Abstract
Multi-channel signals captured by spatially separated sensors often contain a high level of data redundancy. A compact signal representation enables more efficient storage and processing, which has been exploited for data compression, noise reduction, and speech and image coding. This article focuses on the compact representation of speech signals acquired by spherical microphone arrays. A polynomial matrix eigenvalue decomposition (PEVD) can spatially decorrelate signals over a range of time lags and is known to achieve optimum multi-channel data compaction. However, the complexity of PEVD algorithms scales at best cubically with the number of channel signals. In contrast, the spherical harmonic transform (SHT) provides a compact spatial representation at a cost that rises only quadratically. In this work, we propose to exploit the compact representation offered by the SHT to reduce the number of channels used for subsequent PEVD processing. We show that the diagonality factor improves by up to 7 dB over the microphone signal representation with a significantly lower computation cost. Moreover, the proposed method improves STOI and SDR by up to 0.2 and 20 dB, respectively.
Introduction
Spherical microphone arrays (SMAs) are critical for capturing 3D sound fields in applications like virtual reality and robot audition. However, the high number of microphones leads to significant data redundancy and computational challenges. Signal compaction—the process of decorrelating these signals into a more efficient representation—is essential for real-time processing.
Traditional approaches like the Spherical Harmonic Transform (SHT) are computationally efficient but fail to account for temporal correlations caused by reverberation. Conversely, Polynomial Matrix Eigenvalue Decomposition (PEVD) offers superior broadband decorrelation across space and time but suffers from cubic computational complexity ($O(Q^3)$) relative to the number of microphones ($Q$). This paper proposes a hybrid SHT-PEVD framework to bridge this gap.
The SHT-PEVD Framework
The proposed framework leverages the spatial efficiency of SHT to reduce the dimensionality of the input before applying the more intensive PEVD.
As shown in the block diagram above, the pipeline consists of three main stages:
- Dimensionality Reduction (SHT): Raw microphone signals are transformed into a smaller set of $L$ eigenbeam signals.
- Eigenbeam Selection: An algorithm selects the most relevant eigenbeams, further reducing the channel count to $\mathcal{P}$.
- Broadband Decorrelation (PEVD): The PEVD is applied to the reduced set of channels. By processing a compact set of eigenbeams rather than raw microphone data, the computational complexity is drastically reduced (e.g., from $32^3$ to $4^3$).
Experimental Results
The framework was evaluated using simulated and real-world data, including the TIMIT corpus and measured Room Impulse Responses from the ACE corpus. The results demonstrate significant improvements in both signal compaction and speech processing metrics.
As illustrated in the variance plot above, the SHT+PEVD approach (blue circles) achieves superior energy compaction compared to SHT alone or SHT combined with the Karhunen-Loève Transform (KLT). Key performance highlights include:
- Compaction Quality: The diagonality factor improved by up to 7 dB.
- Speech Enhancement: STOI improved by up to 0.2 in noisy environments.
- Source Separation: SDR gains of up to 20 dB were observed, significantly outperforming baseline methods like AuxIVA.
Conclusion
The hybrid SHT-PEVD framework provides a computationally efficient yet highly effective solution for signal compaction in spherical microphone arrays. By combining the spatial strengths of SHT with the temporal decorrelation capabilities of PEVD, the system achieves near-optimal performance for speech enhancement and source separation. Future work will focus on real-time adaptive implementations for dynamic acoustic environments.
Summary updated