AI Pulse · Frontier Labs Watch

Google has put a new speech-to-text model, Gemini 3.5 Transcribe, into public preview. On the surface, it turns spoken audio into text. The more important detail is what happens around that transcription. It can automatically remove filler words, recognise custom vocabulary, identify multiple speakers, and work across more than 85 languages.

Google is also placing it across the Gemini API, AI Studio, Gmail, Docs, Gboard, and the Gemini app. That suggests transcription is no longer being treated as a separate utility that produces a file for someone to review. It is being positioned as an input to other AI tasks. A recorded customer call could become a clean transcript, then feed directly into a Gemini workflow for notes, follow-up drafts, issue tagging, or a summary for the account team.

For operators, the practical question is where spoken information is currently getting lost. Sales calls, site visits, internal meetings, voice notes, support conversations, and interviews often contain useful detail that never makes it into your systems. Better transcription alone does not fix that. The value comes from defining what should happen after the words are captured, who needs the output, and where it should be stored.

This is the kind of thing we build into an AI command centre, connecting the raw conversation to a repeatable business process rather than leaving it as another transcript in a folder.