Vowel speech recognition from rat electroencephalography using long short-term memory neural network.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Speech perception and recognition mechanisms are extensively studied using electroencephalography (EEG) to understand neural underpinnings.
• Deep learning models offer advanced capabilities for automatic feature extraction in speech recognition tasks.
• Identifying neural correlates of phoneme representation is crucial for advancing speech processing technologies.

Purpose of the Study:

• To identify neural components related to phoneme representation in the rat brain.
• To discriminate between different vowel speech stimuli using single-trial EEG data.
• To evaluate the efficacy of bidirectional long short-term memory (BiLSTM) networks for speech recognition in animal models.

Main Methods:

• Utilized EEG recordings from Sprague-Dawley rats' anterior auditory fields during exposure to five distinct vowel stimuli (/a/, /e/, /i/, /o/, /u/).
• Applied minimal preprocessing including z-score normalization to the EEG data.
• Employed a bidirectional long short-term memory (BiLSTM) network and classical machine learning methods for classification, assessed via 10-fold cross-validation.

Main Results:

• The BiLSTM network achieved the highest classification performance, with an overall accuracy of 75.18%, an f1-score of 0.75, and a Cohen's κ of 0.68.
• Demonstrated the capability of BiLSTM to effectively model sequential EEG data for speech recognition.
• Indicated that end-to-end learning with BiLSTM can derive informative features without manual feature engineering.

Conclusions:

• BiLSTM networks are effective for modeling sequential EEG data in the context of speech perception.
• Deep learning approaches, particularly BiLSTM, show significant potential for decoding neural representations of speech stimuli.
• This study provides insights into the neural basis of speech recognition and highlights the utility of AI in neuroscience research.