Robust minimal variance distortionless speech power spectra enhancement using order statistic filter for microphone array

Tao Yu, John H. L. HansenInterspeech 2009

TL;DR

This paper introduces a novel speech enhancement algorithm for microphone arrays that significantly reduces background noise while preserving speech quality. By leveraging order statistics and focusing on the power spectrum, the method offers improved robustness to real-world issues like microphone inaccuracies and diffuse noise, outperforming traditional approaches.

What Problem Does This Paper Address?

Microphone arrays are widely used for applications like teleconferencing and speech recognition, aiming to isolate and enhance desired speech. However, practical implementations face several challenges: achieving sufficient array gain requires many microphones, speech power is more critical than phase for recognition, array steering vectors are prone to errors due to environmental factors and hardware imperfections, and diffuse noise environments are difficult to handle. Traditional methods like the Minimum Variance Distortionless Response (MVDR) beamformer, while effective in ideal conditions, struggle with these real-world complexities, often leading to degraded performance or requiring computationally expensive post-filtering that can distort the speech. This paper aims to address these limitations by proposing a more robust and computationally efficient enhancement algorithm.

Methods

The core of the proposed method lies in operating in the power spectral domain rather than the phase-sensitive domain used by conventional beamformers. This allows for a more robust model, particularly when dealing with noisy speech. The authors model the background noise's magnitude-squared spectra as following an exponential distribution, a departure from the common Gaussian assumption. They then employ an Order Statistics Filter (OSF), inspired by median filtering, to exploit the non-Gaussianity of this noise distribution.

The OSF works by ordering the observed magnitude-squared spectra from the microphone array and applying a weighting scheme. This approach allows the algorithm to adapt to the noise distribution, effectively reducing output noise power while preserving the speech power spectrum. The authors derive optimal weights for this OSF by minimizing the mean squared error between the enhanced speech and the original speech, subject to a distortionless constraint. This method is computationally simpler as it primarily relies on second-order statistics.

Datasets

The paper evaluates its proposed method using:

  • Artificial Gaussian Noise: This is used to test the algorithm in a controlled, spatially uncorrelated white noise environment.
  • Realistic In-Car Noise (CU-Move database): This dataset represents a more challenging scenario with highly correlated noise, simulating real-world conditions inside a moving vehicle.

The speech data used for experiments is drawn from the TIMIT database, comprising both male and female speakers.

Key Insights and Ideas

  • Power Spectrum Focus is Key: Shifting from phase-sensitive models to the magnitude-squared power spectrum significantly improves robustness to real-world issues like steering vector errors.
  • Non-Gaussian Noise is Exploitable: The exponential distribution of noise magnitude-squared spectra, a consequence of Gaussian noise in the time-frequency domain, can be effectively leveraged by order statistics filters for better noise reduction.
  • Order Statistics Filter for Robustness: OSFs offer a powerful way to adapt to noise characteristics and are inherently more robust to outliers and variations than simple averaging methods.
  • Computational Efficiency: The reliance on second-order statistics makes the proposed method computationally less demanding than some traditional approaches, making it suitable for real-time applications.

Key Results

  • Improved IS Distortion: The proposed method consistently achieves lower Itakura-Saito (IS) distortion compared to the Delay-and-Sum Beamformer (DSB) in artificial Gaussian noise, indicating better preservation of speech spectral characteristics.
  • Significant SNR Improvement: In realistic in-car noise, where noise is highly correlated, the proposed method demonstrates substantially higher Signal-to-Noise Ratio (SNR) improvements than DSB, which offers only marginal gains.
  • Robustness to DOA Errors: The algorithm shows remarkable robustness to Direction-of-Arrival (DOA) errors, maintaining high array gain even when the DOA estimation is inaccurate, unlike DSB which degrades significantly.
  • Effective Noise Reduction: Qualitative examples show that the proposed method significantly reduces residual noise power spectral density, outperforming DSB, especially in correlated noise environments.

Limitations and Future Work

The authors acknowledge several limitations:

  • Equal Amplitude Assumption: The current method assumes that the desired speech in each channel has equal amplitude of squared spectra.
  • Directional Interference: The phase-insensitive model does not explicitly handle directional interference noise.
  • Short-Time Processing Precision: While generally valid, the phase-insensitive model might not be perfectly precise for very short-time processing, potentially limiting performance.

Future work could explore incorporating directional interference handling and refining the model for even more precise short-time processing.

Summary updated

BibTeX
@inproceedings{yu2009robust,
  author = {Tao Yu and John H. L. Hansen},
  title = {{Robust minimal variance distortionless speech power spectra enhancement using order statistic filter for microphone array}},
  booktitle = {Interspeech},
  year = {2009},
}

Loading related content...

Figure