Hacker News Reader

An extremely fast implementation of whisper built on top of Apple MLX. It outperforms Whisper CPP by ~10x.

How did you run whisper.cpp? Can you share the command line? What machine did you use to run that benchmark figure?

Will disable beam search hurt performance on long form transcription?

According to the whisper paper, beam search is necessary to achieve good long form audio transcription performance.

What's the word error rate? Is it the same as the distilled whisper models?

Nice. Why was this tested with a 15 min video instead of an audio file? or is it a typo on the graph?

Are you going to have a Swift version of this?

Lightning Whisper MLX