I wanted an easily deployable way of bringing transcription to an underpowered server I have. I used Moondream's excellent parakeet-redux as a base model and adapted Highway SIMD to run it. The result is a single binary with the model embedded than can provide nearly identical transcription quality while running even faster than the Photon/torch model on CPU targets. Benchmarks are in the repository.