Gemini 3.5 Transcribe Turns Raw Audio Directly into Formatted Text

The end of raw transcripts

Google released Gemini 3.5 Transcribe, a new speech-to-text model designed to solve the classic pains of transcription. Instead of just churning out words, it can directly throw out "um" and "uh", handle mid-sentence corrections ("let's meet Tuesday, no, Wednesday"), and format the text right away.

Shifting from transcription to understanding

The difference compared to the older Chirp 3 model is not just in accuracy (Word Error Rate dropped to 4.0% for streaming). The crucial part is the connection to the ecosystem. Transcribe can recognize up to three speakers and return timestamps. More importantly, in the Gemini macOS app, it can directly make function calls from audio and task other models.

Where AI ends and UI begins

Google is pushing this model into its products (Gboard, Antigravity, Chrome), showing that voice input is no longer meant to be just a dictation tool, but a full-fledged way to control the system. For developers, the model is available via the Gemini API in variants for both live streaming and recording processing.

Adoption speed outside Google

The real test will be whether this model convinces developers who are currently building on Whisper or Deepgram. Price, latency, and reliability on edge cases will decide.

Lilith's verdict

Audio transcription is a commodity, so Google is trying to sell intent understanding and machine control directly. The question is whether developers want a smart package from Google, or prefer to build their own logic on top of pure text from Whisper.