Robust Speech Recognition via Large-Scale Weak Supervision