Kuvaus
DESCRIPTION
Synthetic-ReMASC is a collection of synthetically generated multi-channel replay speech datasets for physical-access (replay) attack detection in voice-controlled systems. The data is produced by an acoustic simulation framework that models the full replay chain. Specifically, a talker's speech captured by a verification-system microphone array (genuine), recorded by a nearby spoofing microphone, and re-played through a loudspeaker into a second room and re-captured by the array (replay). By doing so, each simulation iteration yields one balanced genuine/replay pair.
The release contains four device-specific datasets, one per microphone array geometry taken from the real ReMASC corpus:
D1 (Google AIY) — 2 channels — replay_spoofed_D1_2mic
D2 (ReSpeaker 4-Linear / Amlogic 113X1) — 4 channels — replay_spoofed_D2_4mic
D3 (ReSpeaker V2) — 6 channels — replay_spoofed_D3_6mic
D4 — 7 channels — replay_spoofed_D4_7mic
Each dataset holds 4,280 recordings (2,140 genuine + 2,140 replay), split talker-disjointly into 2,960 train / 400 validation / 920 test. Recordings are built from the 107 anechoic EARS talkers, with room impulse responses simulated in Pyroomacoustics, measured loudspeaker/talker directivities, and spatially coherent WHAMR! background noise. Every recording is accompanied by rich per-file metadata (room geometry, RT60, surface materials, source/microphone/loudspeaker positions, device names, and SNR).N.B.: Only datasets and details about the simulation framework are available. No generation scripts are available.
REPORT & REFERENCE
This dataset accompanies the paper:
M. Neri and T. Virtanen, "Acoustic Simulation Framework for Multi-Channel Replay Speech Detection," IEEE International Workshop on Multimedia Signal Processing (MMSP), 2026.
Preprint: arXiv:2509.14789
Code / framework: github.com/michaelneri/synthetic-ReMASC
Authors: Michael Neri, Tuomas Virtanen — Faculty of Information Technology and Communication Sciences, Tampere University, Finland
@inproceedings{neri2026acoustic,
title = {Acoustic Simulation Framework for Multi-Channel Replay Speech Detection},
author = {Neri, Michael and Virtanen, Tuomas},
booktitle = {IEEE International Workshop on Multimedia Signal Processing (MMSP)},
address = {Istanbul, T\"urkiye},
year = {2026}
}
If you use this data, please also cite the underlying resources it is built upon: ReMASC (Gong et al., Interspeech 2019, array geometries), EARS (Richter et al., Interspeech 2024, source speech), WHAMR! (Wichern et al., Interspeech 2019, background noise), the Gallien et al. measured loudspeaker directivities (AES 2024), and Pyroomacoustics (Scheibler et al., ICASSP 2018).
AIM
Real multi-channel replay corpora are scarce and expensive to collect: hardware cost, array synchronization/calibration, storage, and the difficulty of reproducing consistent spatial setups across environments have left ReMASC as the only real multi-channel replay dataset. This release addresses that gap by providing a fully synthetic and controllable alternative. Its aims are to:
Enable controlled experiments on the acoustic factors that affect replay detection (room, SNR, array geometry, directivity, source–array geometry), which are impractical to isolate in real recordings.
Support environment-independent evaluation of detectors trained exclusively on synthetic data and tested on real ReMASC recordings, without any real training data.
Facilitate research on spatial feature exploitation — adaptive beamforming, inter-channel phase difference (IPD) features, and per-channel log-power — for generalization from synthetic to real conditions.
