Hallucination

Hallucination occurs when a language model produces confident-sounding output that is fabricated or contradicts the source material. In speech recognition, hallucination can manifest as the model generating plausible-sounding words or phrases that were never spoken in the audio — especially during silence or background noise.

In text refinement, hallucination might cause the model to add information not present in the original dictation, change the meaning of a sentence, or invent quotes or statistics. Ummless mitigates refinement hallucination by using carefully designed prompts that instruct the model to preserve the speaker's original meaning and only modify style, grammar, and formatting.