OpenAI Whisper - Robust Speech Recognition via Large-Scale Weak Supervision

This course is focused on developing a robust automated speech recognition system. It begins by providing an overview of the research paper related to this system and moves to collecting a large-scale, weakly supervised dataset. The course then addresses evaluation metric issues and Word Error Rate (WER). It then covers effective robustness, scaling laws in progress and discusses how to decode in an effective way. There is a code walk-through to explain the model architecture and diagram, highlighting how this relates to the code. The transcription task is discussed, including how to load the audio, create mel spectrograms and detect language. There is also information on how to suppress token logits and detect voice activity. Finally, the course ends with decoding and heuristics.