How we cut our local transcription error rate roughly in half by switching from whisper.cpp to NVIDIA Parakeet on sherpa-onnx, and what Whisper still does better.
