Home

Awesome

Donate

alt text

Standalone executables of OpenAI's Whisper & Faster-Whisper for those who don't want to bother with Python.

Faster-Whisper executables are x86-64 compatible with Windows 7, Linux v5.4, macOS v10.15 and above.
Faster-Whisper-XXL executables are x86-64 compatible with Windows 7, Linux v5.4 and above.
Whisper executables are x86-64 compatible with Windows 7 and above.
Meant to be used in command-line interface or in programs like Subtitle Edit, Tero Subtitler, FFAStrans, AviUtl.
Faster-Whisper is much faster & better than OpenAI's Whisper, and it requires less RAM/VRAM.

Usage examples:

Notes:

Executables & libs can be downloaded from Releases. [at the right side of this page]
Don't copy programs to the Windows' folders! [run as Administrator if you did]
Programs automatically will choose to work on GPU if CUDA is detected.
For decent transcription use not smaller than medium model.
Guide how to run the command line programs: https://www.youtube.com/watch?v=A3nwRCV-bTU
Examples how to do batch processing on the multiple files: https://github.com/Purfview/whisper-standalone-win/discussions/29

Standalone Whisper info:

Vanilla Whisper, compiled as is - no changes to the original code.
A reference implementation, stagnant development, atm maybe useful for some tests.

Standalone Faster-Whisper info:

Some defaults are tweaked for movies transcriptions and to make it portable.
Features various new experimental settings and tweaks.
Shows the progress bar in the title bar of command-line interface. [or it can be printed with -pp]
By default it looks for models in the same folder, in path like this -> _models\faster-whisper-medium.
Models are downloaded automatically or can be downloaded manually from: Systran & Purfview
beam_size=1: can speed-up transcription twice. [ in my tests it had insignificant impact on accuracy ]
compute_type: test different types to find fastest for your hardware. [--verbose=true to see all supported types]
To reduce memory usage try incrementally: --best_of=1, --beam_size=1, -fallback=None.

Standalone Faster-Whisper-XXL info:

Includes all Standalone Faster-Whisper features +the additional ones, for example:
Preprocess audio with MDX23 Kim_vocal_v2 vocal extraction model.
Alternative VAD methods: 'silero_v3', 'silero_v4', 'pyannote_v3', 'pyannote_onnx_v3', 'auditok', 'webrtc'.
Speaker Diarization.
Read more about new features in the Discussions' thread.

paypal