lamhaourii/darija-speaker-diarization-finetuning

Fine-tuned pyannote speaker diarization model specifically for Darija (Moroccan Arabic). This project adapts the pyannote.audio pipeline to handle Moroccan dialectal speech, improving speaker segmentation and turn-taking detection for conversational Darija audio. # darija-speaker-diarization-finetuning Fine-tuned pyannote speaker diarization model for Darija (Moroccan Arabic) with 0.12 DER. ## Overview This project improves speaker diarization for Darija by fine-tuning `pyannote/segmentation-3.0` on a synthetic dataset of 2-person conversations, achieving a Diarization Error Rate (DER) of **0.12** on Darija audio. ## Pipeline 1. **Data Collection**: Downloaded videos from Moroccan YouTubers 2. **Clip Extraction**: Chunked videos into small speaker clips → darija-speakers-clips 3. **Synthetic Call Generation**: Created realistic 2-person calls with controlled overlap and silence → darija-synthetic-calls 4. **Model Fine-tuning**: Fine-tuned pyannote segmentation model → speaker-segmentation-darija2 ## Model - **Base Model**: `pyannote/segmentation-3.0` - **Fine-tuned Model**: igitsml/speaker-segmentation-darija2 - **Performance**: DER = 0.12 on Darija conversations ## Application We used this model to create an efficient Darija transcription system by accurately segmenting speakers before transcription. ## Datasets - Darija Speaker Clips - Darija Synthetic Calls