Abstract
In this paper, we propose a novel speech enhancement method to improve the performance of the conventional global soft decision which is based on the spectral gradient method applied to the ratio of a priori speech absence and presence probability value ($q$). Conventional global soft decision scheme used a fixed value of $q$ in accordance with the hypothesis assumed, but the proposed algorithm is a technique for improving the speech absence probability which is applied adaptively variable value of $q$ according to the speech presence or absence in the previous two frames and the conditions of the spectral gradient value. Experimental results show that the proposed improved global soft decision method based on the spectral gradient method yields better results compared to the conventional global soft decision technique based on the performance criteria of the ITU-T P. 862 PESQ (Perceptual Evaluation of Speech Quality).
Introduction
In the field of speech enhancement, the "Global Soft Decision" (GSD) framework is widely utilized to estimate the probability of speech presence within a noisy signal. A pivotal parameter in this framework is $q$, defined as the ratio of the a priori probability of speech being absent versus being present.
Traditional GSD implementations typically treat $q$ as a fixed constant. However, real-world acoustic environments are non-stationary; noise levels in a moving vehicle or a crowded street fluctuate significantly. When $q$ is held constant, the system often struggles to differentiate between low-energy speech "tails" and background noise, resulting in either excessive speech suppression or persistent noise artifacts. This research addresses this limitation by introducing an adaptive mechanism that allows $q$ to respond dynamically to the signal's characteristics.
Methods
The proposed algorithm adaptively updates the value of $q$ for every audio frame by integrating two primary metrics: the Spectral Gradient and Temporal Context.
Spectral Gradient ($\Delta(n)$)
The spectral gradient measures the deviation between the current power spectrum and a predicted long-term average. A significant gradient indicates a rapid change in the signal, typically signaling the onset or offset of speech.
Temporal Context and Sigmoid Adaptation
To prevent erratic switching between speech and noise states, the algorithm considers the speech presence status of the two preceding frames. A sigmoid function is then used to smoothly transition the $q$ value based on the spectral gradient.
By utilizing this sigmoid mapping, the system ensures that transitions are natural and less prone to "musical noise" or harsh clipping. If the preceding frames indicate speech, the function adjusts $q$ to favor continued speech presence; conversely, if the context suggests noise, $q$ is adjusted to be more conservative, enhancing noise suppression.
Results and Performance Evaluation
The proposed method was rigorously tested against conventional fixed-$q$ GSD using a dataset of 96 sentences across five noise environments (Babble, Car, Office, Street, and White noise) at SNR levels of 5, 10, and 15 dB.
Key findings include:
- Consistent Quality Gains: The proposed method outperformed the conventional algorithm across all tested noise types and SNR levels. For instance, in car noise at 10 dB SNR, the PESQ score improved from 3.028 to 3.070.
- Improved Signal Retention: Visual analysis of the speech presence probability (as seen in the figure above) demonstrates that the adaptive $q$ method is significantly better at maintaining the signal during word endings, preventing premature cut-offs.
- Robustness to Non-Stationary Noise: By focusing on the spectral gradient rather than absolute energy, the system maintains high performance even when background noise levels shift.
Conclusion
This research demonstrates that replacing fixed probability ratios with an adaptive, sigmoid-based tracking method significantly enhances speech presence uncertainty estimation. By incorporating both spectral gradients and temporal context, the proposed GSD framework provides a more robust solution for speech enhancement in dynamic, real-world environments. Future work will focus on optimizing the computational complexity for real-time implementation in low-power devices such as hearing aids and mobile communication systems.
Summary updated