A Siamese Network-Based Framework for Voice Mimicry Proficiency Assessment Using X-Vector Embeddings

This paper evaluates the quality of mimicked speech by computing X-vector based speaker embeddings derived from spectral and prosodic features. These embeddings are processed through a long short term memory (LSTM)-based Siamese Network.The best mimicry artist is first identified through a perception test, and we then examine whether the Siamese Network predicts the same artist. A correct prediction is counted as a hit when the model assigns the highest similarity score (rank-1) to the artist identified by the mean opinion score (MOS). Performance evaluation is conducted on the MIMICz data set using Top-K hit rate. The results highlight the effectiveness of score-level fusion of spectral and prosodic speaker embeddings in assessing voice mimickry competency, with the fusion approach achieving a Top-1 hit rate of 72% and a Top-2 hit rate of 88%.