Causal speech enhancement using dynamical-weighted loss and attention encoder-decoder recurrent neural network.

Area of Science:

• Signal Processing
• Machine Learning
• Acoustics

Background:

• Speech enhancement (SE) is crucial for applications like robust automatic speech recognition (ASR) and mobile communications.
• Existing SE systems require low-latency and efficient optimization for real-world, real-time performance.
• Single-microphone SE presents unique challenges for noise reduction and intelligibility improvement.

Purpose of the Study:

• To develop a single-microphone speech enhancement model optimized for real-time, low-latency causal processing.
• To improve voice quality, intelligibility, and noise suppression in noisy speech signals.
• To evaluate the model's performance against baseline methods using objective metrics and ASR.

Main Methods:

• Proposed a causal data-driven model employing an attention encoder-decoder Long Short-Term Memory (LSTM) network.
• The model estimates a time-frequency mask to reconstruct clean speech from noisy input.
• Introduced a dynamical-weighted (DW) loss function to enhance model learning efficiency.

Main Results:

• The proposed model demonstrated consistent improvements in voice quality, intelligibility, and noise suppression.
• Achieved significant gains in Short-Time Objective Intelligibility (STOI) compared to baseline models (e.g., 6.6% over LSTM-KF).
• Reduced Word Error Rates (WERs) in Google's ASR from 46.33% (noisy) to 13.11% (proposed model).

Conclusions:

• The attention encoder-decoder LSTM model with DW loss offers superior performance for real-time speech enhancement.
• The proposed causal SE model effectively suppresses noise and enhances speech intelligibility, outperforming existing methods.
• Significant reduction in ASR error rates highlights the practical utility of the developed speech enhancement system.