Real time speech enhancement for the noisy MRI environment

Area of Science:

• Biomedical Engineering
• Signal Processing
• Acoustics

Background:

• Magnetic Resonance Imaging (MRI) environments generate significant noise, often corrupting speech signals.
• Low Signal-to-Noise Ratio (SNR) in MRI audio complicates listener comprehension and increases fatigue.
• Automated speech enhancement is crucial for improving audio quality in noisy medical imaging settings.

Purpose of the Study:

• To evaluate the performance of adaptive and single-channel speech enhancement algorithms in a low-SNR MRI environment.
• To reduce listener fatigue by effectively suppressing strong MRI noise.
• To compare the noise characteristics and convergence speed of different speech enhancement techniques.

Main Methods:

• Testing two adaptive algorithms (Normalized Least Mean Squares - NLMS, Sign-Error LMS) and one single-channel algorithm (LogMMSE).
• Utilizing a floating-point Digital Signal Processor (DSP) for algorithm implementation.
• Using an actual dataset collected from a 3-Tesla MRI machine for experiments.

Main Results:

• The Sign-Error LMS algorithm resulted in residual noise with white noise characteristics, unlike the more structured noise from other methods.
• Sign-Error LMS demonstrated faster convergence compared to NLMS and LogMMSE.
• All tested algorithms showed effectiveness in enhancing speech corrupted by MRI noise.

Conclusions:

• Sign-Error LMS is a promising algorithm for speech enhancement in MRI environments due to its ability to produce white noise residuals and its fast convergence.
• Automated speech enhancement systems can significantly improve audio clarity and reduce listener fatigue in MRI scans.
• Further research can explore optimizing these algorithms for real-time applications in noisy environments.