cMPDR audio demos

Audio samples accompanying the paper: "Cyclic MPDR Beamforming for Suppression of Almost-Cyclostationary Acoustic Interference" [GitHub] [arXiv]
by Giovanni Bologni, Martin Bo Møller, Richard Heusdens and Richard C. Hendriks

The Cyclic Minimum Power Distortionless Response (cMPDR) beamformer is an algorithm for canceling loud tonal interferers such as engines, fans, and musical instruments. 🎶
By exploiting correlations across microphones and frequency components, cMPDR achieves improved noise reduction compared to the standard MPDR beamformer, particularly in low signal-to-noise ratio (SNR) scenarios.

Spectrograms: Noisy, Target, MPDR, cMPDR

Mel spectrograms of clean speech mixed with buzz noise from the MUSAN recording noise-free-sound-0002 (-5 dB SNR), showing the noisy input, the clean target signal, and the outputs of the MPDR and proposed cMPDR. The highlights show that the cMPDR can suppress strong tonal noise thanks to cyclostationarity-aware processing.

Experimental setup

🔬 We evaluate the algorithm on recordings from the DREGON (drone) dataset, synthetic harmonic noise, and non-stationary noise from MUSAN/Freesound. 🚁

⚙️ Parameters: All samples use $M=2$ microphones, maximum number of frequency shifts $C_{\max} = 8$, and coherence threshold $\gamma_{\min} = 0.6$. All settings match Tables II and III in the paper.

🗣️ Target Speech: All mixtures use Harvard sentences (IEEE Std 297-1969) as the target signal.

🎧 Headphones: We recommend listening with headphones for the best fidelity.

Dataset 1: DREGON

📝 Description: DREGON (Drone Dataset for Grounded Object Navigation) features motor-only recordings from a stationary quadrotor UAV. These recordings include individual propeller and combined engine noise at various rotational speeds, exhibiting pseudo-harmonic spectral peaks.

📊 Results: In these experiments, cMPDR achieves significant noise reduction gains over the MPDR benchmark, particularly in low signal-to-noise ratio (SNR) conditions.

SNR -20 to -10 dB
Sample: Motor4_70
Noisy

Mixture

SI-SDR: -19.88 dB | STOI: 0.05
Baseline

MPDR

SI-SDR: -22.83 dB | STOI: 0.13
Spectrogram Motor4_70
Spectrograms: Noisy, Target, MPDR, cMPDR
Reference

Clean target

Reference Audio
Proposed

cMPDR

SI-SDR: -14.08 dB | STOI: 0.25
Sample: Motor2_90
Noisy

Mixture

SI-SDR: -24.86 dB | STOI: 0.12
Baseline

MPDR

SI-SDR: -28.24 dB | STOI: 0.32
Spectrogram Motor2_90
Spectrograms: Noisy, Target, MPDR, cMPDR
Reference

Clean target

Reference Audio
Proposed

cMPDR

SI-SDR: -12.64 dB | STOI: 0.40

SNR -10 to 0 dB
Sample: Motor2_70
Noisy

Mixture

SI-SDR: -10.60 dB | STOI: 0.26
Baseline

MPDR

SI-SDR: -11.93 dB | STOI: 0.39
Spectrogram Motor2_70
Spectrograms: Noisy, Target, MPDR, cMPDR
Reference

Clean target

Reference Audio
Proposed

cMPDR

SI-SDR: -2.36 dB | STOI: 0.54
Sample: Motor2_50
Noisy

Mixture

SI-SDR: -9.99 dB | STOI: 0.18
Baseline

MPDR

SI-SDR: -10.19 dB | STOI: 0.34
Spectrogram Motor2_50
Spectrograms: Noisy, Target, MPDR, cMPDR
Reference

Clean target

Reference Audio
Proposed

cMPDR

SI-SDR: -5.86 dB | STOI: 0.44

SNR 0 to 10 dB
Sample: Motor2_80
Noisy

Mixture

SI-SDR: -0.85 dB | STOI: 0.53
Baseline

MPDR

SI-SDR: -2.27 dB | STOI: 0.65
Spectrogram Motor2_80
Spectrograms: Noisy, Target, MPDR, cMPDR
Reference

Clean target

Reference Audio
Proposed

cMPDR

SI-SDR: 4.71 dB | STOI: 0.76
Sample: Motor1_50
Noisy

Mixture

SI-SDR: -0.62 dB | STOI: 0.35
Baseline

MPDR

SI-SDR: -0.66 dB | STOI: 0.41
Spectrogram Motor1_50
Spectrograms: Noisy, Target, MPDR, cMPDR
Reference

Clean target

Reference Audio
Proposed

cMPDR

SI-SDR: 2.63 dB | STOI: 0.53

DREGON excerpts are available at dregon.inria.fr and are freely available for personal, educational, and academic use only.
Please cite: DREGON: Dataset and Methods for UAV-Embedded Sound Source Localization, Martin Strauss, Pol Mordel, Victor Miguet and Antoine Deleforge, IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) 2018.

Dataset 2: Synthetic Harmonic Noise

📝 Description: This dataset uses random harmonic signals with adjustable spectral correlation to simulate the partial correlation observed in real-world noise. It allows for controlled evaluation of the beamformer's sensitivity to parameters like correlation level, frequency estimation error, and reverberation.

📊 Results: The results demonstrate that cMPDR achieves improved noise reduction as spectral correlation increases, confirming its theoretical advantage.

