Speech enhancement using long short term memory with trained speech features and adaptive wiener filter

Area of Science:

• Signal Processing
• Artificial Intelligence
• Acoustics

Background:

• Speech signals are often degraded by background noise, reducing intelligibility.
• Traditional methods for speech enhancement have limitations in complex noise environments.
• Deep learning offers a promising approach for advanced speech signal processing.

Purpose of the Study:

• To introduce a novel deep learning-based speech enhancement model.
• To improve the clarity and intelligibility of noisy speech signals.
• To leverage Long Short Term Memory (LSTM) for adaptive Wiener filter tuning.

Main Methods:

• A two-phase model: Training and Testing.
• Noise and signal spectra estimation using Non-negative Matrix Factorization (NMF).
• Empirical Mean Decomposition (EMD) feature extraction from the Wiener filter.
• Fractional Delta AMS feature extraction after Bark frequency evaluation.
• Long Short Term Memory (LSTM) model to estimate the Wiener filter tuning factor (η).

Main Results:

• The proposed model successfully enhances de-noised speech signals.
• The LSTM model effectively estimates the Wiener filter tuning factor for improved performance.
• Comparative evaluation shows advantages over existing speech enhancement models.

Conclusions:

• The novel deep learning approach, particularly using LSTM for Wiener filter tuning, significantly enhances speech signals.
• The model demonstrates effectiveness in reducing background noise and improving speech intelligibility.
• This research contributes a robust method for advanced speech enhancement applications.