Non-intrusive speech quality assessment with attention-based ResNet-BiLSTM.

Area of Science:

• Computer Science
• Signal Processing
• Speech Technology

Background:

• Online conferencing speech quality is degraded by noise, reverberation, packet loss, and network jitter.
• Evaluating speech quality in real-world scenarios is challenging due to the absence of a clean reference signal.
• Effective non-intrusive speech quality assessment (NISQA) methods are crucial for online communication platforms.

Purpose of the Study:

• To develop a novel network framework for accurate non-intrusive speech quality assessment (NISQA).
• To address limitations of existing methods, particularly the potential loss of contextual information in feature extraction for NISQA.

Main Methods:

• A hybrid network architecture combining ResNet and Bidirectional Long Short-Term Memory (BiLSTM) was proposed.
• ResNet was employed for extracting local acoustic features from speech signals.
• A modified ResNet variant was developed to preserve temporal information, crucial for speech analysis.
• BiLSTM was utilized to model long-term dependencies and sequential characteristics of speech features.

Main Results:

• The proposed ResNet-BiLSTM framework demonstrated a high correlation with subjective Mean Opinion Scores (MOS).
• The method effectively assessed the quality of clean, noisy, and processed speech signals.
• The modified ResNet component successfully preserved vital time-series information.

Conclusions:

• The developed NISQA method offers a robust solution for evaluating speech quality in challenging online conferencing environments.
• The integration of ResNet and BiLSTM provides an effective approach for capturing both local and sequential features in speech.
• This framework advances the capability of automated speech quality assessment without requiring original clean speech signals.