Bangla Diarizz: Domain-Adapted Speaker Diarization for Bengali Long-Form Audio

Speaker diarization-the task of automatically determining who spoke when in a multi-speaker recording-remains a notable open challenge for low-resource languages, and Bengali (Bangla) is no exception. Despite being among the most widely spoken languages globally, Bengali has historically lacked largescale, publicly available diarization corpora and purpose-built models. This paper presents Bangla Diarizz, a domain-adapted speaker diarization system developed for the DL Sprint 4.0 Bengali Speaker Diarization competition (BUET CSE Fest 2026). Our system surgically replaces both the segmentation and embedding modules of the pretrained Pyannote-Community-1 pipeline: the segmentation model is fine-tuned on the competition dataset using the diarizers library, and the speaker embedding model is replaced with WeSpeaker ResNet34-LM. We furthermore introduce a post-processing stitching mechanism that merges short intra-speaker gaps and reduces fragmentation. Our approach achieves a competitive Diarization Error Rate (DER) on the Bengali-Loop benchmark [1] while reducing inference time from 1 hour 22 minutes to approximately 36 minutes on 14 test audio files, representing a 56% wall-clock speed improvement. This work demonstrates that targeted domain adaptation through fine-tuning on limited in-domain data can yield substantial improvements over language-agnostic baselines for under-resourced speech.