Ensemble of Incremental System Enhancements for Robust Speaker Diarization in Code-Switched Real-Life Audios

Identifying individual speaker utterances in overlapped multi-speaker conversations pose a challenging problem in speaker diarization, specifically under multi-lingual scenarios. Standard speech diarization the system consists of a speech activity detector, a speaker-embedding extractor followed by clustering. We improve each of these components from the standard pipeline to enhance the speaker diarization in such complex cases. Our investigation focuses on addressing key sub-aspects of the task like the presence of noise variations, utterance duration variations, inclusion of enhanced ECAPA-TDNN embeddings for robustness etc. Finally, we use the DOVER-LAP approach to combine these system predictions so that complementary advantages of individual systems are efficiently incorporated. Our best-proposed systems outperform the baseline by achieving DER of 27.7% and 28.6% on Phase-1 and Phase-2 of Track-1 blind evaluation sets, respectively.