Detecting Lombard Speech Using Deep Learning Approach.

Area of Science:

• Speech processing
• Machine learning
• Acoustics

Background:

• The Lombard effect, characterized by altered speech production in noise, poses challenges for accurate speech detection.
• Robust detection of Lombard speech is crucial for applications requiring clear communication in noisy environments.

Purpose of the Study:

• To propose and evaluate a machine learning-based strategy for detecting Lombard speech in noisy conditions.
• To develop a system capable of near real-time Lombard speech detection for practical applications.

Main Methods:

• A framework combining convolutional neural networks (CNNs) with 2D speech signal representations was developed.
• A threshold-based averaging strategy was introduced to reduce computational cost while maintaining accuracy.
• Experiments were conducted on German and Polish Lombard speech recordings, with and without data augmentation (using alpha channel for speaker gender, F0, and MFCCs).

Main Results:

• The proposed method achieved high detection accuracy in distinguishing Lombard speech from neutral speech.
• The system demonstrated capability for near real-time operation.
• Investigations identified effective CNN structures and 2D speech signal representations for Lombard speech detection.

Conclusions:

• The developed machine learning approach offers a robust solution for Lombard speech detection in noise.
• The strategy is suitable for real-time applications, enhancing performance in challenging acoustic environments.
• Data augmentation using the alpha channel proved beneficial for improving detection performance.