A hybrid microphone array post-filter in a diffuse noise field

Junfeng Li, Masato AkagiInterspeech 2005

TL;DR

This paper introduces a novel hybrid post-filter for microphone arrays designed to tackle noise in diffuse sound fields, common in car environments. By intelligently combining different filtering strategies for low and high frequencies, it achieves superior noise reduction and speech preservation compared to existing methods.

What Problem Does This Paper Address?

Hands-free devices and automatic speech recognition systems often struggle with background noise corrupting the audio captured by microphones. While microphone arrays offer a solution through spatial filtering, effectively removing both correlated and uncorrelated noise, especially in complex environments like car interiors, remains a challenge. Traditional post-filters often have limitations: some fail to suppress correlated noise, others perform poorly at low frequencies, and some introduce undesirable artifacts like "musical noise." This paper aims to bridge this gap by proposing a post-filter that leverages the characteristics of diffuse noise fields to achieve robust noise suppression across the entire frequency spectrum.

Methods

The core of this paper is a novel hybrid post-filter that operates in two distinct frequency bands:

  • High Frequencies: A modified Zelinski post-filter is employed. This approach is designed to handle spatially uncorrelated noise. The modification is key: it intelligently utilizes the correlation between noise signals on different microphone pairs, a departure from the original Zelinski filter's assumption of zero correlation. This is achieved by analyzing the coherence function of the noise field and grouping microphone pairs based on their noise correlation characteristics.
  • Low Frequencies: A single-channel Wiener post-filter is used. This is chosen for its ability to suppress spatially correlated noise and, importantly, its reduced tendency to produce "musical noise" due to its decision-directed SNR estimation mechanism.

The overall system first applies a beamformer (specifically, a superdirective beamformer is mentioned in the experiments) to the multi-channel noisy signals, followed by this proposed hybrid post-filter. The theoretical foundation of the hybrid filter aligns with the multi-channel Wiener filter framework.

Datasets

The experiments were conducted using:

  • Real-world multi-channel car noise recordings: These were captured in a car environment at speeds of 50 km/h and 100 km/h, primarily consisting of engine and road noise.
  • Clean speech signals: 50 Japanese sentences from the ATR database were used.

The noisy signals were generated by artificially mixing clean speech and the recorded car noise at various Signal-to-Noise Ratio (SNR) levels (-5 to 20 dB). This method allowed for precise control and evaluation using objective measures, as the mixing conditions were known.

Key Insights and Ideas

  • Diffuse Noise Field Characteristics: The paper highlights that in diffuse noise fields, low-frequency noise tends to be highly correlated across microphones, while high-frequency noise is less correlated. This observation is crucial for designing frequency-specific filtering strategies.
  • Frequency-Dependent Filtering: The success of the hybrid approach lies in tailoring the filtering strategy to the noise characteristics of different frequency bands. High frequencies benefit from exploiting spatial correlations (modified Zelinski), while low frequencies require a different approach (single-channel Wiener).
  • Modified Zelinski for Uncorrelated Noise: By considering the correlation between microphone pairs, the modified Zelinski filter can more effectively suppress spatially uncorrelated noise in the high-frequency range.
  • Wiener Filter for Correlated Noise and Musical Noise Reduction: The single-channel Wiener filter in the low-frequency band effectively tackles correlated noise and, importantly, mitigates the "musical noise" artifact often associated with spectral subtraction methods.
  • Theoretical Alignment: The proposed hybrid filter maintains consistency with the established multi-channel Wiener filter framework, providing a theoretically sound approach.

Key Results

The proposed hybrid post-filter demonstrated superior performance compared to the beamformer alone, Zelinski post-filter, McCowan post-filter, and the single-channel Wiener post-filter.

  • Segmental SNR (SEGSNR): The hybrid post-filter achieved the highest SEGSNR improvements across all tested noise conditions and SNR levels. For instance, in the 50km/h condition, it consistently outperformed other methods.
  • Log-Spectral Distance (LSD): The hybrid post-filter, along with the McCowan post-filter, showed the lowest speech distortion (LSD) across various SNRs. Notably, the proposed filter yielded less residual noise ("musical noise") compared to the McCowan post-filter, as confirmed by informal listening tests.

Limitations and Future Work

  • Limitations: The paper acknowledges that the performance of the Zelinski and McCowan post-filters can be degraded by large instantaneous variations in noise, which is less of an issue for the proposed method. The paper also implicitly assumes that the microphone array is pre-steered towards the desired signal and that time delay compensation has been performed.
  • Future Work: Not explicitly detailed in the provided text, but the success in car environments suggests potential applications in other noisy, diffuse sound field scenarios. Further exploration into adaptive transient frequency determination and more sophisticated noise estimation techniques could also be avenues for future research.

Summary updated

BibTeX
@inproceedings{li2005hybrid,
  author = {Junfeng Li and Masato Akagi},
  title = {{A hybrid microphone array post-filter in a diffuse noise field}},
  booktitle = {Interspeech},
  year = {2005},
}
Figure