Improved Global Soft Decision Incorporating Second-Order Conditional MAP in Speech Enhancement

Jong-Mo Kum, Joon-Hyuk ChangIEICE Trans. Information and Systems 2010 · DOI

Abstract

In this paper, we propose a novel method based on the second-order conditional maximum a posteriori (CMAP) to improve the performance of the global soft decision in speech enhancement. The conventional global soft decision scheme is found through investigation to have a disadvantage in that the global speech absence probability (GSAP) in that scheme is adjusted by a fixed parameter, which could be a restrictive assumption in the consecutive occurrences of speech frames. To address this problem, we devise a method to incorporate the second-order CMAP in determining the GSAP, which is clearly different from the previous approach in that not only current observation but also the speech activity decisions of the previous two frames are exploited. Performances of the proposed method are evaluated by a number of tests in various environments and show better results than previous work.

Introduction

Speech enhancement systems operating in noisy environments must accurately distinguish between speech signals and background noise. A common technique for this is the Global Soft Decision (GSD), which utilizes a Global Speech Absence Probability (GSAP) to weight spectral estimators. However, traditional GSD methods often rely on a fixed parameter ($q$) to represent the prior probability of speech absence. This approach assumes that speech occurrence is independent across frames, ignoring the quasi-stationary nature of human speech. Consequently, these systems often suffer from "clipping" at the ends of words (speech tails) or fail to suppress fluctuating noise effectively.

Proposed Methodology

The proposed method addresses these limitations by replacing the static prior with a dynamic, second-order Conditional Maximum A Posteriori (CMAP) criterion. Instead of a single fixed value, the algorithm employs four conditional parameters ($q'_{00}, q'_{01}, q'_{10}, q'_{11}$) that represent the likelihood of speech absence given the voice activity decisions of the two preceding frames.

By tracking the state of speech presence over time, the system can adapt its sensitivity. For instance, if speech was detected in the previous two frames, the prior for speech presence in the current frame is significantly increased. This temporal "inertia" allows the algorithm to better preserve low-energy speech segments and maintain natural-sounding speech tails. The resulting dynamic GSAP is integrated into a Minimum Mean Square Error (MMSE) log-spectral amplitude estimator to achieve superior noise suppression.

Experimental Setup

The performance of the proposed CMAP-based method was evaluated using 100 phrases from the NTT database, featuring an equal distribution of male and female speakers. Background noise was simulated using White, Car, and F16 cockpit noise from the NOISEX-92 database at Signal-to-Noise Ratios (SNRs) of 5 dB, 10 dB, and 15 dB. The audio signals were sampled at 8 kHz with a 10ms frame shift for processing.

Results and Discussion

Subjective quality was assessed using Mean Opinion Scores (MOS) from 10 listeners. The proposed method consistently outperformed the conventional GSD approach across all tested noise types and SNR levels.

Extracted table from page 3 of Improved Global Soft Decision Incorporating Second-Order Conditional MAP in Speech Enhancement

As shown in the table above, the proposed method provides a measurable improvement in perceived quality. For example, in F16 noise at 10 dB, the MOS increased from 3.43 to 3.62. Visual analysis of spectrograms further confirms these results, demonstrating a significant reduction in "musical noise" and better preservation of the harmonic structure of speech compared to previous methods.

Spectrogram comparison of (a) clean speech, (b) noisy speech (F16 noise at 10dB), (c) speech enhanced by the previous method, and (d) speech enhanced by the proposed method.

The experimental results highlight the importance of temporal correlation in speech enhancement. The wide variance in the optimized conditional parameters—ranging from 0.0246 for consecutive silence to 479 for speech following a transition—underscores the necessity of a dynamic prior that adapts to the local context of the conversation.

Conclusion

By incorporating second-order temporal dependencies into the global soft decision process, the proposed CMAP-based method significantly improves speech enhancement performance. The dynamic adjustment of the speech absence probability allows for better preservation of speech components and more effective noise reduction, particularly in challenging, non-stationary noise environments.

Summary updated

BibTeX
@article{kum2010improved,
  author = {Jong-Mo Kum and Joon-Hyuk Chang},
  title = {{Improved Global Soft Decision Incorporating Second-Order Conditional MAP in Speech Enhancement}},
  journal = {IEICE Trans. Information and Systems},
  year = {2010},
  doi = {10.1587/transinf.e93.d.1652},
}
Figure