Update README.md

README.md

@@ -5,13 +5,17 @@ tags: []
 # MSA-ASR
 Multilingual Speaker-Attributed Automatic Speech Recognition
 
+### Demo
+
+<video src="https://huggingface.co/nguyenvulebinh/MSA-ASR/resolve/main/demo_sa-asr.mp4" width="640" height="480" controls></video>
+
 ### Introduction
 
 This repository provides an implementation of a Speaker-Attributed Automatic Speech Recognition model. The model performs both multilingual speech recognition and speaker embedding extraction, enabling speaker differentiation.
 
 Model architecture
 
-![MSA-ASR Model](https://github.com/nguyenvulebinh/MSA-ASR/blob/679f7016c1b0610c5ae5f85fae2168096491b464/resource/model.png?raw=true)
+![MSA-ASR Model](https://github.com/nguyenvulebinh/MSA-ASR/blob/main/resource/model.png?raw=true)
 
 
 ### Setup
@@ -30,18 +34,39 @@ Test script:
 python infer.py
 ```
 
+### Training Dataset
+
+*From ASR to SA-ASR dataset:*
+
+- Segment ASR data into single-speaker turns.
+- Match turns into group which may come from the same speaker by using speaker embedding cosine similarity.
+- Pick a few groups, each group a few turns.
+- Concatenate turns in random order.
+
+![MSA-ASR Dataset](https://github.com/nguyenvulebinh/MSA-ASR/blob/main/resource/sa_asr_data_pipeline.png?raw=true)
+
+*In total:*
+
+- 15.5M turns
+- 14k audio hours
+- English only
+
+Dataset is open available in [HF Dataset](https://huggingface.co/datasets/nguyenvulebinh/spk-attribute)
+
 ### Citation
 
 ```bibtex
-@misc{nguyen2025msaasrefficientmultilingualspeaker,
-      title={MSA-ASR: Efficient Multilingual Speaker Attribution with frozen ASR Models}, 
-      author={Thai-Binh Nguyen and Alexander Waibel},
-      year={2025},
-      eprint={2411.18152},
-      archivePrefix={arXiv},
-      primaryClass={cs.CL},
-      url={https://arxiv.org/abs/2411.18152}, 
-}
+@INPROCEEDINGS{10889116,
+  author={Nguyen, Thai-Binh and Waibel, Alexander},
+  booktitle={ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)}, 
+  title={MSA-ASR: Efficient Multilingual Speaker Attribution with frozen ASR Models}, 
+  year={2025},
+  volume={},
+  number={},
+  pages={1-5},
+  keywords={Training;Adaptation models;Limiting;Predictive models;Data models;Robustness;Multilingual;Data mining;Speech processing;Standards;speaker-attributed;asr;multilingual},
+  doi={10.1109/ICASSP49660.2025.10889116}}
+
 ```
 
 ### License