SPECIFICATIONS
Audio format: WAV, 16-bit PCM, 16 kHz sampling rate
Duration: ~2.2 s per recording (2 s speech segments, following the ReMASC protocol)
Channels: 2 / 4 / 6 / 7 (per device dataset D1–D4)
Classes: balanced genuine vs. replay
Talkers: 107 anechoic EARS speakers; talker-disjoint 7:1:2 split (no talker leakage across train/val/test)
Per device: 20 excerpts per talker → 4,280 recordings (2,960 / 400 / 920)
Simulator: Pyroomacoustics — shoebox rooms with per-surface materials, simulated RIRs for genuine/spoof/replay paths
Generation ranges (see each config.json): room dimensions U[3.0, 6.0] m; SNR U[5, 40] dB; direction-of-arrival U[−10, 10]°; talker–ASV distance > 1 m; spoofing mic < 1 m from talker; source–surface distance > 1 m
Directivities: measured talker directivity; loudspeaker modeled with either a built-in cardioid or measured SOFA directivities (to avoid overfitting to one loudspeaker); ASV array and spoofing microphones modeled as omnidirectional/flat (MEMS-like)
Background noise: WHAMR!, rendered as spatially coherent multi-channel diffuse noise, globally scaled to the target SNR
Per-dataset contents:
train/, val/, test/ — the WAV recordings
dataset.csv — per-recording metadata: Filename, TalkerID, EARSfilename, subset, EnvARoomDimensions, EnvART60, EnvASurfaceNames, posTalker, posASV, ASVMicName, posSpoof, SpoofMicName, posSpeaker, nameSpeaker, SNRNoise
config.json — full generation parameters and array geometry
Filename convention: pXXX__genuine.wav and pXXX__replay_loudspeaker_.wav (talker ID, excerpt index, condition).
DOWNLOAD SPECIFICATIONS
Four independent ZIP archives — download only the array configuration(s) you need:
replay_spoofed_D1_2mic.zip — D1 (2 ch) — ~273 MB
replay_spoofed_D2_4mic.zip — D2 (4 ch) — ~551 MB
replay_spoofed_D3_6mic.zip — D3 (6 ch) — ~781 MB
replay_spoofed_D4_7mic.zip — D4 (7 ch) — ~883 MB
Total: ~2.5 GB compressed, 17,120 recordings (4 × 4,280).
Each archive unpacks to a self-contained folder with train/, val/, test/, dataset.csv, and config.json.
Loading example is available in the GitHub repository.
Synthetic-ReMASC is a collection of synthetically generated multi-channel replay speech datasets for physical-access (replay) attack detection in voice-controlled systems. The data is produced by an acoustic simulation framework that models the full replay chain. Specifically, a talker's speech captured by a verification-system microphone array (genuine), recorded by a nearby spoofing microphone, and re-played through a loudspeaker into a second room and re-captured by the array (replay). By doing so, each simulation iteration yields one balanced genuine/replay pair.
The release contains four device-specific datasets, one per microphone array geometry taken from the real ReMASC corpus:
D1 (Google AIY) — 2 channels — replay_spoofed_D1_2mic
D2 (ReSpeaker 4-Linear / Amlogic 113X1) — 4 channels — replay_spoofed_D2_4mic
D3 (ReSpeaker V2) — 6 channels — replay_spoofed_D3_6mic
D4 — 7 channels — replay_spoofed_D4_7mic
Each dataset holds 4,280 recordings (2,140 genuine + 2,140 replay), split talker-disjointly into 2,960 train / 400 validation / 920 test. Recordings are built from the 107 anechoic EARS talkers, with room impulse responses simulated in Pyroomacoustics, measured loudspeaker/talker directivities, and spatially coherent WHAMR! background noise. Every recording is accompanied by rich per-file metadata (room geometry, RT60, surface materials, source/microphone/loudspeaker positions, device names, and SNR).N.B.: Only datasets and details about the simulation framework are available. No generation scripts are available.
REPORT & REFERENCE
This dataset accompanies the paper:
M. Neri and T. Virtanen, "Acoustic Simulation Framework for Multi-Channel Replay Speech Detection," IEEE International Workshop on Multimedia Signal Processing (MMSP), 2026.
