Abstract
We propose a novel sparse representation for heavily underdetermined multichannel sound mixtures, where the number of sources significantly exceeds the number of microphones. Operating in the complex Fourier domain, our approach preserves spatial characteristics by jointly estimating a dictionary of spectral and spatial features, a sparse activation matrix, and instantaneous source phases. This Phase-Optimized K-SVD (PO-KSVD) method is applied to the challenging problem of robot ego-noise reduction, demonstrating superior performance over conventional dictionary-based techniques in real-world environments.
Introduction: The Challenge of Ego-Noise
In robot audition, "ego-noise"—the sound generated by a robot's own motors and mechanical movements—presents a significant barrier to effective speech recognition. This noise is particularly difficult to suppress because it is highly non-stationary and spatially complex, often forming a "cloud" of sound sources that outnumber the available microphones.
Traditional approaches like spectral subtraction or beamforming often fail in these "heavily underdetermined" scenarios. Furthermore, common sparse representation methods like Non-negative Matrix Factorization (NMF) typically discard phase information, which is essential for capturing the spatial cues necessary to separate overlapping signals in multichannel audio. Our work addresses this by treating phase as a variable to be optimized rather than ignored.
Methodology: Phase-Optimized Dictionary Learning
We introduce a framework that operates directly in the complex Fourier domain, allowing for the preservation and optimization of phase information.
PO-OMP and PO-KSVD
The core of our approach lies in modeling the captured signal as a sparse combination of atoms from a dictionary, where each atom represents a source's spectral shape and spatial transfer function.
- Phase-Optimized Orthogonal Matching Pursuit (PO-OMP): This greedy algorithm performs sparse coding by selecting the best dictionary atoms to explain a mixture while simultaneously optimizing their instantaneous phases.
- Phase-Optimized K-SVD (PO-KSVD): This learning algorithm alternates between sparse coding (via PO-OMP) and a dictionary update step. The update uses Singular Value Decomposition (SVD) to refine dictionary atoms while accounting for the estimated phases, effectively "learning" the robot's unique noise profile from examples.
- PO-KSVD+: An enhanced variant that utilizes the estimated noise to generate a time-frequency mask, providing further refinement of the target speech signal.
Experimental Setup
The method was evaluated using 4-channel recordings from a NAO humanoid robot performing "waving" (arm movement) and "walking" (walking in place) motions. Target speech was drawn from the GRID corpus, a standard dataset for audio-visual speech recognition. Test mixtures were synthesized by combining clean speech with recorded ego-noise at various levels to simulate realistic robot audition challenges.
Results
The proposed PO-KSVD+ method significantly outperformed standard K-SVD and NMF across all primary metrics, including Signal-to-Distortion Ratio (SDR) and Correct Keyword Recognition (CKR).
As shown in the table above, in the "waving noise" scenario, PO-KSVD+ achieved an SDR of 2.31 dB, a substantial improvement over the -3.91 dB of standard K-SVD and 0.07 dB of NMF. This technical improvement translated directly to functional performance: the CKR rate for the robot jumped from roughly 44% (unprocessed mixture) to 82.9% using PO-KSVD+.
While training the dictionary is computationally intensive (approximately 10 seconds per second of signal), the extraction process (test phase) is efficient, running at roughly 0.59 seconds per second of audio on a standard PC, supporting the potential for near-real-time implementation.
Discussion and Future Work
The primary insight of this work is that phase information contains critical spatial signatures that can be optimized to unmix complex signals. By learning these signatures in the complex domain, PO-KSVD can handle mixtures where sources far outnumber sensors by relying on the assumption of temporal sparsity.
Future research will focus on addressing the non-convex nature of the optimization to avoid local minima and incorporating explicit spatial constraints, such as microphone geometry, to further enhance source localization and separation performance.
Summary updated