Microphone Array Post-filter based on Spatially-Correlated Noise Measurements for Distant Speech Recognition

Ken'ichi Kumatani, Bhiksha Raj, Rita Singh, John W. McDonoughInterspeech 2012 · DOI

Introduction

Distant Speech Recognition (DSR) in automotive environments presents a significant challenge due to complex, spatially-correlated noise fields. Traditional speech enhancement techniques often rely on microphone array beamforming followed by a post-filter to suppress residual noise. However, standard post-filters, such as the Zelinski or McCowan filters, typically assume idealized noise fields—either perfectly uncorrelated or perfectly diffuse. In real-world scenarios, such as a car with wind noise or directional interference, these assumptions fail, leading to suboptimal performance. This paper proposes a novel observation-based post-filtering approach that adapts to the actual noise environment in real-time.

Methodology: Observation-Based Estimation

The proposed system shifts from model-based assumptions to direct observation of the noise field. The architecture utilizes a microphone array to isolate both the target speech and the surrounding noise through a dual-path approach.

The proposed system architecture, highlighting the Super-Directive Maximum Negentropy (SD-MN) beamformer and the null-steering path for noise estimation.

As shown in the architecture above, the process involves several key stages:

  1. Target Extraction: A Super-Directive Maximum Negentropy (SD-MN) beamformer is employed to capture the target signal while suppressing reverberation and interference without the signal cancellation issues common in adaptive beamformers.
  2. Noise Fingerprinting: Simultaneously, a null-steering beamformer ($w_{null}^H$) creates a spatial "null" in the direction of the speaker. This allows the system to capture a reference noise signal ($Y_{null}$) that is largely free of target speech.
  3. Dynamic Post-filtering: By comparing the power spectral densities (PSD) of the beamformer output and the null-steered noise reference, the system dynamically calculates a Wiener filter ($H_p$). This filter adapts to the specific, immediate acoustic characteristics of the environment rather than relying on a static mathematical model.

Experimental Setup

The method was evaluated using real-world recordings from a two-microphone array (3.8 cm spacing) installed in a vehicle. The testing suite included 1,000 utterances across eight distinct "stress test" conditions, ranging from engine idling to high-speed highway driving (65 mph) with windows open (the "Wind" condition). To ensure robustness, the acoustic models were trained on the Wall Street Journal (WSJ1) corpus augmented with representative car noise.

Results and Analysis

The proposed observation-based post-filter demonstrated superior performance across nearly all noise conditions compared to traditional methods.

Comparison of Word Error Rates (WER) for different enhancement techniques across various driving conditions.

As illustrated in the performance chart, the "New PF" (Proposed Filter) consistently achieves lower Word Error Rates (WER). Key findings include:

  • Overall Improvement: The proposed method achieved a 26.7% WER, representing a 20.8% relative improvement over a single distant microphone baseline (33.7% WER).
  • Robustness to Directional Noise: In the most challenging "65-Wind" condition, where noise is highly directional and intense, the proposed filter significantly outperformed the McCowan and Zelinski filters, which struggled with the non-diffuse nature of the wind.
  • Synergy: The results confirm that while the SD-MN beamformer is effective at handling reverberation, the addition of the specialized post-filter provides a critical secondary boost in high-noise environments.

Conclusion

By replacing rigid noise field models with real-time measurements via null-steering, this research provides a pragmatic and effective solution for DSR in complex acoustic environments. While the current implementation focuses on a two-microphone array, the observation-based approach offers a robust framework that avoids the pitfalls of static environmental assumptions. Future work will explore scaling this approach to larger microphone geometries and optimizing the computational overhead for embedded automotive systems.

Summary updated

BibTeX
@inproceedings{kumatani2012microphone,
  author = {Ken'ichi Kumatani and Bhiksha Raj and Rita Singh and John W. McDonough},
  title = {{Microphone Array Post-filter based on Spatially-Correlated Noise Measurements for Distant Speech Recognition}},
  booktitle = {Interspeech},
  year = {2012},
  doi = {10.21437/interspeech.2012-107},
}
Figure