Homophone Disambiguation Reveals Patterns of Context Mixing in Speech Transformers