Efficient Self-Attention Model for Speech Recognition-Based Assistive Robots Control

Area of Science:

• Robotics
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Conventional assistive robot interfaces like joysticks are often unintuitive for users with upper body disabilities.
• Vocal commands offer a more natural and accessible alternative for controlling assistive robots and performing Activities of Daily Living (ADL).

Purpose of the Study:

• To develop a novel, lightweight vocal command recognition system for assistive robots.
• To enhance the usability of assistive robots for individuals with upper limb disabilities through an intuitive vocal interface.

Main Methods:

• A novel lightweight vocal command recognition system was developed, leveraging the MobileNet2 architecture and a new self-attention mechanism.
• The system achieved state-of-the-art performance on the Google Speech Commands Dataset (GSCD) for Keyword Spotting (KWS).
• Transfer Learning (TL) was employed to adapt the model to a new French Speech Commands Dataset (FSCD), demonstrating cross-language efficacy.

Main Results:

• The proposed model achieved state-of-the-art Keyword Spotting (KWS) performance on the Google Speech Commands Dataset (GSCD).
• Transfer Learning (TL) significantly improved the model's performance on the cross-language French Speech Commands Dataset (FSCD).
• Real-life robotic arm control was successfully demonstrated using the vocal interface by healthy participants.

Conclusions:

• The developed lightweight vocal command recognition system offers a viable and intuitive interface for assistive robots.
• The novel approach and the French Speech Commands Dataset (FSCD) advance the field of voice-controlled assistive technology.
• This technology holds significant potential for improving the autonomy and quality of life for individuals with upper limb disabilities.