Google Launches Gemini 3.5 Transcribe: A Closer Look at Its New Audio AI Lineup
Google introduced Gemini 3.5 Transcribe on August 26, 2026, calling it its most precise speech-to-text model yet and a completely new addition to the Gemini family. The model takes raw audio and converts it directly into formatted, polished text. It supports more than 85 languages and includes specialized jargon detection along with a customizable vocabulary that adapts to unique spelling needs. Google said the model delivers a major improvement over Chirp 3, its previous transcription model, especially in multilingual performance and word error rates. Google AI Blog
Beyond straight transcription, Gemini 3.5 Transcribe lets users edit text using their voice. It can automatically format output and strip out filler words such as "um" and "uh." For audio with multiple speakers, the model can tell apart up to three voices in pre-recorded files and provides word-level timestamps. It is already powering several of Google's own products. 9to5Google
Alongside Transcribe, Google updated the broader Gemini Audio lineup with two live models: Gemini 3.5 Live and Gemini 3.5 Live Experimental. Both build on the speech recognition technology behind Gemini's voice chat mode. Gemini 3.5 Live improves how the model handles mid-sentence interruptions, automatic language detection, and live visual processing. Gemini 3.5 Live Experimental takes a different approach to complex tasks: it narrates its reasoning step by step in real time, letting users hear the model's intermediate thinking before it produces a final answer. The Verge
The rollout is staged. The Gemini Audio updates are available now in English for all macOS Gemini app users. On Android, the Rambler dictation feature is live in select countries and languages. Chrome support is coming soon, and Google said Gemini 3.5 Transcribe will eventually let users apply speech-to-text in any web field within the browser. For developers, all three new models are available in public preview through the Gemini API via AI Studio and Antigravity. Engadget
The broader context here is that Google has been steadily building an audio model stack across both its proprietary Gemini line and its open Gemma family. Gemini 3.5 Live Translate, announced in June 2026, delivers near real-time speech-to-speech translation across more than 70 languages. On the open-source side, Gemma 3n introduced an audio encoder for on-device speech-to-text on clips up to 30 seconds, and the more recent Gemma 4 12B is designed to bring agentic, multimodal intelligence to the laptop, including fully local transcription. Gemini 3.5 Transcribe sits at the top of that stack as the cloud-side, full-featured model.
What sets Gemini 3.5 Transcribe apart from conventional speech recognition is not only accuracy but the post-processing layer. Automatic filler-word removal, voice-driven editing, and formatting all happen inside the model itself rather than requiring a separate language-processing step or user-scripted cleanup. For developers building transcription into applications, that consolidation reduces integration complexity. The customizable vocabulary feature matters for domains where generic speech recognition consistently fails: medical terminology, legal Latin, product code names, and other specialized words that typically force downstream correction workflows.
The speaker attribution and word-level timestamp capabilities position the model for use cases beyond dictation. Pre-recorded audio with up to three speakers covers a meaningful slice of meeting, interview, and podcast transcription, though it stops short of the multi-speaker diarization (the task of separating and labeling many speakers throughout a long recording) that enterprise call-center and media-production workloads often require.
Worth flagging is the gap between the current feature set and what Google has said is coming. Speech-to-text in any Chrome web field would put Gemini 3.5 Transcribe directly into the browser's input layer, extending the model's reach from app-embedded use to the broader web. Google has not provided a date for that capability or for broader Chrome support. The staged rollout, currently English-only and macOS-first on the consumer side, suggests Google is validating the model's real-world performance before widening access.
For developers evaluating the API, the public preview in AI Studio and Antigravity is the practical entry point. The model is already production-deployed in Google's own products, which provides some signal about its readiness, though first-party usage conditions and third-party API performance can diverge in areas like latency under load and rate limits.
