Whisper to me like lovers do

How I learned to stop worrying and use transformers/pipeline-ai

December 19, 2023

Dun-dun-dun-dun-dun-dun whirrrrr…




Longer story

So, I was in need of a transcribing solution, and stumbled upon mystic.ai: they are nice, AND they currently give $50/mo (that’s bordering with I love you) in free credits.

Obviously, the first choice was OpenAI’s Whisper as it’s good, has a lot of optimized solutions etc.

Stumbling upon insanely-fast-whisper was a blessing: as mystic.ai charges per compute time, I could potentially make my psychotherapy diary infinite (yeah, like the glove, but the stones’re in my head).

Someone’s gotta tell him it’s not safe to raise your hand like that…

Whisper optimizations

Why all the fuss?

I’ll omit the quirks of pipeline management, as those were largely low skills, but will post some benchmarks of the same 5-minute .oga file.

OpenAI’s default Whisper v2:

My version:

Okay, it’s not a 10x gain, but still.




BibTeX citation:
  author = {Kogan, Zakhar},
  title = {Whisper to Me Like Lovers Do},
  date = {2023-12-19},
  url = {https://teleogenic.com/posts/231219-whisper},
  langid = {en}
For attribution, please cite this work as:
Kogan, Zakhar. 2023. “Whisper to Me Like Lovers Do.” December 19, 2023. https://teleogenic.com/posts/231219-whisper.