SoundSense AI - EEND-OLA Lite Diarizer

• Architecture inspired by EEND-OLA (End-to-End Neural Diarization with Online Attractors), https://arxiv.org/abs/2006.02616
• Custom implementation: 4-block Transformer encoder + dual LSTM heads (OD + PSE)

• Simulated diarization mixtures generated from LibriSpeech train-clean-360 (921 unique speakers, 3000 training samples)

• Part of SoundSense AI hackathon submission (Stage 3: Diarization).
• Determines who is speaking at each 32ms audio frame using Power Set Encoding (8 classes covering up to 3 simultaneous speakers).

• Built for prototype/demo use.
• Trained on simulated mixtures, not real conversational/overlapping speech (e.g. CALLHOME, AMI). Real-world diarization error rate (DER) not yet benchmarked.
• PSE accuracy: 46.9% (vs 12.5% random chance); OD accuracy: 77.5% (vs 50% random chance).
• Performance should be verified on the target environment before deployment.