Benchmarking MOSS-Transcribe-Diarize 0.9B Model Locally on Apple Silicon
A developer benchmarked the MOSS-Transcribe-Diarize 0.9B model locally on an Apple Silicon MacBook Pro using a 17.5-minute four-person meeting audio file from the AMI corpus. The model successfully identified all four speakers and completed the task in 12 minutes with an extremely low speaker confusion rate of 0.08%. This benchmark demonstrates that lightweight multimodal models can perform joint transcription and speaker diarization locally on consumer hardware with high accuracy. Such capabilities enable efficient, privacy-friendly voice processing workflows without relying on costly cloud APIs. Running in fp32 mode via Apple's mps framework on an M5 Pro MacBook Pro with 64 GB of RAM, the model achieved about 0.68x real-time speed. It also effectively handled overlapping speech by cleanly labeling interrupted segments from different speakers.
## BACKGROUND
Speaker diarization is the process of partitioning an audio stream into homogeneous segments based on speaker identity to answer 'who spoke when'. The AMI Meeting Corpus is a multi-modal dataset consisting of 100 hours of meeting recordings widely used for benchmarking speech recognition and diarization systems.