I've been testing the 'skip filler words' feature in Fireflies.ai, and I think it might be too aggressive. It seems to be removing small but important words, which changes the meaning of some sentences in the transcripts.
For example, in a meeting about access controls, "Don't grant access to the server" became "grant access to the server" because the word "don't" was removed. Has anyone else run into this? Is there a setting to adjust the sensitivity of this filter, or is it just something the AI needs to learn over time?
Yeah, I saw something similar. It turned "We cannot approve that budget" into "We approve that budget." Big difference, right?
Is there a way to give feedback on specific transcripts? Maybe if we flag the mistakes, it'll help the AI learn not to cut out those small negatives.
Learning the ropes
Exactly, and your example highlights a critical class of error. Removing "cannot" or "don't" doesn't just change meaning, it reverses intent. This is a serious flaw for any meeting about compliance, budgets, or security controls.
On the feedback mechanism, yes, there's usually a 'Report an issue' or feedback button on the transcript interface. Flagging these is essential. However, from an architecture perspective, this suggests the model's training data likely underrepresented negations in spoken conversation, treating them as filler noise rather than semantic operators.
I'd advise manually reviewing any transcript with financial or policy implications before distribution. The risk of a reversed statement is currently too high to rely on the automated filter for those.
Latency is the enemy.