Abstract
Traditional speech enhancement algorithms often rely on the assumption that speech signals in the frequency domain follow a Gaussian distribution. However, this model frequently fails to capture the sparse, "peaky" nature of human speech, leading to artifacts like musical noise. This paper proposes a speech enhancement framework based on a complex Laplacian probability density function (PDF). By empirically validating the Laplacian fit and integrating it into a soft-decision gain modification rule, the proposed method achieves superior noise suppression and speech quality across various environments compared to standard Gaussian-based approaches.
Introduction
Speech enhancement is a critical component in modern digital signal processing, essential for applications ranging from mobile telephony to voice-controlled smart devices. Most classical algorithms operate in the Discrete Fourier Transform (DFT) domain and rely on a fundamental statistical assumption: that clean speech and background noise follow Gaussian (normal) distributions. While the Gaussian model is mathematically tractable, it often fails to accurately represent the true nature of speech, which tends to be more heavy-tailed. This statistical mismatch leads to sub-optimal noise suppression and the introduction of "musical noise" artifacts. This research seeks to identify a more accurate statistical template to better distinguish between silence, noise, and human speech.
Methodology
The paper introduces a speech enhancement framework centered on the complex Laplacian distribution. The methodology involves several key technical innovations:
- Statistical Validation: The author employs the Kolmogorov-Smirnov (KS) test to empirically demonstrate that the Laplacian model fits noisy speech spectra more accurately than the conventional Gaussian model.
- Complex Laplacian Modeling: The real and imaginary components of the DFT coefficients are modeled as independent Laplacian variables. Although these components are technically "weakly dependent," the author justifies this simplification through scatter plot analysis, showing that the resulting modeling error is negligible.
- Global Speech Absence Probability (GSAP): Utilizing the Laplacian model, a new formula for the Likelihood Ratio (LR) is derived. This ratio is used to calculate the probability of speech presence in specific frequency bins over time.
- Soft Decision Gain Modification: Rather than using a binary filter, the algorithm employs a "soft decision" approach. It modifies the standard Ephraim-Malah suppression rule by weighting the gain based on the Laplacian-derived speech presence probability.
Experimental Setup
The proposed method, designated CL-SEGSD (Complex Laplacian - Spectral Enhancement based on Global Soft Decision), was evaluated using the following resources:
- Datasets: The NOISEX-92 Database, a standard benchmark containing various real-world noise profiles.
- Noise Conditions: Testing was conducted against white noise, babble noise (crowd chatter), and vehicular noise.
- Speech Samples: 10 test sentences, balanced between male and female speakers.
- Evaluation Metrics: Performance was measured at Signal-to-Noise Ratios (SNR) of 5dB, 10dB, and 15dB using PESQ (Perceptual Evaluation of Speech Quality) scores, which correlate closely with human auditory perception.
Results and Discussion
The CL-SEGSD method demonstrated consistent improvements over the standard Gaussian-based SEGSD across all tested noise types and SNR levels.
- Babble Noise: In difficult babble noise environments at 5dB SNR, the PESQ score improved from 1.97 (Gaussian) to 2.09 (Laplacian).
- White Noise: At 10dB SNR in white noise, the score rose from 1.85 to 1.90.
The results indicate that by switching to a Laplacian-based Likelihood Ratio, the system becomes significantly more sensitive to the spectral peaks of speech. This allows the algorithm to preserve the speech signal more effectively while suppressing the surrounding noise floor. Furthermore, the Laplacian model is used to update noise spectrum estimates, creating a mathematically consistent enhancement loop.
Conclusion
This research confirms that the Laplacian distribution is a superior descriptor for speech signals compared to the traditional Gaussian model. By adopting the complex Laplacian PDF, the CL-SEGSD framework provides a more robust solution for speech enhancement in diverse noisy environments. Future work may involve exploring joint distribution models to account for the weak dependency between real and imaginary signal components and conducting subjective listening tests to further validate the perceived naturalness of the enhanced audio.
Summary updated