FireRedAudio 参考音频质检 · T8star-Aix

Every voice-cloning workflow has the same failure pattern: the output sounds wrong, so you blame the sampler, the seed, the model - and it was the reference clip all along. ReferenceQuality is the honesty check that runs before you spend an hour regenerating. It measures the things that actually break clones - duration, sample rate, channels, clipping, silence ratio, loudness, and DC offset - and hands you a JSON report. In this pack it's the first node of the reference-prep chain, and in the wider ecosystem it fills the exact gap the KB's audio essay keeps hitting: reference audio is 80% of the result, but almost no tooling tells you whether your clip qualifies.

How it works

The node sends your audio to the isolated worker's analyze_audio, which computes the metrics over the waveform. The model input is required because the analysis runs inside the worker - though it's cheap: no generation, just a few seconds of measurement, and the lite model profile is plenty. Clipping is reported as a ratio (samples at full scale over total samples), silence as the proportion below a noise floor, loudness in LUFS, and DC offset in whatever the source carries.

Outputs are minimal and to the point: 原始音频 (your clip passed through untouched - the node never modifies anything) and 质检报告 JSON with every metric plus the worker details.

What to look for

• Clipping ratio > ~0 - if the waveform is flat-topping, the model inherits the distortion. High clipping on a reference is a genuine "get a better recording" signal.
• Silence ratio - a clip that's half silence means the model spends context on dead air. That's what PrepareReference's trim is for.
• Duration - FireRedAudio clones best from a few clean seconds; the pack's own guidance and the reference-mining defaults converge on 3–15.
• DC offset - a small DC bias shows up as a low-frequency rumble and can make cloned audio sound "off" in a way you can't quite name.
• Sample rate / channels - 24 kHz mono is the pack's target; if your clip is 48 kHz stereo, note it and let PrepareReference handle the conversion.

Where it fits

Right before PrepareReference: Load Audio → this node → 准备参考音频 → the clone's prompt_audio. Run it once on a candidate and you'll know whether to feed it to the clone, run it through cleanup first, or go back to the recording. It's also genuinely useful for auditing a voice reference someone else handed you - before you build an entire batch on a clipped, half-silent clip, let the report tell you.

Installing

Pack-wide install (ComfyUI Manager search comfyui-fireredaudio-T8, or clone + python scripts\setup_runtime.py). The model download is required for the worker even though this node doesn't generate - if the worker can't start, analyze_audio fails with an RPC error. That's the one thing that trips people up: a "just measuring" node that still needs the runtime standing up. Once it's running, this is the cheapest diagnostic in the whole pack, and the one you'll use most often.