Gemini 3.5 Transcribe: Google's Most Precise Speech AI Yet

Gemini 3.5 Transcribe is the most precise speech-to-text model Google has shipped, by its own benchmarks and by independent testing, and it entered public preview on August 26, 2026. The model ships as two separate APIs: a real-time streaming option for live voice apps and a pre-recorded option that adds speaker labels and word-level timestamps for podcasts, meetings, and interviews.

Google built the new model around a smart transcription mode that catches mid-sentence corrections, drops filler words, and formats the output automatically, on top of the traditional word-for-word transcript most speech tools default to. Independent testing from Artificial Analysis puts word error rate at 2.6% for pre-recorded audio and 4.0% for live streaming, alongside a 70% cut in time-to-final-transcript compared with Chirp 3, the model it replaces. The rollout already reaches Gboard, Google Antigravity, AI Studio, and the Gemini app on macOS.

Real-Time or Recorded: Inside the Gemini Transcribe API

Gemini 3.5 Transcribe splits into two APIs built for different jobs. The Live API, gemini-3.5-transcribe-live, streams audio continuously in both directions and returns text with sub-second latency, for voice agents and other apps that can’t wait for a pause in speech. The Interactions API, gemini-3.5-transcribe, processes pre-recorded audio and adds speaker attribution with word-level timestamps, the format podcast editors and meeting notes need.

Both APIs share the same behavior around messy real speech. The default mode, verbatim, transcribes exactly what was said, filler words included. A separate smart mode resolves the kind of self-correction that shows up in ordinary conversation, like someone saying “let’s meet Tuesday, no, Wednesday,” strips out “ums” and “ahs,” and formats the result instead of leaving a flat wall of text.

Custom vocabulary is supported too. Feed it jargon, product names, or unusual spellings ahead of time and it adapts instead of guessing. The Interactions API can also identify separate speakers and timestamp each one.

How Accurate Is Gemini 3.5 Transcribe?

Independent testing backs up Google’s accuracy claim. Artificial Analysis measured a 2.6% word error rate on pre-recorded audio through the Interactions API and 4.0% on live streaming through the Live API, with the model holding up well in noisy environments and correctly capturing alphanumeric strings like postal codes and order IDs.

On the FLEURS multilingual benchmark, which tests accuracy across a wide language set, the model scored 5.04% word error rate for non-streaming audio and 5.50% for streaming.

Chirp 3, Google’s prior transcription model, is the baseline for that 70% speed figure. Google says the newer model brings new capabilities on top of a lower word error rate, but the clearest practical difference is speed: Artificial Analysis clocked a 70% improvement in time-to-final-transcript against Chirp 3, which matters most in the Live API’s real-time use cases, where lag is felt immediately.

Where Can You Use Gemini 3.5 Transcribe Today?

Availability splits by audience. Developers can test Gemini 3.5 Transcribe now through the Gemini API in Google AI Studio and inside Google Antigravity. Enterprises get it in public preview through the Gemini Enterprise Agent Platform, with Gemini Enterprise for Customer Experience listed as coming soon. Consumers already have it running inside four products:

Chrome support is coming soon, for talking instead of typing into any web form or text field. On macOS, the model can also hand a task to a different Gemini model through function calling: asking it to generate an image or analyze a file mid-conversation routes the request automatically, with Gemini 3.5 Transcribe acting as the interface rather than doing that work itself.

For a working creator, the practical uses line up with tools already in daily rotation: turning a raw podcast recording into a searchable transcript, drafting captions before a video export, dictating a rough script instead of typing one, or feeding a clean transcript into one of the best AI text-to-speech tools to produce a dubbed voiceover in another language. Developer platforms including Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel, and Vision Agents already support it, and early users like Vivo, Intellitek Health, and Lingopal cited latency, accuracy, and language coverage in feedback Google published at launch.

Google’s Bet on Transcription as Infrastructure

Google didn’t give the August 26 launch a standalone product page or a price list. It shipped the model as a layer inside things people already use: a keyboard, a coding tool, an operating system’s built-in assistant. That’s a different pitch than the one dedicated voice AI platforms have been making.

ElevenLabs built its business the other way, as a platform where transcription is one feature among several rather than a background utility folded into other apps. Its own speech-to-text model, Scribe, sits inside a workspace that also handles text-to-speech, voice cloning, and dubbing, aimed at people who want one tool built specifically for audio production. A developer wiring a voice agent together needs something different than a video editor who just wants a finished interface, not an API to learn.

What Gemini 3.5 Transcribe changes is the baseline. When a model this accurate ships free inside a keyboard app, decent transcription stops being something creators pay a separate subscription for by default, and platforms competing for that budget need a clearer reason than accuracy alone. Google has moved fast on Gemini features this quarter, from a watermark toggle for AI-generated media to this transcription launch. Same pattern both times: capability shipped as infrastructure, not sold as a product on its own. It’s worth watching alongside our AI news feed.

FAQ

---

Sources

1. Google: Gemini 3.5 Transcribe - Google, August 26, 2026
2. 9to5Google: Gemini 3.5 Transcribe - 9to5Google, August 26, 2026
3. Business Today: Google launches Gemini 3.5 Transcribe with smarter speech-to-text and 85 language support - Business Today, August 27, 2026
4. Google AI for Developers: Transcribe audio - Google AI Documentation