whisper-jax

Why we included this project

For teams that have already standardized on Whisper but are hitting its speed limits, this JAX port is worth a serious look. It runs the same model over 70x faster than the original PyTorch code, and on TPUs or multi-GPU machines it can transcribe a half-hour of audio in about thirty seconds. The FlaxWhisperPipeline class handles preprocessing, chunking, and stitching results together, so you get the speedup without managing parallel inference yourself. It also runs on CPU, GPU, and TPU, and works as a standalone script or as an inference endpoint, making it a practical drop-in for transcription services and meeting tools that outgrow the baseline Whisper runtime.