"Sortformer: Seamless Integration of Speaker Diarization and ASR by Bridging Timestamps and Tokens,"in Proc. ICML, 2025. (NVIDIA) [Pub.] [HuggingFace] [Code]

👁️‍🗨️ 1. Introduction / 2. Related works

[❓⚠️ Research Question]

—SD+ multi-speaker ASR 의 integration —**Permutation (**speaker-label alignment)

[📎 Related Work] Limitations of previous methods

[🔬 Experiment] Real-word dataset (~4spk)

🗼 3. Proposed Approach: SortFormer

⚠️ 3.1 Permutation Problem in Diarization

3.2 Diarization Model as a Multi-label Binary Classifier

🖐️3.3 Loss Calculation — Arrival-Time Sorting

3.4 Transformer Encoder Learns to Sort — PE

🗼4. Bridging Timestamps and Tokens

🧪 5. Experimental Results