Denoise Audio (ClearerVoice Studio)

How it fits

Typical backends

Caveats

• Hard constraintmodel_name must be exactly `MossFormerGAN_SE_16K`, `FRCRN_SE_16K`, or `MossFormer2_SE_48K`; any other value raises an unsupported model error.
• Latencymodel_name and device are resolved once at startup and are NOT mutable; changing either requires a redeploy. Weights download on first load and a warm-up inference runs at startup, so initial startup is slow but the first real call is fast.
• I/O contractOutput sample rate is the model's native rate (16 kHz or 48 kHz); the component does NOT match the input rate — downstream consumers must handle the rate change.
• I/O contractInput is flattened to mono before resampling; multi-channel input is mixed down without weighting.
• I/O contractAn empty input AudioFrame is returned unchanged without invoking the model.
• Fallbackdevice starting with `cuda` silently falls back to CPU when CUDA is unavailable at startup; CPU is much slower, especially for the 48 kHz model.
• I/O contractOutput amplitude is whatever the model emits and is NOT loudness-matched to the input; pair with a loudness normaliser downstream if level matters.