CNN-Based Audio Word Comparison: Exploring MobileNet for Similarity Assessment

Deep learning approaches are heavily invested to develop the objective to improve speech communication and impairment problems. One such approach is to calculate the similarity between two audio files with the use of a Mel spectrogram and a convolutional neural network (CNN). The distinctive characteristics of a speaker’s speech must be precisely identified to diagnose the difference in a spoken word. Mel spectrogram provides a time-frequency representation of the audio signal and the CNN-based MobileNet model is used to learn features from these time-frequency representations. To calculate the similarity score of audio files 30 words have been taken for the task to train the MobileNet model from the Google speech command data set version1. The model has been tested by real-time data collected by considering some benchmarks to prove whether this model is suitable for finding similarity scores in spoken words or not.