Audio models for speech not understanding pitch, seems similar to how text LLMs often don't understand spelling: it's not what they were trying to recognize.