Pattern learning with deep neural networks in EMG-based speech recognition.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Speech Technology

Background:

• Facial electromyographic (EMG) data offers a non-invasive method for capturing speech-related muscle activity.
• Silent Speech interfaces aim to decode speech from physiological signals.
• Accurate classification of phonetic features is crucial for effective Silent Speech communication.

Purpose of the Study:

• To classify phones and phonetic features using facial EMG data.
• To evaluate the performance of Deep Neural Networks (DNNs) against conventional models for this task.
• To visualize and understand the patterns learned by DNNs in EMG data.

Main Methods:

• Utilized facial electromyographic (EMG) data.
• Applied Deep Neural Networks (DNNs) for classification.
• Compared DNN performance with Gaussian Mixture Models (GMMs).
• Developed methods for visualizing learned neural network patterns.

Main Results:

• DNNs achieved significant improvements in classifying phones and phonetic features from EMG data compared to GMMs.
• The study successfully visualized complex patterns learned by the neural network.
• Increasing network depth correlated with the representation of more intricate EMG activity.

Conclusions:

• Deep Neural Networks are highly effective for EMG-based Silent Speech interfaces.
• Visualizing learned patterns provides valuable insights into the DNN's processing of electromyographic signals.
• This approach advances the development of robust and accurate Silent Speech communication systems.