Skip to content

v2.6.0

Choose a tag to compare

@rotemdan rotemdan released this 04 May 06:15
· 72 commits to main since this release

Enhancements

  • whisper.cpp: Update packages to use latest whisper.cpp builds. Add new build for CuBLAS GPU support on Linux x64 (requires system-wide installation of the CUDA Toolkit 2.4+, available on path)
  • whisper: Use optimized Top-K selection method, instead of sorting. Should significantly improve decoding speed on smaller models, especially tiny, on fast CPUs with high core counts
  • Lexicons: add support for object valued entries
  • Extend and update English lexicons. Convert word lexicon to object values entries for simplicity and faster load times

Fixes

  • Phone conversion: Fix IPA to Kirshenbaum phone conversion to correctly handle multi-character phonemes

Full Changelog: v2.5.2...v2.6.0