Download for Windows x64 10.0 MB
Portable Windows version 8.0 MB
Installation guide
This program requires model files to work; the recommended model is 3.4 GB. A downloader is included.
The installer usually asks for administrator rights. Approving the prompt is optional – see the installation guide for details.
Most changes in this release apply to the file transcription application.
The main improvement is a reworked high-level approach to transcribing audio files.
Based on my tests, this version now produces better text quality than the original OpenAI implementation.
Even though both use the same ML model and weights.
Moved most preferences from the test screen into a new “Advanced options” dialog.
Added specialised narrow matrix BLAS compute shaders,
improving performance when using beam search or best of decoding.
Beam search decoding is now the default.
While transcribing a file, the progress bar shows voice probability in the upper half,
and decoded word timing and probability in the lower half.
When transcribing multiple files in batch mode, the overall progress bar is now based on the combined duration of all audio files.
Corrected a systematic bias in word timestamp generation.
Added a logit filter to greatly reduce invalid UTF-8 sequences in the output text, controlled by a checkbox.