WEAVE — Cross-Device, Array-Agnostic Wearable Speech Enhancement Benchmark

WEAVE is a benchmark for target-speaker speech enhancement on wearable microphone arrays that measures the hardest generalization question in the field: does a model trained on one set of device geometries still work zero-shot on unseen array geometries it never saw in training? Unlike prior wearable benchmarks that fix a single microphone layout (e.g. a 4-mic glasses array), WEAVE spans glasses, earbuds, and neckbands with 2–8 microphones.

Generator code, docs, and evaluation: https://github.com/pujariaditya/WEAVE

⚠️ Licensing — please read first

WEAVE mixtures are rendered from three source corpora that are non-commercial / custom-licensed: EARS (clean speech), WHAM! (ambient noise), and TAU-SRIR DB (real room impulse responses, used only to derive reverberation statistics). The audio here is derivative of those sources and therefore inherits their non-commercial terms — released under CC BY-NC 4.0 for non-commercial research use only. By using it you agree to comply with the upstream licenses:

Corpus	Role in WEAVE	License	Source
EARS	clean target + interferer speech	non-commercial research	https://sp-uhh.github.io/ears_dataset/
WHAM! (high_res)	ambient diffuse noise	CC BY-NC 4.0	http://wham.whisper.ai/
TAU-SRIR DB	reverberation statistics (RT60, DRR)	non-commercial research	https://zenodo.org/record/6408611

For a license-clean path, the generator (Apache-2.0) at the GitHub repo regenerates this benchmark byte-identically from corpora you obtain yourself.

The benchmark

• 13,000 utterances, 48 kHz. Each is a simulated scene: one frontal (0°) target speaker + up to 5 interferers + diffuse noise + sensor noise, recorded by a device's mic array.
• Zero-shot cross-device split — the only variable between seen and unseen is device geometry:

Config / split	Rows	Device classes	Mic counts
seen / train	10,000	glasses, earbuds	2, 4, 6
seen / dev	600	glasses, earbuds	2, 4, 6
seen / test	1,200	glasses, earbuds	2, 4, 6
unseen / test	1,200	glasses, earbuds, neckband	3, 4, 6, 8

Train on head-worn devices (glasses/earbuds); the unseen split holds out a torso-worn neckband (a different acoustic body) and an unseen mic count (3).

How it was built

Array-agnosticism lives in the renderer, not the corpora, so the same acoustic scene can be materialized for any mic layout. Per utterance: mix = reverberant_target + Σ interferers + diffuse noise + sensor noise. The direct path is a parametric DRTF (rigid-sphere head for glasses/earbuds, two-sphere "snowman" torso+head for neckbands, + device coloration + propagation delay). Late reverb is an isotropic diffuse field synthesized from real TAU-SRIR statistics (per-octave RT60, DRR), with inter-mic coherence sinc(2·f·d/c). TAU audio never enters a mixture — only its statistics.

Audio format

mix, target, and enroll are stored as the HuggingFace Audio feature (WAV, 48 kHz), so you can listen to every clip in the dataset viewer. mix is a multichannel WAV (one channel per microphone); target (the clean dry reference at mic 0) and enroll are mono. Each signal is peak-normalized to 0.95 so nothing clips — this does not affect the benchmark because SI-SDR is scale-invariant.

Loading

from datasets import load_dataset, Audio
import soundfile as sf, io, numpy as np

ds = load_dataset("RootAccess4Life/WEAVE", "seen", split="test")   # or config "unseen"
row = ds[0]
target = np.asarray(row["target"]["array"])       # [T] clean dry target at mic 0 (label)
enroll = np.asarray(row["enroll"]["array"])        # [Te] enrollment of the target speaker

# full multichannel mix (all microphones): read the raw WAV bytes
raw = ds.cast_column("mix", Audio(decode=False))
mix, sr = sf.read(io.BytesIO(raw[0]["mix"]["bytes"]))   # [T, n_mics]
mic_positions = np.asarray(row["mic_positions"], np.float32)  # [n_mics, 3], device-frame, centroid-relative

Columns

Column	Type	Notes
id	string	utterance id
device_class	string	glasses / earbuds / neckband
array_id	string	topology id (e.g. 0, eb2, nb8)
condition	string	<device_class>:<array_id>
n_mics, n_samples	int	array shape
mic_positions	list<list<float>>	[n_mics, 3], device-frame, centroid-relative (metres)
mix	Audio	multichannel WAV — the array recording (input)
target	Audio	mono WAV — dry direct-path target at mic 0 (the enhancement label / SI-SDR reference)
enroll	Audio	mono WAV — a different utterance of the target speaker
sir_db, snr_db, sensor_snr_db, rt60, drr_db	float	per-scene acoustic metadata

Task & evaluation

Recover target from mix (single-channel estimate of the frontal target). The headline metric is the cross-device generalization gap — the drop from the seen-test score to the zero-shot unseen-test score (small = geometry-robust) — alongside SI-SDR / STOI / PESQ improvement over the noisy mic-0 input. An evaluation module (weave-eval) is provided in the GitHub repo.

Limitations

Pure simulation (no real wearable-array recordings); parametric acoustics, not measured HRTFs; the neckband "snowman" omits inter-sphere multiple scattering (documented v1 approximation); reverb is a diffuse field synthesized from real room statistics, not full measured-RIR convolution; the dry reference is a modeled target. Real-measured-device validation is future work.

Citation

See CITATION.cff in the GitHub repository. Please also cite the source corpora (EARS, WHAM!, TAU-SRIR DB).