Preprint: arXiv:2509.14789
Code / framework: github.com/michaelneri/synthetic-ReMASC
Authors: Michael Neri, Tuomas Virtanen — Faculty of Information Technology and Communication Sciences, Tampere University, Finland
@inproceedings{neri2026acoustic,
title = {Acoustic Simulation Framework for Multi-Channel Replay Speech Detection},
author = {Neri, Michael and Virtanen, Tuomas},
booktitle = {IEEE International Workshop on Multimedia Signal Processing (MMSP)},
address = {Istanbul, T\"urkiye},
year = {2026}
}
If you use this data, please also cite the underlying resources it is built upon: ReMASC (Gong et al., Interspeech 2019, array geometries), EARS (Richter et al., Interspeech 2024, source speech), WHAMR! (Wichern et al., Interspeech 2019, background noise), the Gallien et al. measured loudspeaker directivities (AES 2024), and Pyroomacoustics (Scheibler et al., ICASSP 2018).
AIM
Real multi-channel replay corpora are scarce and expensive to collect: hardware cost, array synchronization/calibration, storage, and the difficulty of reproducing consistent spatial setups across environments have left ReMASC as the only real multi-channel replay dataset. This release addresses that gap by providing a fully synthetic and controllable alternative. Its aims are to:
Enable controlled experiments on the acoustic factors that affect replay detection (room, SNR, array geometry, directivity, source–array geometry), which are impractical to isolate in real recordings.
Support environment-independent evaluation of detectors trained exclusively on synthetic data and tested on real ReMASC recordings, without any real training data.
Facilitate research on spatial feature exploitation — adaptive beamforming, inter-channel phase difference (IPD) features, and per-channel log-power — for generalization from synthetic to real conditions.
SPECIFICATIONS
Audio format: WAV, 16-bit PCM, 16 kHz sampling rate
Duration: ~2.2 s per recording (2 s speech segments, following the ReMASC protocol)
Channels: 2 / 4 / 6 / 7 (per device dataset D1–D4)
Classes: balanced genuine vs. replay
Talkers: 107 anechoic EARS speakers; talker-disjoint 7:1:2 split (no talker leakage across train/val/test)
Per device: 20 excerpts per talker → 4,280 recordings (2,960 / 400 / 920)
Simulator: Pyroomacoustics — shoebox rooms with per-surface materials, simulated RIRs for genuine/spoof/replay paths
Generation ranges (see each config.json): room dimensions U[3.0, 6.0] m; SNR U[5, 40] dB; direction-of-arrival U[−10, 10]°; talker–ASV distance > 1 m; spoofing mic < 1 m from talker; source–surface distance > 1 m
Directivities: measured talker directivity; loudspeaker modeled with either a built-in cardioid or measured SOFA directivities (to avoid overfitting to one loudspeaker); ASV array and spoofing microphones modeled as omnidirectional/flat (MEMS-like)
Background noise: WHAMR!, rendered as spatially coherent multi-channel diffuse noise, globally scaled to the target SNR
Per-dataset contents:
train/, val/, test/ — the WAV recordings
dataset.csv — per-recording metadata: Filename, TalkerID, EARSfilename, subset, EnvARoomDimensions, EnvART60, EnvASurfaceNames, posTalker, posASV, ASVMicName, posSpoof, SpoofMicName, posSpeaker, nameSpeaker, SNRNoise
config.json — full generation parameters and array geometry
Filename convention: pXXX__genuine.wav and pXXX__replay_loudspeaker_.wav (talker ID, excerpt index, condition).
DOWNLOAD SPECIFICATIONS
Four independent ZIP archives — download only the array configuration(s) you need:
replay_spoofed_D1_2mic.zip — D1 (2 ch) — ~273 MB
replay_spoofed_D2_4mic.zip — D2 (4 ch) — ~551 MB
replay_spoofed_D3_6mic.zip — D3 (6 ch) — ~781 MB
replay_spoofed_D4_7mic.zip — D4 (7 ch) — ~883 MB
Total: ~2.5 GB compressed, 17,120 recordings (4 × 4,280).
Each archive unpacks to a self-contained folder with train/, val/, test/, dataset.csv, and config.json.
Loading example is available in the GitHub repository.
| Koska saatavilla | 21 elok. 2026 |
|---|---|
| Julkaisija | Zenodo |
Field of science, Statistics Finland
- 113 Tietojenkäsittely ja informaatiotieteet
Siteeraa tätä
- DataSetCite