What happened
Google DeepMind has introduced Gemini 3.5 Transcribe, a new speech-to-text model. According to the official announcement, the model is designed to convert raw audio into polished, formatted text, automatically handling background noise, jargon, and filler words.
It is available in public preview via the Gemini API in Google AI Studio. The model offers two APIs — one for real-time streaming and another for pre-recorded audio. Key features include custom vocabulary, support for over 85 languages, and multi-speaker identification.
How the room's reading it
The initial reaction from the developer ecosystem is focused on the model's intelligence beyond simple transcription. Early partners like LangChain and Vercel are highlighting its potential for building high-performance voice interfaces without heavy infrastructure management. The consensus is that the 'smart transcription' features — which remove filler words and handle self-corrections — are the main attraction for anyone building interactive voice applications.
Infra teams are noting the performance claims from third-party testing, which cite an average Word Error Rate of 4.0% for streaming. This is seen as a significant improvement over Google's previous models, especially for real-world, noisy environments.
Sailfish's take
We see this as more than just a better transcription API. The real value is that the model handles the messy, human parts of speech — the 'ums', 'ahs', and self-corrections — before the text even hits your application logic. This cleans up the input signal in a way that typically requires brittle, custom post-processing.
We've shipped enough voice products to know that transcription errors are a primary cause of user frustration. By moving the cleanup into the model itself, Google is lowering the barrier to building voice agents that feel less like rigid command-lines and more like patient collaborators. This isn't just an upgrade for captioning tools. If you've been putting off a voice feature because of transcription jank, now is the time to revisit it.