Audio Recognition Using Deep Learning for Edge Devices

This paper has proposed a methodology that creates an automatic speech recognition system, the task for which would be to recognize keywords. Deep learning was deployed for classifying the spoken words. The created audio data set consisted of short audio clips which were then converted to a Spectrogram by computing Short Time Fourier Transform (STFT) of each audio sample from the data. Spectrogram is a picture of spectrum of frequencies of a signal. Convolutional neural network is a deep learning algorithm, prominently used for classifying image data. In our case Spectrogram, which is the audio representation, was used to train the CNN model and the model which achieved a higher recognition rate was deployed on the hardware. The proposed research has been motivated from the requirement of an audio classification system that can be deployed on hardware and further based on the classification the hardware has been assigned certain task to be completed.