In-Ear Microphone Speech Data Segmentation and Recognition using Neural Networks

Speech collected through a microphone placed in front of the mouth has been the primary source of data collection for speech recognition. However, this set-up also picks up any ambient noise present at the same time. As a result, locations which may provide shielding from surrounding noise have also been considered. This study considers an ear-insert microphone which collects speech from the ear canal to take advantage of the ear canal noise shielding properties to operate in noisy environments. Speech segmentation is achieved using short-time signal magnitude and short-time energy-entropy features. Cepstral coefficients extracted from each segmented utterance are used as input features to a back-propagation neural network for the seven isolated word recognizer implemented. Results show that a backpropagation neural network configuration may be a viable choice for this recognition task and that the best average recognition rate (94.73%) is obtained with mel-frequency cepstral coefficients for a two-layer network