YouTube already generates transcripts of uploaded videos (though not super accurate ones). A simple and straightforward implementation would just pattern match in the transcript text.
No need for fancy AI, on top of whatever technology powers the transcription feature.
It will be the STT they use for their automatic captioning so it should be as accurate as whatever that produces. It would be the height of inefficiency to do a separate STT run over the entire audio just for this gimmick.
It might even just be conventional speech-to-text, which can struggle with accents and poor acoustics, etc.