Voice Command Recognition Using Biologically Inspired Time-Frequency Representation and Convolutional Neural...

Area of Science:

• Biomedical Engineering
• Computer Science
• Signal Processing

Background:

• Voice command technology is crucial for human-computer interaction in healthcare, enabling hands-free control of medical devices and patient care systems.
• Convolutional Neural Networks (CNNs), originally for image analysis, show promise in speech classification tasks.

Purpose of the Study:

• To investigate the efficacy of cochleagrams, a time-frequency representation mimicking human auditory processing, as input for CNNs in voice command recognition.
• To explore the potential of multi-view CNNs for integrating information from diverse time-frequency speech representations.

Main Methods:

• Utilized cochleagrams, derived using gammatone filters that model the human cochlea's frequency selectivity, as input for CNN classifiers.
• Implemented and evaluated multi-view CNN architectures to combine multiple time-frequency speech representations.
• Tested the proposed methods on a large-scale voice command dataset.

Main Results:

• The cochleagram-based approach achieved high classification accuracy for voice commands.
• Multi-view CNNs demonstrated effectiveness in leveraging different speech representations for improved performance.
• The proposed methods show significant potential for enhancing voice command recognition in healthcare settings.

Conclusions:

• Cochleagrams offer a biologically plausible and effective time-frequency representation for CNN-based speech classification.
• Multi-view CNNs can further enhance voice command recognition by integrating complementary speech features.
• This research advances the application of AI in healthcare through improved voice interface technology.