Self-Supervised Voice Denoising Network for Multi-Scenario Human-Robot Interaction.

Area of Science:

• Robotics
• Artificial Intelligence
• Speech Processing

Background:

• Human-robot interaction (HRI) using voice commands has advanced with Vision-Language-Action (VLA) models.
• Current VLA systems struggle with environmental noise and overlapping speech in multi-speaker scenarios.
• A specialized denoising network is needed for robust voice command isolation.

Purpose of the Study:

• To enhance voice command-based HRI in noisy environments.
• To improve the performance of self-supervised denoising networks for mixed-noise audio.
• To increase the real-world applicability of voice-guided robot control.

Main Methods:

• Leveraging synthetic data to train a self-supervised denoising network.
• Scaling training data to improve network performance in denoising mixed-noise audio.
• Developing a method to enhance voice command recognition in challenging acoustic conditions.

Main Results:

• The proposed method outperforms existing approaches in simulated environments.
• Achieved 7.5% higher accuracy compared to the state-of-the-art in noisy real-world environments.
• Demonstrated enhanced voice-guided robot control in practical settings.

Conclusions:

• The developed approach effectively enhances voice command recognition for HRI in noisy conditions.
• Synthetic data and self-supervised learning are crucial for improving denoising network performance.
• The method offers a significant advancement for robust voice-guided robot control.