A dual microphone coherence based method for speech enhancement in headsets

Mohsen Rahmani, Ahmad Akbari, Beghdad AyadInterspeech 2008

TL;DR

This paper introduces a novel dual-microphone speech enhancement technique that improves upon existing Cross Power Spectrum Subtraction (CPSS) methods, particularly in scenarios with correlated noise. By proposing a new method for estimating the speech cross-power spectrum and integrating it into CPSS, the authors achieve superior performance in both speech recognition preprocessing and as a standalone enhancement system.

What Problem Does This Paper Address?

In modern communication devices like mobile headsets, the distance between the speaker and the microphone is often significant, leading to a low signal-to-noise ratio (SNR). Speech enhancement systems are crucial for improving the quality of noisy speech. While multi-microphone approaches offer advantages over single-microphone methods by leveraging spatial information for noise reduction, they can be computationally expensive. Dual-microphone systems strike a balance. Coherence-based methods, a subset of dual-microphone techniques, work well when noise is uncorrelated but degrade when noise sources are correlated. Existing CPSS methods attempt to address correlated noise by subtracting estimated noise cross-power spectral density (CPSD) from the noisy signal's CPSD. However, the accuracy of these estimations, especially for the speech CPSD, can be a bottleneck. This paper aims to improve the estimation of the speech cross-power spectrum within the CPSS framework to enhance performance in correlated noise environments.

Methods

The core of this paper lies in a modification of the Cross Power Spectrum Subtraction (CPSS) method. The authors first outline the standard CPSS approach, which relies on estimating the coherence function between two microphone signals to derive a spectral modification filter. They then introduce a new technique for estimating the speech cross-power spectrum, denoted as $|P_{S_1S_2}(f,n)|$. This estimation is achieved by introducing a correction term, $v(f,n)$, derived from a Taylor expansion of the relationship between the noisy signal CPSD and the individual speech and noise CPSDs. This term accounts for the phase difference between speech and noise CPSDs, which is often ignored in simpler CPSS methods. The derived $|P_{S_1S_2}(f,n)|$ is then used to compute a more accurate spectral modification filter, $H(f,n)$, which is applied to one of the microphone channels. The paper also details implementation aspects, including estimating the noise CPSD using noise-only frames and approximating the phase of the noise CPSD.

Datasets

The paper utilizes existing datasets for evaluation:

  • TIDIGIT database: Used for generating clean speech signals in the simulated multi-microphone database.
  • Simulated multi-microphone database: Created using an image method technique [10] to simulate various noise sources and a speech source in a virtual room with different microphone configurations.
  • Car noise and babble noise: These specific noise types were used to generate noisy speech signals for PESQ evaluation.

No new datasets were introduced in this paper.

Key Insights and Ideas

  • Correlated Noise is a Major Hurdle: The paper highlights that standard coherence-based methods and even basic CPSS struggle significantly when noise signals from different microphones are correlated.
  • Phase Matters in CPSD Estimation: The authors demonstrate that accurately estimating the phase relationship between speech and noise CPSDs is crucial for improving CPSS performance, especially in correlated noise. Their Taylor expansion-based correction term ($v(f,n)$) directly addresses this.
  • Improved Speech CPSD Estimation Leads to Better Enhancement: By refining the estimation of the speech cross-power spectrum, the proposed method leads to a more effective spectral modification filter, resulting in cleaner speech.
  • Practical Applicability: The proposed method is shown to be effective not only as a standalone speech enhancement tool but also as a valuable preprocessor for Automatic Speech Recognition (ASR) systems, boosting recognition accuracy.

Key Results

The proposed "modified CPSS" method consistently outperformed the baseline "coherence based" and standard "CPSS" methods in both speech recognition and objective quality measures.

  • Speech Recognition: In experiments using a simulated multi-microphone database, the modified CPSS method achieved higher correct recognition rates across various input SNRs compared to the other methods. For instance, at 20 dB input SNR, the modified CPSS significantly improved recognition rates.
  • Objective Quality (PESQ): Using PESQ scores for car noise and babble noise, the modified CPSS consistently yielded higher scores than the baseline methods for all tested input SNRs and microphone pairs. For example, with car noise and a 10 dB input SNR using microphone pair (1,3), the modified CPSS achieved a PESQ score of 2.914, compared to 2.786 for CPSS and 2.505 for coherence-based methods. The paper also notes that performance generally improves with increased microphone distance and higher input SNRs.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that for very low input SNRs (e.g., 20 dB), the enhanced signal might exhibit some distortion, leading to a slightly lower recognition rate than the noisy signal itself. This suggests that while the method reduces noise, it can introduce artifacts.
  • The assumption of a diffuse noise field for estimating the noise CPSD phase might not hold true in all real-world scenarios.
  • Future Work:
    • The authors suggest that further improvements could be made by refining the estimation of the noise CPSD and exploring more advanced Voice Activity Detection (VAD) techniques to better distinguish between speech and noise frames.
  • Investigating the performance of the proposed method in more complex acoustic environments with reverberation and multiple interfering speakers could be a valuable next step.

Summary updated

BibTeX
@inproceedings{rahmani2008dual,
  author = {Mohsen Rahmani and Ahmad Akbari and Beghdad Ayad},
  title = {{A dual microphone coherence based method for speech enhancement in headsets}},
  booktitle = {Interspeech},
  year = {2008},
}
Figure