Speaker Identification using Triplet Loss Function Combined with Clustering Techniques

Speaker identification plays a critical role in many applications like robotics specially the applications that focus on humanoid robotics. The speaker identification includes comparing unknown utterances against pre-stored utterances of speakers. In general, the encoded features are stored from the pre-known speakers database and 1:N comparisons between the extracted encoded features of the unknown utterances and the pre-stored N known speakers are implemented. Different techniques can be used for these types of comparisons of which cosine similarity is the most used one. However, the more the number of the pre-stored known speakers, the longer the execution time the model will need to finish these comparisons, and hence it may not be suitable for real-time applications. In this paper, we combined previously published Triple Neural Network for speaker identification with clustering techniques on the speakers dataset. We employed different clustering techniques and presented two different methods for comparing unknown utterances against pre-stored utterances. The obtained results showed a significant enhancement in the comparisons time with a few reductions in the obtained accuracy. The proposed approach provided a framework that can represent a trade-off between execution time and obtained accuracy.