G-STAR-SA-ASR/gstar_streaming_sortformer_8spk_v0

Streaming Sortformer – 8‑Speaker Fine‑tuned Checkpoint

This repository provides a fine‑tuned checkpoint of Streaming Sortformer (Medennikov et al., 2025) that supports up to 8 concurrent speakers in a meeting or conversational scenario. The model performs online speaker diarization and tracking with an Arrival‑Order Speaker Cache (AOSC), which maintains consistent speaker identities across chunks without re‑indexing.

Key features

• Supports up to 8 speakers per meeting.
• Cross‑chunk identity persistence via AOSC.
• Fine‑tuned on English and Chinese real‑world data.
• Designed for chunk‑wise / streaming inference.

Usage

Load the checkpoint with the official Streaming Sortformer codebase. The model accepts audio chunks and outputs frame‑synchronous speaker activity with stable global labels.