TL;DR
This paper introduces a novel "binaural heterophasic superdirective beamformer" that tackles the issue of white noise amplification in traditional superdirective beamformers. By designing two orthogonal sub-filters, it separates the desired signal (in-phase) from amplified white noise (random phase) in the binaural outputs, leading to improved speech intelligibility, especially in noisy environments.
What Problem Does This Paper Address?
Superdirective beamformers are highly attractive for acoustic signal processing because they can achieve high directivity, effectively suppressing unwanted sounds. However, a significant drawback is their tendency to amplify white noise, particularly at low frequencies. This amplification necessitates the use of extremely low self-noise microphones, which are difficult and expensive to obtain, limiting the practical application of these powerful beamformers. Existing solutions often compromise the directivity factor (DF) to mitigate white noise amplification, resulting in beamformers that are no longer truly "superdirective." This paper aims to overcome this limitation by leveraging psychoacoustic principles to improve speech intelligibility without sacrificing the superdirective property.
Methods
The core innovation lies in the proposed "binaural heterophasic superdirective beamformer." Instead of a single monaural output, this approach generates two binaural outputs. These outputs are derived from two sub-beamforming filters that are designed to be orthogonal to each other. This orthogonality is key: it ensures that the desired acoustic signal in the binaural outputs remains in phase, while the amplified white noise components are rendered in random phase. This phase difference is crucial for the human auditory system to better distinguish the signal of interest from the noise. The design leverages the concept of interaural coherence (IC) to achieve this separation, aiming to minimize the IC of white noise while maximizing the IC of diffuse noise in the binaural outputs. The paper details the mathematical derivation of these orthogonal filters, involving diagonalization of the noise covariance matrix and manipulation of eigenvalues and eigenvectors.
Datasets
Not discussed in the paper. The experiments utilize simulated data and real-world acoustic environments with a loudspeaker and microphone array.
Key Insights and Ideas
- Leveraging Psychoacoustics for Signal Processing: The paper demonstrates how understanding human auditory perception (specifically, how the brain processes interaural phase differences for sound localization and intelligibility) can directly inform the design of signal processing algorithms.
- Orthogonal Filters for Noise Separation: The key insight is that orthogonal filters can decorrelate noise components in binaural outputs, effectively separating them from the coherent desired signal. This is a novel approach to tackling white noise amplification.
- Binaural Output for Enhanced Intelligibility: By creating binaural outputs where the signal is in-phase and noise is random-phase, the system mimics natural auditory cues that aid in sound source separation, leading to improved speech intelligibility.
- Interaural Coherence as a Design Tool: The paper effectively uses interaural coherence (IC) as a metric to guide the design of the beamformer, aiming to minimize noise IC and maximize signal IC.
Key Results
- Theoretical Superiority: The proposed binaural heterophasic superdirective beamformer is theoretically shown to be superior to its conventional monaural counterpart.
- White Noise Decoupling: Experiments demonstrate that the binaural superdirective beamformer achieves an output IC of zero for white noise, while the conventional beamformer has an IC of one. This indicates effective decorrelation of the amplified white noise.
- Improved Intelligibility: Listening tests, based on psychoacoustic studies, suggest an expected improvement in intelligibility of approximately 4 dB at low SNR conditions compared to conventional methods.
- Frequency Invariance: The beampatterns of the proposed beamformer are shown to be largely frequency invariant, which is desirable for processing broadband signals like speech.
- Trade-off Management: The parameter N allows for flexibility in achieving a compromise between white noise gain (WNG) and directivity factor (DF).
Limitations and Future Work
- White Noise Amplification Still Present: While the binaural approach helps in perceptual separation, the paper acknowledges that white noise amplification still occurs, particularly at low frequencies. The output IC in white noise approaches zero, but the underlying amplified noise is still present.
- Dependence on Input SNR: The paper notes that the localization cues of the estimated signals depend more on the desired signal at high SNRs and more on the noise at low SNRs.
- Future Work: The paper suggests that further investigation into maximizing the binaural DF subject to distortionless constraints could lead to alternative forms of the binaural superdirective beamformer. The authors also imply that the parameter N offers a tunable trade-off that can be optimized for specific applications.
Summary updated