The hard part of filler word removal is not finding "um." Any audio tool can find "um." The hard part is deciding which ones to remove.
Take three occurrences of "um" in a 45-minute interview. The first appears at the start of an answer after a technical question -- the speaker is searching for where to begin. The second appears in the middle of a long explanation -- the speaker is managing a complex thought. The third appears right before the speaker says something personal and hesitant. All three are the same word. They are doing completely different things.
Remove all three, and the episode is technically cleaner. But the third one was doing work. Its absence makes the personal moment sound rehearsed, which makes it feel less true.
What filler words actually do
Filler words are placeholders -- verbal signals that the speaker is still holding the floor while they process what to say next. "Um," "uh," "like," "you know," "so" -- all of these are fundamentally the same kind of content: they tell the listener "I am still here, I am still talking, stay with me."
In casual conversation, these are usually inaudible in any meaningful sense. Listeners process them as background. In recorded audio, they become more noticeable because the listener is not doing the active social work of maintaining the conversation -- they are passively following someone else's speech.
But the reason filler words appear matters. A speaker who is thinking -- genuinely working through a complicated idea -- uses different filler word patterns than a speaker who is nervous, or a speaker who has a habitual verbal quirk. The thinking patterns are contextually meaningful. The habitual patterns are not.
How Reverbwell's detection works
Detecting a filler word is a speech recognition problem. We identify the word boundaries and the phoneme sequence. That is solved. The classifier that decides what to do with it is where the real work lives.
The model looks at a window around the detected filler -- the content before and after, the speech rate, the prosodic features (pitch and energy contours), and the conversational context. A filler word that appears at the end of a multi-sentence answer, before a topic shift, has a different context signature than a filler word that appears in the middle of a run-on explanation of a complex topic.
We trained the classifier against human-edited ground truth: pairs of raw recordings and professionally edited versions, with timestamps for what was removed and what was kept. This gave the model examples of editorial decisions, not just acoustic classification. It learned when editors keep fillers, not just how to find them.
The sensitivity dial: what it actually controls
Reverbwell's filler sensitivity control adjusts the threshold at which the classifier acts on detected fillers. At lower sensitivity, only high-confidence removal cases are acted on -- fillers that appear in patterns strongly associated with "clearly habitual" across the training data. At higher sensitivity, the classifier acts on a wider range of cases, including some that require more context to judge.
The practical effect: lower sensitivity removes fewer fillers but removes them more confidently. Higher sensitivity removes more fillers but with a higher rate of cases where a thoughtful editor might have made a different call.
For most solo podcast formats -- narration-heavy, single speaker, rehearsed material -- higher sensitivity works well. The speaker knows the material; the fillers are mostly habitual rather than cognitive. For interview format, lower sensitivity is usually the right choice, for the reasons discussed above. Guests are working through ideas in real time, and the fillers that appear during that process are doing work.
What the model does not do
The model does not remove personality. It removes instances of specific acoustic patterns in contexts where the training data suggests those patterns are not carrying meaningful content.
The concern that "the AI will make me sound like a robot" is understandable and worth addressing directly. It comes from experience with early-generation filler removal tools that did two things badly: they classified all fillers equally (no context sensitivity), and they created audible artifacts at the cut points (the transition from filler to the next word was not smoothed). The result was speech that had the natural verbal texture removed and left audible seams where the cuts were made.
The classifier context-sensitivity addresses the first problem. For the second problem -- smooth transitions -- Reverbwell applies a brief crossfade at cut points, softening the transition between the removed filler and the continuing speech. The default crossfade is 30ms, which is long enough to prevent a hard cut but short enough to be completely inaudible in the final output.
When to adjust and when to leave it
Check the filler removal result on a five-minute segment before committing to the full episode export. Listen specifically for two things: places where the pacing suddenly feels unnatural (a sign of over-removal), and places where fillers are still clearly present and distracting (a sign of under-removal).
If pacing feels unnatural, reduce sensitivity one step and re-check. If fillers are still distracting, increase sensitivity one step. The transcript editor in Reverbwell also lets you manually restore or remove specific instances -- useful for the edge cases where the classifier's call was wrong in a specific moment.
For most episodes, the medium sensitivity setting produces output that sounds like an edited podcast, not like an over-processed one. The personality stays. The ums that were doing cognitive work stay. The ones that were just verbal static are gone.
