TL;DR
This paper introduces EC-Beam, a novel Direction of Arrival (DOA) estimation algorithm that uses only two microphones. By integrating the Equalization-Cancellation (EC) model into a beamforming framework, EC-Beam effectively removes the target signal from a specific direction, allowing for accurate DOA estimation even in noisy environments. This approach offers a practical and robust solution for applications like binaural speech enhancement and humanoid robots.
What Problem Does This Paper Address?
Accurate estimation of the Direction of Arrival (DOA) for sound sources is crucial for various applications, including enhancing speech intelligibility in noisy environments and enabling humanoid robots to interact more naturally with their surroundings. Existing DOA estimation methods often suffer from significant limitations: many require a large number of microphones, making them impractical for compact systems. Others struggle with robustness in noisy conditions or fail to account for the effects of the human auditory system, such as the Head-Related Transfer Function (HRTF), which alters sound as it travels to the ears. This paper aims to bridge this gap by proposing a two-microphone DOA estimation algorithm that is robust to noise and adaptable to system effects.
Methods
The core of the EC-Beam algorithm lies in its integration of the Equalization-Cancellation (EC) model with a beamforming strategy. The EC model, inspired by human auditory perception, aims to "equalize" and then "cancel" the target sound signal.
Here's a breakdown of the EC-Beam process:
- Beamforming with EC-Model: For a given direction, two equalizers are trained to compensate for the amplitude and phase differences between the left and right ear signals. These trained equalizers are then used to cancel the target signal from the observed binaural signals. The idea is that if the beam is pointed directly at the sound source, the target signal will be maximally canceled, resulting in minimal remaining energy.
- Interpolation for Non-Beamformed Directions: To avoid the computational expense of exhaustive beamforming across all possible directions, the algorithm employs cube spline interpolation. This technique estimates DOA values for directions that were not directly beamformed, using values from neighboring beamformed directions. This significantly reduces computational load while maintaining search resolution.
- DOA Estimation: The final DOA is determined by finding the direction where the energy of the canceled signal (i.e., the remaining energy after cancellation) is at its minimum. This minimum energy point corresponds to the estimated direction of the sound source.
Datasets
The experiments utilized the following:
- KEMAR HRTF database: This database, measured at MIT, was used to synthesize speech signals with HRTF effects.
- Recorded speech samples: 110 raw audio samples from 11 Japanese speakers were used. These were convolved with HRIRs to create 4070 signals for 4070 directions.
- Real noise recordings: Noise recorded in a cafeteria at Japan Advanced Institute of Science and Technology (JAIST) was used to evaluate performance in realistic noisy conditions.
- HRIR database from University of Oldenburg: This database, featuring recordings from both in-ear and behind-the-ear microphones, was used to test the algorithm's adaptability to different system configurations.
Key Insights and Ideas
- EC Model for DOA: The paper demonstrates that the psychoacoustically-inspired Equalization-Cancellation model, typically used for speech enhancement, can be effectively repurposed for DOA estimation. This offers a novel perspective on sound localization.
- Two Microphones are Enough: A significant contribution is showing that accurate DOA estimation is achievable with just two microphones, a substantial improvement over methods requiring larger arrays.
- Robustness to Noise: The EC-Beam algorithm shows remarkable resilience to various noise conditions, including single-source, multi-source, and real-world cafeteria noise, maintaining high accuracy even at low SNRs.
- Adaptability to System Effects: The algorithm's performance is evaluated across different microphone placements (in-ear vs. behind-the-ear), indicating its adaptability to various acoustic systems and HRTF effects.
- Interpolation for Efficiency: The use of interpolation is a clever way to balance accuracy and computational cost, making the algorithm more practical for real-time applications.
Key Results
- High Accuracy in Clean Conditions: EC-Beam achieved an accuracy of 98.21% in clean conditions with a standard deviation of only 1.29.
- Robustness to Noise: Even at a 5 dB SNR with one-source noise, the accuracy remained high at 89.95%. With two-source noise and cafeteria noise, accuracies were 93.19% and 92.11% respectively at 5 dB SNR. As SNR increased, accuracy approached that of clean conditions.
- Comparison to GCC-PHAT: EC-Beam significantly outperformed the GCC-PHAT algorithm, especially in HRTF-affected regions, demonstrating its superiority for binaural applications.
- Adaptability: When tested with behind-the-ear microphone signals, EC-Beam achieved an accuracy of 92.97%, showing good adaptability.
Limitations and Future Work
- HRTF Dependency: While the EC-Beam is designed to be adaptable, its performance is still influenced by the HRTF. The paper notes that the accuracy decreased slightly when tested with behind-the-ear microphones compared to in-ear microphones, suggesting that HRTF characteristics play a role.
- Computational Cost: Although interpolation is used to reduce computational expense, the initial beamforming stage can still be intensive.
- Future Work: The authors suggest further research into optimizing the interpolation technique and exploring more advanced EC model variations for even greater accuracy and robustness. Investigating the algorithm's performance with more complex acoustic environments and a larger number of simultaneous sound sources would also be valuable.
Summary updated