Denoise Audio (FRCRN)

How it fits

Typical backends

Caveats

• I/O contractout_0 is always 16 kHz mono regardless of the input rate or channel layout; multi-channel in_0 is averaged to mono without weighting. Pair with a downstream resample if the original rate matters.
• FallbackAn empty input AudioFrame is returned unchanged without invoking the model, so every input still yields exactly one output frame.
• CompatibilityThe FRCRN weights are downloaded by the ClearVoice runtime on first run and cached; there is no binding for a local checkpoint, so the cold start needs network access.
• LatencyFRCRN is markedly slower than a lightweight real-time denoiser at the same real-time factor; the device is fixed at startup with no hot-swap path, and device starting with `cuda` falls back to CPU when no GPU is present.
• I/O contractOutput amplitude is whatever the model emits and is NOT loudness-matched to the input; add a loudness normaliser downstream if level matters.