If you've ever noticed words or phrases appearing in your Voicenotes transcripts that you definitely didn't say, don't worry—you're not seeing someone else's notes or a technical glitch. You're experiencing what's called AI hallucination.
What is AI Hallucination?
AI hallucination happens when an AI transcription model mistakenly creates words or sentences that weren't actually spoken or recorded. Essentially, the AI model "fills in" what it believes it hears, even if nothing was said at all.
Why Does AI Hallucination Occur?
There are several common reasons why AI hallucinations typically happen:
Long Periods of Silence or Pauses:
Extended quiet moments can confuse the AI, prompting it to hallucinate by inserting random words or phrases.
High Background Noise Levels:
Moderate to high levels of background noise (such as traffic, crowded rooms with multiple people speaking, or loud equipment) make it challenging for the AI model to differentiate clearly between your voice and surrounding sounds, causing it to transcribe similar-sounding but incorrect words.
Default Language Settings:
Ensuring the correct language is set in your transcription settings greatly helps reduce hallucinations and improves overall accuracy. If the setting is in 'Detect language' and there is prolonged silence or high background noise, the AI models struggle to decipher what it picks up.
"Names to Remember" Feature:
You might have noticed that words or names you list in the 'Names to Remember' feature frequently show up during hallucinations. That's because the AI prioritizes these listed terms and tends to "hear" them even when they aren't clearly spoken.
Tips to Reduce AI Hallucination:
Set your preferred language accurately before transcribing.
Pause your recording if there's an extended silence, then resume when you're ready to speak again.
Record in quieter environments whenever possible, minimizing background distractions. While Voicenotes enhances audio and can capture even whispers, too noisy recordings might still be challenging for the AI models, even the best-in-class models we use.
Understanding these behaviors can help you better interpret your transcripts and ensure a smoother, more accurate note-taking experience.