whisper.cpp v1.8.5

v1.8.5
Added 4
  • Add streaming detect and explicit state reset to VAD
  • Add support for controlling token_timestamps directly in server
  • Add support for carry_initial_prompt in server
  • Return speaker information in JSON in server
Changed 3
  • Set bench data for each iteration in whisper
  • Merge tokens split across UTF-8 boundaries in JSON output in CLI
  • Ruby bindings now transcribe without GVL, accept more MemoryViews, and support Windows
Fixed 7
  • Fix dangling pointers, memory leak, and SEGV on parallel transcription in Ruby bindings
  • Fix max_tokens skipping remaining audio issue
  • Fix params leak between requests in server
  • Fix no_speech_thold not being read in server
  • Fix memory leak in read_audio_data in examples
  • Fix server /inference failing to decode in-memory audio regression
  • Fix incorrect timestamps
Overview

Maintenance release + performance improvements all around:

https://github.com/ggml-org/whisper.cpp/blob/master/scripts/bench-all-gg.txt

What's Changed
New Contributors

Full Changelog: https://github.com/ggml-org/whisper.cpp/compare/v1.8.4...v1.8.5

View original

Upgraded? How did it go?

Discussion