Denoise Audio (MossFormer2)

How it fits

Typical backends

Caveats

• I/O contractOutput is ALWAYS mono at 48 kHz regardless of input rate or channel count; multi-channel input is mixed to mono without weighting. Add a downstream `resample_audio_frame` if the original rate matters.
• I/O contractEvery input AudioFrame yields exactly one output AudioFrame. An empty input frame is returned unchanged at its original rate without invoking the model, so it is the one case the output is not 48 kHz.
• LatencyEach input is denoised independently with no cross-call buffer. The ClearVoice runtime decodes clips at or under its ~20 s threshold in a single window and sliding-window decodes longer clips, so there is no external padding and no fixed minimum-length cost. The output sample count tracks the 48 kHz resampled input length, not the original input sample count.
• State lifetimedevice is fixed at startup with no runtime mutation; changing it requires a redeploy, and it falls back to CPU when no GPU is present — CPU mode is well below real-time for this model. The MossFormer2_SE_48K weights are fetched by the ClearVoice runtime on first deploy and cached, so the first call also pays the download and cold-kernel cost.
• I/O contractOutput amplitude is whatever the model emits and is not loudness-matched to the input; add a downstream `normalize_audio_loudness_ebu_r128` if level matters.
• Compatibilitydenoise_audio_clearervoice_studio exposes the same MossFormer2 runtime plus the lighter 16 kHz models; `denoise_audio_frcrn` and `denoise_audio_deepfilternet3` are faster lower-latency denoisers; `upscale_audio_mossformer2` is the super-resolution sibling.