Starstruk Logo MarkStarstruk

v3.1.0

m-bain/whisperXv3.1.0May 7, 2023by m-bain

AI Summary

Highlights major performance improvements (70x speed) and reduced memory usage. It introduces sentence-level segmentation for transcripts and reassigns speaker diarization to the sentence level for SRT output.

Key Highlights

  • Achieved 70x real-time transcription speed with <8GB GPU memory requirement
  • Each transcript segment is a sentence (using nltk.sent_tokenize)
  • Diarization now assigned per sentence (and outputted to srt)
  • Improved alignment logic

New Features

  • Sentence-level transcript segmentation
  • Sentence-level speaker diarization

Full Release Notes

- 70x real time transcription, <8GB gpu memory requirement  ⚡️⚡️
- each transcript segment is a sentence (using `nltk.sent_tokenize`)
- diarization now assigned per sentence (and outputted to srt)
- clean up on alignment logic