Denoise Audio (DeepFilterNet3)

How it fits

Typical backends

Caveats

• I/O contractOutput is always 48 kHz regardless of input rate, and multi-channel input is mixed to mono without weighting before inference; pair with a downstream resample if the source rate must be preserved.
• I/O contractAn empty input AudioFrame is returned unchanged without invoking the model, and a pure-silence (zero-energy) frame is emitted as model-rate silence without invoking the model; both still emit exactly one output frame.
• Parameter interactionattenuation_db caps the maximum noise reduction (the full value applies the model's complete output), and post_filter adds residual-noise suppression at a small cost while slightly attenuating quiet speech; both are fixed at startup.
• Parameter interactionA positive chunk_seconds splits the input into fixed-length pieces to bound memory before denoising; chunk boundaries can introduce minor audible discontinuities. It is captured once at startup.
• FallbackThe denoiser falls back to CPU when no GPU is present, which is fine here since the model is real-time on a mid-range CPU.
• 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.