Speech Recognition for the iCub Platform.

Area of Science:

• Robotics
• Artificial Intelligence
• Computational Neuroscience

Background:

• Developing effective automatic speech recognition (ASR) systems for human-robot interaction presents significant challenges.
• Existing ASR systems often require extensive labeled data and struggle in complex acoustic environments.
• Bridging the gap between human speech perception and machine recognition is crucial for intuitive robot communication.

Purpose of the Study:

• To present an open-source software toolkit for building and deploying ASR systems within the YARP platform.
• To enable non-experts to create custom ASR systems for the iCub robot.
• To develop advanced deep learning-based ASR models addressing challenges in human-iCub verbal interactions.

Main Methods:

• The toolkit integrates Python, C++, and shell scripts within the YARP framework.
• It offers codebases for both general ASR system development and specialized bio-inspired/developmental learning approaches.
• Two distinct ASR paradigms are implemented: articulatory (motor cortex-inspired) and unsupervised speech recognition.

Main Results:

• The developed articulatory speech recognition systems demonstrate superior performance compared to strong deep learning baselines.
• Unsupervised speech recognition systems are provided, mimicking infant language acquisition without requiring labeled data.
• Pre-trained deep learning models and the VoCub dataset (2.5h spoken commands) are released to facilitate ASR system adaptation.

Conclusions:

• The open-source toolkit simplifies ASR system creation for non-experts, particularly for the iCub platform.
• Novel articulatory and unsupervised ASR approaches show promise for robust human-robot communication.
• The provided resources, including models and datasets, accelerate research and development in robot speech recognition.