I would like to understand how the --sentence option in whisper-standalone-win is implemented.

[despairTK]

Recently, I have started developing new ASR AI model tools to replace Whisper. I am curious about the implementation of the --sentence flag in whisper-standalone-win: does it rely on NLP-based sentence segmentation?
Currently, I am working on a transcription tool called VibeVoice-ASR to improve transcription accuracy. I have attempted to replicate the --sentence functionality of whisper-standalone-win using NLP-based sentence splitting. However, the results are not quite matching; in segments with high speech density, the timestamps are often inaccurate, typically off by one to two seconds.