One voice is all you need: a one-shot approach to recognize you

Abstract

Human knowledge can quickly learn any unfamiliar concepts based on what they have previously learned. Keeping this in mind, researchers tested training models with limited training data in machine learning classification functions.One-shot learning has proven to be effective in the researches of Computer Vision sector, as it works accurately with a single labeled training example and a small number of training sets. By using a single input example from each class, one-shot learning can work more efficiently and quickly. For training the architecture of neural networks to predict similarities between two inputs, one-shot learning employs the Siamese network as neural network architecture. This architecture has been successfully used for various audio-related problems, but its use of one-shot learning in speaker recognition has received less attention. The goal of this thesis is to apply the concept of one-shot learning to classify speakers by extracting specific features, where it uses triplet loss to train the model to learn through Siamese network and calculates the rate of similarity while testing via support set and a query set to recognize the speaker accurately and faster. The proposed system is trained on the LibriSpeech dataset, which contains different audio recordings of speakers. The final one-shot is performed on few previously unseen classes, utilizing only a single sample of each type while making the classification by extracting features from training data and calculating the similarity ratio to recognize the speaker through the proposed model trained by the Siamese network. As we tested for several classes, the accuracy varied: for two classes, we got 100%, for three classes 95%, for four classes 84%, and for five classes 74%, which is significantly better than the other algorithms we tested for our solution. The results suggest that Siamese networks are a viable solution to the challenging one-shot audio classification issue.