It could be something like a small LLM helping the transcription do its job, and because "Trump" and "racist" appear frequently together in online discourse, they're tied closely together in the model.
You can actually get something similar to happen yourself, too, if you transcribe enough text with Whisper. I've had Whisper reword sentences in ways that don't sound like source audio at all, but still work as what someone might actually say.
You can actually get something similar to happen yourself, too, if you transcribe enough text with Whisper. I've had Whisper reword sentences in ways that don't sound like source audio at all, but still work as what someone might actually say.