RESEARCH · Deep Research
Long-Audio Meeting Transcription with Overlapped Speech Diarization via Open-Source MOSS-TD
MOSS-TD by Fudan University is an open-source long-form speech transcription and diarization system engineered for messy multi-speaker scenarios. Even during intense debates where people interrupt and talk over each other, MOSS-TD cleanly isolates overlapping audio tracks and assigns timestamps to every unique speaker.
CORE TOOLS
MOSS-TD (Open Source), PyTorch, ONNX Runtime, Python
BEST FOR
Executive assistants, legal paralegals, investigative journalists, investment analysts, consultants, conference scribes
REQUIREMENTS
PC or server with Python 3.10 and an Nvidia GPU with 8GB+ VRAM (RTX 3060/4060 or CPU mode).
✦ WHY THIS WORKFLOW
Traditional ASR completely fails when speakers talk over one another, outputting garbled nonsense. MOSS-TD excels at overlapping speech disentanglement and long-form speaker tracking under real-world noise.
✦
Interactive Workflow Animation Preview
Step-by-Step Action Checklist
0 / 4 Completed
READY-TO-USE PROMPT TEMPLATE
I am using Fudan University's open-source MOSS-TD for challenging multi-speaker meeting transcription with heavy interruptions. Please provide: 1) [Speaker Enrollment Script]: Python code using 10-second reference clips for 4 key executives to map diarized clusters directly to real names instead of generic "Speaker 0/1"; 2) [Cross-Talk Synthesis Prompt]: LLM prompt to dissect messy multi-party arguments, isolating objections, concessions, and unanimous approvals; 3) [Long-Audio Memory Optimization]: Best practices for batch chunking to process 3+ hour recordings without GPU OOM errors.