SNR -20 to -10 dB
Sample: 10
Noisy

Mixture

SI-SDR: -21.06 dB | STOI: 0.32
Baseline

MPDR

SI-SDR: -17.90 dB | STOI: 0.29
Spectrogram 10
Spectrograms: Noisy, Target, MPDR, cMPDR
Reference

Clean target

Reference Audio
Proposed

cMPDR

SI-SDR: -8.58 dB | STOI: 0.54
Sample: 16
Noisy

Mixture

SI-SDR: -19.89 dB | STOI: 0.27
Baseline

MPDR

SI-SDR: -15.91 dB | STOI: 0.37
Spectrogram 16
Spectrograms: Noisy, Target, MPDR, cMPDR
Reference

Clean target

Reference Audio
Proposed

cMPDR

SI-SDR: -3.96 dB | STOI: 0.57

SNR -10 to 0 dB
Sample: 5
Noisy

Mixture

SI-SDR: -9.86 dB | STOI: 0.41
Baseline

MPDR

SI-SDR: -10.87 dB | STOI: 0.43
Spectrogram 5
Spectrograms: Noisy, Target, MPDR, cMPDR
Reference

Clean target

Reference Audio
Proposed

cMPDR

SI-SDR: -1.05 dB | STOI: 0.62
Sample: 14
Noisy

Mixture

SI-SDR: -10.28 dB | STOI: 0.46
Baseline

MPDR

SI-SDR: -5.46 dB | STOI: 0.50
Spectrogram 14
Spectrograms: Noisy, Target, MPDR, cMPDR
Reference

Clean target

Reference Audio
Proposed

cMPDR

SI-SDR: 2.41 dB | STOI: 0.67

SNR 0 to 10 dB
Sample: 11
Noisy

Mixture

SI-SDR: 0.34 dB | STOI: 0.64
Baseline

MPDR

SI-SDR: 4.32 dB | STOI: 0.76
Spectrogram 11
Spectrograms: Noisy, Target, MPDR, cMPDR
Reference

Clean target

Reference Audio
Proposed

cMPDR

SI-SDR: 5.21 dB | STOI: 0.81
Sample: 12
Noisy

Mixture

SI-SDR: 0.82 dB | STOI: 0.62
Baseline

MPDR

SI-SDR: 3.78 dB | STOI: 0.71
Spectrogram 12
Spectrograms: Noisy, Target, MPDR, cMPDR
Reference

Clean target

Reference Audio
Proposed

cMPDR

SI-SDR: 4.52 dB | STOI: 0.72

Dataset 3: MUSAN / Freesound

📝 Description: MUSAN provides a diverse set of non-stationary environmental recordings that do not follow the cyclostationary assumptions of the proposed model. This dataset is used to verify that the beamformer remains robust and does not suffer performance degradation when the noise structure is absent.

📊 Results: Results confirm that cMPDR effectively reverts to standard MPDR behavior in these conditions, preserving baseline performance.

SNR -20 to -10 dB
Sample: noise-free-sound-0458
Noisy

Mixture

SI-SDR: -20.24 dB | STOI: 0.66
Baseline

MPDR

SI-SDR: -17.99 dB | STOI: 0.53
Spectrogram noise-free-sound-0458
Spectrograms: Noisy, Target, MPDR, cMPDR
Reference

Clean target

Reference Audio
Proposed

cMPDR

SI-SDR: -13.42 dB | STOI: 0.55
Sample: noise-free-sound-0826
Noisy

Mixture

SI-SDR: -17.18 dB | STOI: 0.04
Baseline

MPDR

SI-SDR: -16.39 dB | STOI: 0.16
Spectrogram noise-free-sound-0826
Spectrograms: Noisy, Target, MPDR, cMPDR
Reference

Clean target

Reference Audio
Proposed

cMPDR

SI-SDR: -15.53 dB | STOI: 0.17

SNR -10 to 0 dB
Sample: noise-free-sound-0829
Noisy

Mixture

SI-SDR: -11.04 dB | STOI: 0.45
Baseline

MPDR

SI-SDR: -6.40 dB | STOI: 0.54
Spectrogram noise-free-sound-0829
Spectrograms: Noisy, Target, MPDR, cMPDR
Reference

Clean target

Reference Audio
Proposed

cMPDR

SI-SDR: -0.71 dB | STOI: 0.56
Sample: noise-free-sound-0002
Noisy

Mixture

SI-SDR: -10.27 dB | STOI: 0.38
Baseline

MPDR

SI-SDR: -9.46 dB | STOI: 0.55
Spectrogram noise-free-sound-0002
Spectrograms: Noisy, Target, MPDR, cMPDR
Reference

Clean target

Reference Audio
Proposed

cMPDR

SI-SDR: 2.33 dB | STOI: 0.72

SNR 0 to 10 dB
Sample: noise-free-sound-0046
Noisy

Mixture

SI-SDR: 0.61 dB | STOI: 0.81
Baseline

MPDR

SI-SDR: -1.25 dB | STOI: 0.82
Spectrogram noise-free-sound-0046
Spectrograms: Noisy, Target, MPDR, cMPDR
Reference

Clean target

Reference Audio
Proposed

cMPDR

SI-SDR: -0.59 dB | STOI: 0.77
Sample: noise-free-sound-0329
Noisy

Mixture

SI-SDR: 0.38 dB | STOI: 0.72
Baseline

MPDR

SI-SDR: 4.75 dB | STOI: 0.81
Spectrogram noise-free-sound-0329
Spectrograms: Noisy, Target, MPDR, cMPDR
Reference

Clean target

Reference Audio
Proposed

cMPDR

SI-SDR: 4.79 dB | STOI: 0.81

MUSAN noise excerpts are in the US Public Domain or under a Creative Commons license; see Snyder et al. (2015). Freesound clips are used under their respective CC licenses.