Cross-lingual Speaker Verification: Evaluation on X-Vector Method

Automatic Speaker Verification (ASV) systems accuracy is based on the spoken language used in training and enrolling speakers. Language dependency makes voice-based security systems less robust and generalizable to a wide range of applications. In this work, a study on language dependency of a speaker verification system and experiments are performed to benchmark the robustness of the x-vector based techniques to language dependency. Experiments are carried out on a smartphone multi-lingual dataset with 50 subjects containing utterances in four different languages captured in five sessions. We have used two world training datasets, one with only one language and one with multiple languages. Results show that performance is degraded when there is a language mismatch in enrolling and testing. Further, our experimental results indicate that the performance degradation depends on the language present in the word training data.