TLDR: Neuro-MSBG is a novel, lightweight, and end-to-end neural model designed for hearing loss simulation. It significantly reduces computational time (up to 46x faster than traditional methods) and latency, enabling real-time applications. The model incorporates a personalized audiogram encoder and phase-aware modeling, which are crucial for maintaining speech intelligibility and perceptual quality. Its differentiable framework allows seamless integration with speech processing systems, demonstrating its potential for training personalized hearing aid compensators.
Hearing loss affects millions worldwide, and accurately simulating how sound is perceived by individuals with hearing impairment is crucial for developing effective hearing aids and speech processing technologies. Traditionally, models like the Moore, Stone, Baer, and Glasberg (MSBG) model have been widely used for this purpose. However, these existing models often come with significant drawbacks, including high computational demands, noticeable delays, and difficulty integrating into modern, real-time speech systems.
These limitations mean that while traditional models are valuable for research, they struggle to keep up with the demands of real-time applications, such as those needed for everyday hearing aids. They often process sound sequentially, leading to delays that can be disruptive, and their complex structures make them hard to combine with advanced speech enhancement systems.
To address these challenges, researchers have developed a new model called Neuro-MSBG. This innovative model is designed to be lightweight and operate from start to finish (end-to-end), making it highly efficient. A key feature of Neuro-MSBG is its personalized audiogram encoder, which allows it to tailor the simulation precisely to an individual’s unique hearing profile, as captured by their audiogram (a chart showing a person’s hearing sensitivity).
Neuro-MSBG offers several significant advantages. Firstly, it supports parallel processing, meaning it can handle multiple parts of the audio simultaneously. This dramatically speeds up the simulation process. For instance, it can simulate one second of audio 46 times faster than the original MSBG model, reducing the processing time from 0.970 seconds to just 0.021 seconds. This speed makes it highly practical for real-time applications.
Secondly, Neuro-MSBG seamlessly integrates with modern speech processing systems. It resolves the inherent delay issues found in older models, allowing it to be directly incorporated into advanced speech compensator training pipelines. This means that hearing aid algorithms can be trained more effectively, leading to better personalized sound amplification.
Thirdly, and crucially, Neuro-MSBG incorporates phase information in its modeling. While many previous hearing loss models focused only on the magnitude (loudness) of sound, this new model also considers the phase (timing) of sound waves. This phase-aware approach is vital because it significantly improves the fidelity of the simulation, maintaining the intelligibility and perceptual quality of the original MSBG model. Experiments showed strong correlations with objective measures of speech intelligibility (STOI) and perceptual quality (PESQ).
How Neuro-MSBG Works
The model takes normal speech signals and an individual’s audiogram as input. The audiogram is processed by a special ‘Audiogram Encoder’ that transforms it into personalized hearing features. Meanwhile, the speech signal is converted into its time-frequency components (magnitude and phase). These three sets of features – personalized hearing, magnitude, and phase – are then combined and fed into a neural network. This network then predicts the specific magnitude and phase shifts that occur due to hearing loss. Finally, these predicted shifts are used to reconstruct the speech signal as it would be perceived by someone with hearing impairment.
The researchers explored different neural network architectures within Neuro-MSBG, including Mamba, Transformer, LSTM, and CNN blocks. The Mamba-based variant consistently showed the best performance, demonstrating its potential for low-latency, high-fidelity speech modeling. Ablation studies further confirmed the importance of both phase prediction and the specialized Audiogram Encoder for accurate simulation.
Also Read:
- Enhancing Medical Rehabilitation with AI: A New Framework for Motion Data Optimization and Anomaly Detection
- Spiking Neural Networks: Bringing Brain-Inspired AI to the Edge
Real-World Application
Beyond simulation, Neuro-MSBG has practical applications in training hearing aid compensators. By connecting a trainable compensator to a fixed Neuro-MSBG model, the compensator can learn to adjust audio to match an individual’s hearing condition. This setup allows for personalized hearing enhancement. Initial results show that this approach significantly improves speech intelligibility, as measured by the Hearing-Aid Speech Perception Index (HASPI).
In conclusion, Neuro-MSBG represents a significant step forward in hearing loss simulation. Its lightweight, parallelizable, and phase-aware design addresses critical limitations of traditional models, paving the way for more efficient, accurate, and personalized hearing aid technologies. For more technical details, you can refer to the full research paper: Neuro-MSBG: An End-to-End Neural Model for Hearing Loss Simulation.
