Gemini 3.5 Transcribe Scrubs Filler Words from Audio

Google is rolling out an update to Gemini Audio, featuring the new Gemini 3.5 Transcribe model. Beyond supporting over 85 languages, it does one crucial thing: during transcription, it can independently recognize and delete filler words like “um” and “uh” without breaking the sentence's context.

Mechanical work disappears for editors and assistants

Anyone who has ever transcribed an interview or a meeting knows the phase of manually cutting through the fluff so the text makes sense. The new model crosses out this layer of work. It also allows defining a custom vocabulary for specific jargon, so the automation doesn't trip over technical terms it might otherwise consider noise.

Clean text straight out of the box as a native API feature

While speech-to-text transcription is handled by many, a clean transcript that can be immediately published or inserted into minutes was previously the domain of specialized third-party software. Google is now packaging this layer into a single API.

Speaker detection only works for up to three people

While the tool can distinguish who is currently speaking, it has a limit. It can reliably assign voices to a maximum of three speakers in pre-recorded audio. It is not yet sufficient for larger panel discussions or chaotic meetings and will continue to require human oversight.

The purity of corporate databases will determine adoption

How quickly the tool catches on will be decided by its reliability on local data. If it turns out that Transcribe doesn't delete pieces of technical names along with the filler, it will become an invisible part of the daily workflow in other applications.

Lilith's verdict

A clean transcript without 'ums' sounds like a triviality, but in reality, it saves hours on every podcast or research interview.