### What is Speaker Diarization?

Speaker diarization segments audio by speaker identity — determining *when* each person speaks, not *what* they say.

| Task | Output |
|------|--------|
| Speech Recognition (ASR) | Transcript of words |
| Speaker Identification | Identity of known speakers |
| **Speaker Diarization** | Time segments per speaker |

### The Challenge

Given **long-form Bangla audio** (lectures, interviews, conversations), output speaker-wise segments:

```
Audio Input: conversation.wav
     ↓
Model Output:
  [0.0 - 5.2]   → SPEAKER_1
  [5.3 - 12.8]  → SPEAKER_2  
  [13.0 - 18.5] → SPEAKER_1
  [18.6 - 25.0] → SPEAKER_3
```



### Competition Phases

| Phase | Description |
|-------|-------------|
| **Phase I (Online)** | Submit on Kaggle, evaluated on Public/Private test sets |
| **Phase II (Final)** | Hidden test set + On-site presentation for top teams |

### Scoring Breakdown

```
Total Score = 0.70 × Online Score + 0.30 × Offline Score

Online Score  = 0.30 × Inference Time Score + 0.70 × Model Performance
Offline Score = 0.20 × Presentation + 0.20 × Paper + 0.60 × Novelty
```

### Key Rules

| Allowed  | Prohibited  |
|-----------|--------------|
| Open-source pretrained models | External APIs (OpenAI, Claude, etc.) |
| Fine-tuning on competition data | Cloud-based inference services |
| External public datasets (cited) | Manual test set annotation |
| External compute for training | Private code sharing outside team |

**Requirements**: Inference notebook (Kaggle-runnable), IEEE research paper, max 4 members/team, 5 submissions/day.