Microsoft AEC Challenge 16kHz

Microsoft AEC Challenge dataset converted from 16kHz WAV to FLAC (lossless compression) and packed into tar shards.

Source: the datasets/ directory of the microsoft/AEC-Challenge Git LFS repo. Covers all challenge years (2021, ICASSP 2022, ICASSP 2023).

Structure

Real recordings

Paired loopback (far-end reference) and microphone recordings from real devices.

• real/ — 37,578 files, single playback real recordings
• real_doubled/ — 10,531 files, double playback real recordings

Filenames preserve the GUID-based naming convention: {GUID}_{scenario}_{signal}.flac

Scenarios: farend_singletalk, farend_singletalk_with_movement, nearend_singletalk, doubletalk, doubletalk_with_movement, sweep

Signals: lpb (loopback/far-end reference), mic (microphone recording)

Synthetic data (10,000 clips)

• synthetic_echo/ — Echo signal component
• synthetic_farend/ — Far-end reference signal
• synthetic_nearend_mic/ — Mixed microphone signal (echo + near-end + noise)
• synthetic_nearend_speech/ — Clean near-end speech
• meta.csv — Synthetic data metadata

Test sets

• test_set/ — Original test set (clean + noisy)
• test_set_icassp2022/ — ICASSP 2022 test set
• blind_test_set/ — Original blind test set
• blind_test_set_icassp2022/ — ICASSP 2022 blind test set
• blind_test_set_icassp2023/ — ICASSP 2023 blind test set
• blind_test_set_interspeech2021/ — Interspeech 2021 blind test set

Usage

from huggingface_hub import snapshot_download
import tarfile
from pathlib import Path

# Download
local = snapshot_download("richiejp/aec-challenge-16k", local_dir="/data/aec", repo_type="dataset")

# Extract all shards
for tar_path in sorted(Path(local).rglob("*.tar")):
    with tarfile.open(tar_path) as tf:
        tf.extractall(tar_path.parent)

Source

Original data from Microsoft's AEC Challenge:

• https://github.com/microsoft/AEC-Challenge
• License: CC-BY-4.0 (see original repo for details)