Direction of Arrival Estimation Based on Subband Weighting for Noisy Conditions

Wei Xue, Wenju LiuInterspeech 2012 · OpenReview

TL;DR

This paper introduces a novel method for estimating the direction of arrival (DOA) of sound sources in noisy environments by leveraging subband processing and adaptive weighting. By decomposing the audio signal into frequency subbands and emphasizing those with higher speech content, the proposed method significantly improves DOA estimation accuracy and robustness compared to existing techniques, especially under low Signal-to-Noise Ratio (SNR) conditions.

What Problem Does This Paper Address?

Accurately determining the direction from which a sound originates is crucial for many applications, including video conferencing, robotics, and audio signal processing. However, existing Direction of Arrival (DOA) estimation methods often struggle in environments with significant background noise (low SNR) or reverberation. Traditional approaches like MUSIC (based on subspace analysis) and SRP-PHAT (a beamforming technique) can degrade in performance under these challenging conditions. Time Difference of Arrival (TDOA) based methods, while effective in moderate noise, also suffer when noise levels increase. This paper aims to bridge this gap by developing a DOA estimation technique that is more resilient to noise.

Methods

The proposed method builds upon a Time Difference of Arrival (TDOA) estimation framework. The core innovations are:

  1. Gammatone Filterbank Decomposition: The incoming broadband audio signal from each microphone is passed through a bank of 64 overlapping Gammatone filters. This decomposes the signal into narrowband subband signals, mimicking human auditory processing. The key idea is that different frequency bands might be affected by noise differently.
  2. Subband TDOA Estimation: TDOA is estimated independently within each frequency subband. To handle the discrete nature of digital signals and achieve precise TDOA estimation, a subsample shifting technique based on the Discrete Fourier Transform (DFT) and Inverse DFT (IDFT) is employed.
  3. Novel Cost Function: A cost function based on the determinant of the spatial correlation matrix (SCM) is used for TDOA estimation within each subband. This function is designed to be maximized when signals are well-aligned. The SCM is smoothed over time to improve stability.
  4. Adaptive Subband Weighting: This is the central contribution. The paper proposes calculating a weight for each subband based on its energy. The assumption is that subbands with higher energy are more likely to contain speech signals and thus be more reliable for DOA estimation. These weights are then used to create a fullband cost function as a weighted sum of the subband cost functions. This effectively emphasizes the estimation results from more "speech-like" subbands.
  5. Normalization: Before weighting, the subband cost functions are normalized to ensure that subbands with little or no speech content (and thus low cost function values) have minimal impact on the final estimation.

Datasets

The experiments were conducted using synthetic data generated to simulate a realistic acoustic environment.

  • Room Impulse Responses: Modeled using an image-source method for a 6x4x3 meter room.
  • Microphone Array: A linear array of four omnidirectional microphones with 0.08m spacing.
  • Speech Source: Simulated speech signals with durations of 10 seconds, sampled at 8KHz.
  • Noise: Both white Gaussian noise and colored pink noise were used to evaluate performance under different noise conditions.
  • SNR Range: Experiments were performed across a range of SNRs from -10dB to 20dB, with a step size of 5dB.
  • DOA Range: DOA was varied from -90° to 90° in 5° steps.
  • Evaluation Metrics: Accuracy (defined as the percentage of frames where the estimated DOA is within 5° of the true DOA) and Root Mean Square Error (RMSE) were used.

Key Insights and Ideas

  • Frequency-Dependent Noise Robustness: The core insight is that not all frequency bands are equally affected by noise. By analyzing and weighting subbands, we can leverage the cleaner parts of the spectrum.
  • Energy as a Proxy for Speech Presence: The assumption that higher energy in a subband implies a higher probability of speech is a practical heuristic that proves effective. This allows for adaptive weighting without explicit speech detection.
  • Subband Processing for Broadband Signals: Decomposing broadband signals into subbands is a powerful technique for handling complex acoustic scenarios, allowing for more targeted processing.
  • Improved Performance in Low SNR: The subband weighting approach demonstrates a significant advantage in low SNR conditions, where traditional methods falter.

Key Results

  • The proposed subband weighting method consistently outperforms SRP-PHAT and broadband MUSIC algorithms in terms of accuracy and RMSE, especially in highly noisy environments (low SNR).
  • Under white Gaussian noise, the proposed method shows significant accuracy improvements at low SNRs and comparable performance to broadband MUSIC at higher SNRs.
  • Similar improvements are observed when using colored pink noise.
  • The algorithm exhibits less performance degradation than the compared methods when moving from white to colored noise.
  • Qualitative analysis of DOA estimates for a continuous speech signal shows the proposed method provides more stable and accurate tracking of the true DOA.

Limitations and Future Work

  • Control Factor 'p': The paper uses a control factor 'p' (set to 2) to adjust the influence of subband energy on the weighting. While a tradeoff is discussed, the optimal selection of 'p' for various scenarios might require further investigation.
  • Assumption of Noise Energy Distribution: The method relies on the assumption that noise has a flatter energy distribution across subbands than speech. While generally true, highly structured or non-stationary noise might challenge this assumption.
  • Computational Complexity: While not explicitly detailed, subband decomposition and processing can increase computational load compared to simpler single-band methods.
  • Future Work: The authors suggest that further research could explore more sophisticated subband weighting strategies, potentially incorporating explicit speech detection or more advanced noise modeling. Investigating the performance in reverberant environments would also be a valuable extension.

Summary updated

BibTeX
@inproceedings{xue2012direction,
  author = {Wei Xue and Wenju Liu},
  title = {{Direction of Arrival Estimation Based on Subband Weighting for Noisy Conditions}},
  booktitle = {Interspeech},
  year = {2012},
  url = {https://openreview.net/forum?id=E2aQs-UaP3p},
}
Figure