minute-or-so model-loading cost; warm runs are quick.