Transcribe speech 100x faster and 100x cheaper with open models