- Multimodal LLM are bad at hearing 'CB beeps' -> (example)
- Models seem tuned to human vocal frequency range (makes sense)
- Solution: Just record yourself saying 'beep-boop' and Morse an LLM!
- Run
python LLMorse.py-> type your prompt (plain text) -> outputs morse-voice translation
- To insert your own voice 'beep-boop':
- I recommend Audacity for this, record:
- DIT (short.wav) duration: ~0.1s, DAH (long.wav) duration ~0.2s
- Adjust pitch (higher=better) if LLM struggles to distinguish short / long (as included)
- Reroute audio (loopback) using e.g. VB-Cable, mix: Voicemeeter
- Enjoy bleeping at the AI -> hallucination-galore ensues. π
- AI: DOG: major doubledot november charlie double-duh. Gee!
- AI: -.MOM..--AND-..-POP
- AI: The word was "compute"!! \π€/
π€£π Link to Video of this incident