Google’s new AI transcription edits out your ‘ums’ and ‘ahs’

What happened

Google has updated Gemini Audio with three new models built on its Gemini 3.5 generation: Gemini 3.5 Live, Gemini 3.5 Live Experimental, and Gemini 3.5 Transcribe. The update brings automatic detection of specialized jargon, support for more than 85 languages, and improved handling of background noise and imperfect speech — including the filler words, false starts, and verbal tics that tend to clutter raw transcripts.

This piece breaks down exactly what shipped in the Gemini Audio update, how the three models differ, and what developers building voice-controlled or transcription-oriented products need to know before integrating them. Whether you are maintaining an existing Gemini-powered pipeline or evaluating it for a new project, the details below cover what changed, what to verify, and what Google is likely to ship next.

Google rolled out three distinct models under the Gemini 3.5 umbrella for audio work. Gemini 3.5 Live is the production-ready variant aimed at real-time, voice-controlled AI features. Gemini 3.5 Live Experimental sits alongside it as a preview channel, presumably offering access to capabilities not yet fully stabilized for production workloads. Gemini 3.5 Transcribe is the dedicated transcription model, optimized for converting spoken audio into clean written text rather than driving conversational back-and-forth.

How it works

All three models share the core improvements: automatic jargon detection, support for more than 85 languages, and robustness against background noise and non-fluent speech. The filler-word removal — stripping "ums," "ahs," and similar verbal noise from output — appears to be a feature of the transcription pipeline rather than a post-processing step bolted on afterward. Google positioned the update as improving precision across its voice-controlled AI features broadly.

For developers already using earlier Gemini Audio models, the shift to Gemini 3.5 changes the transcription output in a meaningful way. Filler words and disfluencies that previously passed through into raw transcripts will now be suppressed automatically, which alters the character of the output text downstream code may be parsing, displaying, or feeding into further model calls. Any pipeline that previously stripped those tokens itself will need to account for potential double-removal logic or simply remove that preprocessing step.

Why it matters

Jargon detection is the other significant behavioral change. Gemini 3.5 Transcribe is described as automatically detecting specialized terminology, which implies the model is making domain inference decisions at transcription time rather than requiring a vocabulary hint or custom model fine-tune. Builders in legal, medical, technical, or other jargon-heavy verticals should test whether the model's automatic domain detection matches their specific corpus before treating it as a drop-in replacement.

Google has not published a separate SDK version bump announcement alongside this model release, so the upgrade path runs through the Gemini API model selector rather than a package manager update. Developers using the Google AI SDK or Vertex AI should update the model identifier string in their API calls from whichever prior Gemini Audio model they were targeting to gemini-3.5-live, gemini-3.5-live-experimental, or gemini-3.5-transcribe, depending on their use case.

Gemini 3.5 Live Experimental carries the usual experimental-tier caveats: rate limits, potential breaking changes between updates, and no SLA guarantees. Teams running production workloads should target Gemini 3.5 Live or Gemini 3.5 Transcribe and treat the experimental model as a staging ground for evaluating upcoming changes before they roll into the stable tier.

Who is affected

The automatic filler-word removal is the most likely source of subtle regressions. If your application surfaces verbatim transcripts to end users — legal depositions, journalistic interviews, accessibility captions — removing ums and ahs changes the record. Verify whether the suppression is configurable before assuming it can be toggled off, and test edge cases where a word like "um" might appear legitimately in technical contexts or proper nouns.

The 85-plus-language claim covers detection and transcription, but language support quality is rarely uniform across all 85. Builders serving non-English-primary users should run their own accuracy benchmarks on target languages rather than treating the number as a quality floor. Similarly, the background noise robustness improvement does not specify what noise conditions were tested, so applications in extreme acoustic environments — factory floors, outdoor events, phone audio with heavy compression — warrant dedicated evaluation before shipping.

What to watch next

Gemini 3.5 Live Experimental is the leading indicator of where the stable Live model is heading. Watching what Google adds or adjusts in that experimental tier over the coming weeks will telegraph the next stable update before it ships. The language support count is also worth tracking: the more-than-85-languages figure is a floor, and Google has a pattern of expanding language coverage quietly between major announcements.

The broader Gemini 3.5 release cycle is also in motion. Gemini 3.5 Flash and Gemini 3.5 Pro have both shipped in other modalities, and the audio-specific models arriving now suggest Google is systematically applying the 3.5 generation across all input types. Developers should expect the experimental transcription features visible today in Gemini 3.5 Live Experimental to become default behavior in a future stable release without a separate opt-in step.