TL;DR
This paper introduces FRIDA, a novel algorithm for estimating the direction of arrival (DOA) of multiple wideband sound sources. FRIDA achieves state-of-the-art resolution even at very low signal-to-noise ratios and works with arbitrary microphone array layouts without requiring computationally expensive grid searches.
What Problem Does This Paper Address?
Estimating the direction of arrival (DOA) of sound sources is crucial for many applications, from audio conferencing and surveillance to robotics and virtual reality. Traditional DOA estimation methods often face a trade-off between resolution (ability to distinguish closely spaced sources), robustness to noise, and flexibility in array geometry. High-resolution methods like MUSIC can distinguish close sources but are typically narrowband and sensitive to noise. Beamforming methods (like Steered Response Power) are more robust but struggle with resolving closely spaced sources. Wideband extensions often require coherent processing, which can be sensitive to initial source location guesses or rely on incoherent processing, losing valuable information. Furthermore, many methods assume specific array geometries (like linear arrays) or require extensive grid searches over possible source locations, which is computationally demanding. This paper aims to develop a DOA estimator that is high-resolution, robust to low SNRs, works with arbitrary array layouts, and avoids grid searches.
Methods
FRIDA leverages recent advances in Finite Rate of Innovation (FRI) sampling. The core idea is that for any arbitrary array layout, the spatial covariance matrix entries can be linearly transformed into a uniformly sampled sum of sinusoids. This transformation allows FRIDA to:
- Coherently combine multi-band information: Unlike incoherent methods, FRIDA uses information across different frequency bands in a unified way, improving robustness and resolution.
- Work with arbitrary array layouts: The method is not restricted to specific geometries like uniform linear arrays.
- Avoid grid search: This is a significant computational advantage over many high-resolution subspace methods.
- Utilize cross-correlations: FRIDA operates on cross-correlations between microphone pairs, which effectively increases the number of measurements and allows for using fewer microphones than sources.
The reconstruction algorithm is framed as a constrained optimization problem, where the goal is to find an "annihilating filter" that, when applied to the transformed spatial covariance matrix, results in zeros. This filter's properties are directly related to the source locations.
Datasets
The paper uses simulated data for evaluating performance across various SNRs and source separations. For experiments with real-world signals, they used recordings from the Pyramic compact array (an EPFL-developed array with 24 microphones arranged in a triangular configuration) with sound sources from loudspeakers.
Key Insights and Ideas
- Spatial covariance matrix as uniformly sampled sinusoids: The key insight is that the spatial covariance matrix, regardless of array geometry, can be expressed as a sum of sinusoids sampled uniformly in the angular domain. This is the foundation for applying FRI sampling techniques.
- Annihilating filters for DOA: The concept of annihilating filters, borrowed from FRI sampling, is adapted to directly estimate source locations from the transformed covariance matrix. The roots of the annihilating filter polynomial correspond to the source DOAs.
- Cross-correlations as a powerful measurement: Using cross-correlations instead of raw microphone signals provides more information and allows for estimating more sources than microphones, a capability not typically found in other subspace methods.
Key Results
- State-of-the-art resolution at low SNRs: FRIDA demonstrates superior performance in terms of reconstruction error, especially at low SNRs (below -20 dB), outperforming methods like MUSIC, SRP-PHAT, CSSM, WAVES, and TOPS.
- Excellent resolution of close sources: FRIDA can resolve sources separated by as little as 11.2°, significantly outperforming other methods that struggle with separations beyond 22.5°.
- Ability to resolve more sources than microphones: In experiments, FRIDA successfully reconstructed 10 sources using only 9 microphones, a feat not achievable by conventional subspace methods.
- Robustness to noise: FRIDA shows strong resilience to additive white Gaussian noise.
Limitations and Future Work
- Handling correlated sources: The current FRIDA formulation, which relies on cross-correlations, cannot handle completely correlated sources. The authors mention a straightforward modification that operates on raw signals to address this, but it requires more microphones.
- Computational complexity: While FRIDA avoids grid search, improving its computational efficiency for practical, real-time applications is identified as an important future direction.
- Extension to full spherical case: The current work focuses on 2D DOA estimation (azimuth). Extending the algorithm to handle full 3D DOA (azimuth and elevation) is a logical next step.
- Direct plane wave processing: The authors aim to extend the algorithm to work directly with plane waves rather than cross-correlation coefficients.
Summary updated