Dereverberate Audio (NARA WPE)

@fgnt1 VersionEnhance

Deterministic WPE dereverberation (no neural network, no GPU): each AudioFrame is mixed to mono, resampled to 16 kHz, dereverberated in the STFT domain, and emitted as a 16 kHz AudioFrame. Removes late room reflections only — it does not remove noise.

How it fits

Typical backends

Pre-ASR cleanup of a reverberant recording: removing late reflections sharpens speech so transcription is more accurate.

Full cascade: dereverberation removes room reflections and a neural denoiser then strips broadband noise before transcription, since WPE alone does not denoise.

Caveats

I/O contractOutput is always 16 kHz regardless of input rate; pair with a downstream resample to restore the original rate.
LimitationMulti-channel input is mixed to mono — this is a single-channel WPE implementation and cannot exploit multi-microphone spatial cues.
I/O contractAn empty or effectively-silent input frame emits an empty / silent 16 kHz mono frame without invoking the STFT or filter.
Parameter interactioniterations and taps both raise dereverberation strength at roughly linear CPU cost, while delay protects direct and early-reflection energy — set too small it attenuates the dry signal.
Parameter interactionstft_size and stft_shift set the analysis window and hop; their defaults assume 16 kHz, and changing one without the other shifts the effective time / frequency resolution the taps and delay are tuned against.
LimitationWPE attenuates late reflections only and does not remove broadband background noise; cascade with a denoiser for full cleanup.
State lifetimeAll five settings are captured once at import (the filter geometry is baked into the deployment), so changing iterations, taps, delay, stft_size, or stft_shift requires a redeploy.

Versions

52c2a7b1latestdefault
Automated release
6/29/2026