A frequency domain approach for speech enhancement with directionality using compact microphone array

Heng Zhang, Qiang Fu, Yonghong YanInterspeech 2008 · OpenReview

Abstract

In this paper, a novel two-element-microphone-array-based speech enhancement algorithm is proposed. This algorithm is designed to achieve better overall performance with relatively small array size. A frequency domain adaptive null-forming is used, in which adaptive noise cancelation is implemented in auditory subbands. And an OM-LSA based postfiltering stage further purifies the output. The algorithm also features interaction between the array processing and the postfilter to make the filter adaptation more robust. This approach achieves considerable improvement on signal-to-noise ratio (SNR) and subjective quality of the desired speech. Experiments confirm the effectiveness of the proposed system.

Introduction

Directional speech enhancement is a critical requirement for modern communication devices like mobile phones and hearing aids. While large microphone arrays can utilize beamforming to isolate a target speaker, compact devices face significant physical constraints. When microphones are placed very close together (often less than 3 cm), traditional spatial filtering methods struggle due to high noise correlation between the sensors. Furthermore, implementing complex filterbanks in the time domain is often too computationally expensive for battery-powered handheld hardware.

This research introduces a frequency-domain speech enhancement system designed to overcome these hurdles. By leveraging auditory subbands and a novel feedback mechanism between array processing and post-filtering, the system achieves robust noise suppression in reverberant environments with minimal computational overhead.

System Architecture

The proposed system, termed Auditory Subband Adaptive Null-forming in Frequency Domain (ASANF-FD), operates in three distinct stages as illustrated in the block diagram below.

Figure 1: Block diagram of the proposed ASANF-FD speech enhancement system, showing the integration of auditory grouping, adaptive null-forming, and post-filtering.

1. Frequency Domain Auditory Subbanding

To reduce computational complexity, the system employs a Short-Time Fourier Transform (STFT) rather than expensive time-domain filterbanks. Frequency bins are grouped into "auditory subbands" based on the Bark scale. This perceptual grouping provides higher resolution at lower frequencies, matching the natural energy distribution of human speech and allowing adaptive filters to converge more effectively.

2. Adaptive Null-Forming (ANF)

Within each auditory subband, the system adaptively generates a "null" (a spatial blind spot) directed at noise sources. Because this process occurs independently across subbands, the system can simultaneously suppress multiple noise sources coming from different directions at different frequencies.

3. SPP-Based Post-filtering and Feedback

The final stage utilizes an Optimally Modified Log-Spectral Amplitude (OM-LSA) filter. The core innovation here is the calculation of Signal Presence Probability (SPP). This SPP value is fed back to the adaptive null-former to control its learning rate. When speech is likely present, the adaptation slows down to prevent the "leakage" of desired speech into the noise-cancellation channel, ensuring the system does not accidentally cancel out the target speaker.

Experimental Evaluation

The system was evaluated in a simulated virtual conference room (6m x 5m x 3m) with a reverberation time ($T_{60}$) of 300ms. The setup used two omnidirectional microphones spaced just 2.125 cm apart. The acoustic scene included desired speech at 0°, a competing speaker at 90°, and a music source at -135°, all positioned 1 meter from the array.

Results and Discussion

The proposed ASANF-FD method was compared against standard Delay-and-Sum (DS) beamforming and traditional time-domain ANF implementations. Performance was measured across Segmental SNR (SSNR), Noise Reduction (NR), and Log-Spectral Distance (LSD).

Figure 2: Performance comparison of the proposed method against benchmarks across (a) Segmental SNR, (b) Noise Reduction, and (c) Log-spectral distance.

As shown in the results, the proposed system (indicated by the diamond markers) consistently outperformed the benchmarks:

  • Noise Reduction: The system maintained a high, stable NR of approximately 30 dB across various input SNR levels, significantly higher than the ~17 dB achieved by standard ANF.
  • Speech Quality: Subjective quality tests using PESQ-MOS showed that in a 5dB SNR environment, the proposed system improved the score from 1.89 (noisy) to 2.81, surpassing both the standard ANF (2.56) and the time-domain version (2.63).
  • Signal Preservation: The lower Log-spectral distance (LSD) values confirm that the SPP-constrained feedback loop successfully preserved the spectral characteristics of the desired speech while removing interference.

Conclusion

The ASANF-FD framework provides a computationally efficient and effective solution for speech enhancement on compact hardware. By integrating perceptual auditory grouping with a robust feedback loop between the spatial processor and the post-filter, the system achieves superior noise suppression and speech quality. Future work will focus on adapting the system for dynamic look directions and testing in more complex, non-stationary real-world environments.

Summary updated

Abstract

In this paper, a novel two-element-microphone-array-based speech enhancement algorithm is proposed. This algorithm is designed to achieve better overall performance with relatively small array size. A frequency domain adaptive null-forming is used, in which adaptive noise cancelation is implemented in auditory subbands. And an OM-LSA based postfiltering stage further purifies the output. The algorithm also features interaction between the array processing and the postfilter to make the filter adaptation more robust. This approach achieves considerable improvement on signal-to-noise ratio (SNR) and subjective quality of the desired speech. Experiments confirm the effectiveness of the proposed system.
BibTeX
@inproceedings{zhang2008frequency,
  author = {Heng Zhang and Qiang Fu and Yonghong Yan},
  title = {{A frequency domain approach for speech enhancement with directionality using compact microphone array}},
  booktitle = {Interspeech},
  year = {2008},
  url = {https://openreview.net/forum?id=o3XQhcWd0w},
}
Figure