Meta-AF Demos

We release the inputs, targets, and results for five samples in the test set of each adaptive filter task for all models. These samples are all generated using the Meta-AF codebase in this GitHub repo. Please consult our paper for any experimental configuration details as well as descriptions of our baselines.

Acoustic Echo Cancellation

Single-Talk

Magnitude STFT of the First Sample* *STOI is zero since the near-end speech is silent.

Near-end	Far-end	Near-end Speech	Speex	LMS	RMSProp	NLMS	BD-RLS	D-KF	Meta-AEC

Double-Talk

Magnitude STFT of the First Sample

Near-end	Far-end	Near-end Speech	Speex	LMS	RMSProp	NLMS	BD-RLS	D-KF	Meta-AEC

Double-Talk with Path Change

Magnitude STFT of the First Sample

Near-end	Far-end	Near-end Speech	Speex	LMS	RMSProp	NLMS	BD-RLS	D-KF	Meta-AEC

Double-Talk with Path Change & Nonlinearities

Magnitude STFT of the First Sample

Near-end	Far-end	Near-end Speech	Speex	LMS	RMSProp	NLMS	BD-RLS	D-KF	Meta-AEC

Equalization

Unconstrained

Magnitude STFT of the First Sample

Target	Input	LMS	RMSProp	NLMS	D-RLS	Meta-EQ

Constrained

Magnitude STFT of the First Sample

Target	Input	LMS	RMSProp	NLMS	D-RLS	Meta-EQ

Dereverberation

One, Four, Eight Mic.

Note, this task is a failure-mode of Meta-AF, where perceptual quality is poor despite solving the optimization task well. Magnitude STFT of the First Sample

Reverberant	Anechoic	NARA 1 Mic.	Meta-WPE 1 Mic.	NARA 4 Mic.	Meta-WPE 4 Mic.	NARA 8 Mic.	Meta-WPE 8 Mic.

Beamforming

Diffuse Interferer

Magnitude STFT of the First Sample

Clean Speech	Mixture	LMS	RMSProp	NLMS	BD-RLS	Meta-GSC

Directional Interferer

Magnitude STFT of the First Sample

Clean Speech	Mixture	LMS	RMSProp	NLMS	BD-RLS	Meta-GSC

Note that all samples have been scaled to [-1,1] and saved as .mp3 files for playback. If you would like the raw files, please either follow the directions in the GitHub repository or contact me.