G-STAR: End-to-End Global Speaker-Tracking Attributed Recognition
Fuente:
arXiv
Gespeichert in:
| Hauptverfasser: | , , , , , , , , , |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2026
|
| Schlagworte: | |
| Online-Zugang: | |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866917547998183424 |
|---|---|
| author | Peng, Jing Chen, Ziyi Li, Haoyu Wang, Yucheng Ma, Duo Li, Mengtian Du, Yunfan Xu, Dezhu Yu, Kai Wang, Shuai |
| author_facet | Peng, Jing Chen, Ziyi Li, Haoyu Wang, Yucheng Ma, Duo Li, Mengtian Du, Yunfan Xu, Dezhu Yu, Kai Wang, Shuai |
| contents | We study timestamped speaker-attributed automatic speech recognition (SA-ASR) for long-form, multi-party speech with overlap. In this setting, chunk-wise inference must preserve meeting-level speaker identity consistency while producing time-stamped, speaker-labeled transcripts. Prior Speech-LLM systems tend to prioritize either local diarization or global labeling, lacking the ability to jointly model fine-grained temporal boundaries and robust cross-chunk identity linking. We propose G-STAR, an end-to-end framework that couples a cache-conditioned speaker-tracking module with a Speech-LLM transcription backbone. The tracker provides structured speaker cues with temporal grounding, and the LLM generates attributed text conditioned on these cues. G-STAR supports component-wise optimization and joint end-to-end training, enabling flexible learning under heterogeneous supervision and domain shift. Under chunk-wise decoding protocols, experiments on both oracle-segmented local evaluation and full-meeting global evaluation show strong speaker-attributed transcription performance. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2603_10468 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | G-STAR: End-to-End Global Speaker-Tracking Attributed Recognition Peng, Jing Chen, Ziyi Li, Haoyu Wang, Yucheng Ma, Duo Li, Mengtian Du, Yunfan Xu, Dezhu Yu, Kai Wang, Shuai Audio and Speech Processing Artificial Intelligence Human-Computer Interaction Multimedia Sound We study timestamped speaker-attributed automatic speech recognition (SA-ASR) for long-form, multi-party speech with overlap. In this setting, chunk-wise inference must preserve meeting-level speaker identity consistency while producing time-stamped, speaker-labeled transcripts. Prior Speech-LLM systems tend to prioritize either local diarization or global labeling, lacking the ability to jointly model fine-grained temporal boundaries and robust cross-chunk identity linking. We propose G-STAR, an end-to-end framework that couples a cache-conditioned speaker-tracking module with a Speech-LLM transcription backbone. The tracker provides structured speaker cues with temporal grounding, and the LLM generates attributed text conditioned on these cues. G-STAR supports component-wise optimization and joint end-to-end training, enabling flexible learning under heterogeneous supervision and domain shift. Under chunk-wise decoding protocols, experiments on both oracle-segmented local evaluation and full-meeting global evaluation show strong speaker-attributed transcription performance. |
| title | G-STAR: End-to-End Global Speaker-Tracking Attributed Recognition |
| topic | Audio and Speech Processing Artificial Intelligence Human-Computer Interaction Multimedia Sound |
| url | https://arxiv.org/abs/2603.10468 |