Denoise Audio (MP SENet)

How it fits

Typical backends

Caveats

• I/O contractOutput is always 16 kHz regardless of input rate, and multi-channel input is mixed to mono without weighting; pair with a downstream resample to restore the source rate.
• I/O contractAn empty input AudioFrame is returned unchanged without invoking the network; every input frame produces exactly one output frame.
• Hard constraintThe denoiser checkpoint is bound as the model file (a bundle holding `config.json` and `model.safetensors`); it is loaded once from the local binding at startup, so no network access is needed at run time.
• Parameter interactionA positive chunk_seconds sets the length of the model's internal inference segments to bound memory; 0 keeps the model's built-in 2 second segment. Larger segments can improve quality at higher memory cost.
• AccuracyEnabling half_precision halves VRAM but can amplify artefacts in quiet sections; verify on representative audio before relying on it.
• Fallbackdevice starting with `cuda` silently falls back to CPU when CUDA is unavailable, which is functional but impractical for live audio.
• State lifetimeThe bound weights and all configuration keys are captured once at startup; runtime changes have no effect and require a redeploy.
• I/O contractOutput amplitude is not loudness-matched to the input; add a loudness normaliser downstream if level matters.