Filler Words
Spoken language is full of filler words: “um”, “uh”, “er”, “hmm”. Whether you want them in the transcript depends on what you do with it. The filler_mode query parameter controls this.
By default (filler_mode=natural), you get no guarantee either way: the model keeps some filler words and leaves others out. To make the output consistent, set filler_mode=clean to remove them or filler_mode=verbatim to keep them.
Filler mode is available on the Prerecorded, Realtime, and Turns APIs.
| Mode | Behavior |
|---|---|
clean |
Instructs the model to leave filler words out of the transcript |
natural |
Default. Gives the model no instruction, so it transcribes in its default style - filler words may appear in some places and be left out in others |
verbatim |
Instructs the model to transcribe filler words as they are spoken |
filler_mode=verbatim So, um, I think we should, uh, move the meeting to Friday.filler_mode=clean So I think we should move the meeting to Friday.Choosing a Mode
Section titled “Choosing a Mode”Use clean when people read the transcript or a system processes it further:
- Subtitles, captions, and meeting notes
- Dictation, where the transcript is the document
- Input for an LLM, summarization, search, or analytics
Use verbatim when hesitations carry information:
- Legal, court, or interview transcripts that must reflect exactly what was said
- Linguistic research, speech therapy, and language learning
- Coaching and call quality review, where fluency and confidence are being assessed
- Voice agents that treat “um” as a sign the speaker is still thinking
Use natural if you have no preference either way. Because the model decides per utterance, filler words are not handled consistently. If your application depends on them being always present or always absent, choose clean or verbatim explicitly.
See the API reference for Prerecorded, Realtime, and Turns.