OpenAI's Whisper is the default answer for speech-to-text: tens of thousands of GitHub stars, deployed everywhere from research notebooks to production transcription pipelines. It is also, out of the box, not optimised for the hardware most people run it on. If you are building real-time transcription or processing large volumes of audio, you meet that ceiling quickly.
The model everyone runs
We pointed Artemis at Whisper running on NVIDIA hardware with one goal: make it faster without changing what it produces. No model surgery, no retraining, no swapping in a different architecture.
What Artemis changed
Artemis profiled the real workload, found where the runtime actually went, and evolved candidate code changes against measured benchmarks. The result was a 25% runtime improvement on NVIDIA GPUs, with zero manual tuning by us and no change to transcription output.
Why it matters
A quarter of the runtime back means lower cloud compute costs on the same fleet, and faster turnaround for latency-sensitive use cases. The interesting part is not the number: it is that the number was sitting there, in one of the most-read open-source AI codebases in the world, waiting for something that measures rather than reads.
