Summary
Fahd Mirza installs and tests MOSS-Transcribe-Diarize from OpenMOSS on a local Ubuntu system with an NVIDIA RTX 6000 GPU, demonstrating the model’s ability to transcribe audio and separate speakers in a single end-to-end pass. Unlike traditional pipelines that run an ASR system and a diarization system separately and then attempt to align their outputs, MOSS-Transcribe-Diarize produces one time-stamped transcript with anonymous speaker labels (S01, S02) in a single inference run — targeting long-form multi-speaker recordings such as meetings, calls, podcasts, and lectures.
The architecture is a standard speech-LLM setup: a Whisper medium-style audio encoder converts the waveform to features, a small adapter compresses and maps those into the embedding space of a Qwen 3.6B text decoder, and the decoder autoregressively writes the labeled transcript. Total parameter count is under one billion, allowing it to run on a single GPU or on CPU with sufficient RAM. Mirza tests it on a noisy single-speaker clip and a two-speaker English telephone conversation from the CallHome dataset, finding accurate speaker separation and tight timestamp alignment.
The video includes the full installation walkthrough from cloning the repo through running inference, and briefly covers the model’s acoustic event annotation capability. Mirza notes that multilingual support beyond Chinese and English is unconfirmed in the HuggingFace model card. The model and code are publicly available on HuggingFace and GitHub.
📺 Source: Fahd Mirza · Published July 12, 2026
🏷️ Format: Tutorial Demo
