Abstract
This paper presents a hybrid speech enhancement (SE) method that combines dual-microphone spatial coherence with a Super Gaussian Joint Maximum a Posteriori (SGJMAP) single-microphone gain function. Designed for implementation on smartphones acting as assistive devices for hearing aids, the proposed method introduces a user-adjustable weighting parameter to balance noise suppression and speech distortion. Objective and subjective evaluations demonstrate significant improvements over standard techniques across various noisy conditions.
The Challenge: The "Cocktail Party Problem"
Hearing Aid Devices (HADs) are often limited by their physical size, which constrains battery life and computational power. This makes it difficult to implement complex signal processing needed to solve the "cocktail party problem"—isolating speech in noisy environments.
Existing methods typically fall into two categories, each with distinct drawbacks:
- Statistical Methods (e.g., SGJMAP): These model speech patterns accurately but often introduce "musical noise"—distracting artifacts caused by over-suppression.
- Spatial Methods (Coherence-based): These use dual microphones to filter uncorrelated noise. While effective at suppression, they often result in mechanical-sounding, distorted speech.
Proposed Hybrid Method
The researchers propose a "weighted union" of these two approaches, leveraging the superior processing power and dual-microphone array of a modern smartphone (e.g., Google Pixel).
As shown in the system architecture (Figure 1), the process involves:
- SGJMAP Gain ($G_k$): Uses a Super Gaussian model to capture the natural peaks of human speech more effectively than standard Gaussian models.
- Coherence-Based Gain ($G_{coh}$): Exploits the spatial relationship between two microphones to suppress uncorrelated background noise.
- Weighted Union ($G_{final}$): The core innovation is the final gain function: $G_{final} = \varpi G_k + (1-\varpi) G_{coh}$. The parameter $\varpi$ (omega) allows for a personalized balance; a higher $\varpi$ favors natural speech quality, while a lower $\varpi$ prioritizes aggressive noise suppression.
The system achieved a computational latency of just 12ms per frame on a Google Pixel, making it suitable for real-time assistive hearing applications.
Key Results and Evaluation
The method was tested using the IEEE corpus for clean speech, mixed with machinery and multi-talker babble noise at SNRs of -5 dB, 0 dB, and 5 dB.
Subjective Performance
In Mean Opinion Score (MOS) tests involving 20 human subjects, the proposed method consistently outperformed benchmarks like log-MMSE and pure coherence filters.
As illustrated in Figure 2, the proposed method (red striped bars) maintains a "fair" quality score even at a very noisy -5 dB SNR, where other methods dip into "poor" territory. For machinery noise at 5 dB, the MOS approached a "very good" rating of 4.5.
Objective Metrics
Objective evaluations using PESQ (Perceptual Evaluation of Speech Quality) and CSII (Coherence Speech Intelligibility Index) confirmed these findings. The hybrid approach provided the best balance, with a weighting factor of $\varpi = 0.7$ generally identified as optimal for most users.
Conclusion
By offloading heavy digital signal processing to a smartphone, this research provides a viable path for enhancing the performance of hearing aids in challenging environments. The introduction of a user-controllable weighting factor transforms the hearing aid from a static filter into a personalized, environment-aware assistive tool. Future work may involve automating this weighting adjustment through environment-aware AI.
Summary updated