Google Unveils Gemini 3.5 Transcribe — Betting on Precision With a 2.6% Word Error Rate

Google launched Gemini 3.5 Transcribe, its latest speech-to-text model, on Aug. 26. It posts a 2.6% average word error rate for non-streaming and 4.0% for streaming as measured by Artificial Analysis, beating its own Chirp 3, and auto-detects more than 85 languages. It ships as two separate APIs — one for real-time streaming, one for pre-recorded audio — and already powers Gboard Rambler, the Gemini app on macOS, and Antigravity.

Google introduced Gemini 3.5 Transcribe on Aug. 26. As the "most precise speech-to-text model yet," accuracy is the pitch: as measured by third-party firm Artificial Analysis, it hits a 2.6% word error rate for non-streaming and 4.0% for streaming, and it auto-detects more than 85 languages. It goes beyond raw dictation — stripping fillers like "um" and "ah" and cleaning up self-corrections such as "let's meet Tuesday—no, Wednesday." It ships as two separate APIs, one for real-time and one for pre-recorded audio, now open in public preview for developers and enterprises.

The contest over voice interfaces is shifting from "how well does it hear you" to "how cleanly does it write you down." Announcing Gemini 3.5 Transcribe on Aug. 26, Google said that unlike conventional speech recognition that struggles with background noise, jargon, and disfluency, the model "converts raw audio directly into accurate, polished, formatted text." It already powers first-party products including Rambler on Android's Gboard and the Gemini app on macOS.

Why Two Separate APIs

The model comes not as one endpoint but as two distinct APIs, because the jobs differ.

The real-time gemini-3.5-transcribe-live runs continuous, bidirectional streaming with sub-second latency through the Live API — aimed at voice agents and live captioning where instant response matters. The pre-recorded gemini-3.5-transcribe, by contrast, uses the Interactions API to transcribe meetings, call logs, and more with speaker attribution and word-level timestamps — suited to batch work like post-call analytics pipelines.

Non-streaming WER 2.6% (measured by Artificial Analysis)
Streaming WER 4.0%
Languages 85+ auto-detected
Speaker attribution Up to 3 in pre-recorded audio (3+ experimental)
Real-time latency Sub-second

What Changed Versus Chirp 3

Google calls it a "major advancement" over 2025's Chirp 3. By the numbers, time to final transcription improves by 70%. On the multilingual FLEURS benchmark, across a set of top languages and locales, it posts 5.50% WER in streaming and 5.04% non-streaming, beating Chirp 3.

Beyond accuracy, the standout is task delegation. Through function calling, the model can hand complex jobs — image generation, file analysis — to other Gemini models. In the macOS Gemini app, that means summarizing a local file or generating an image at the cursor using only your voice.

Metric	Gemini 3.5 Transcribe	vs. prior (Chirp 3)
Non-streaming WER	2.6% (Artificial Analysis)	Improved
Streaming WER	4.0%	Improved
FLEURS streaming WER	5.50%	Beats Chirp 3
Time to final transcription	—	70% faster
Speaker attribution	Up to 3 (with timestamps)	Newly strengthened

Word Error Rate (WER) is better when lower. Note that Google's 2.6% and 4.0% figures come from third-party Artificial Analysis rather than Google itself — and that on the multilingual FLEURS benchmark the numbers climb into the 5% range. In other words, real-world performance in English-heavy settings may differ from performance across all 85+ languages.

Where You Can Use It

Availability splits across developers, enterprises, and everyone else. Developers can access it in public preview through the Gemini API in Google AI Studio and in Antigravity. Enterprises get preview access via the Gemini Enterprise Agent Platform, with Gemini Enterprise for Customer Experience coming soon. Consumers can already try it in the Gemini app on macOS (English) and Rambler on Android, with a Chrome feature — talk to type in any web field — arriving next.

The easiest way to try it is Rambler on Android's Gboard or the Gemini app on macOS. Developers should call the two model names distinctly in Google AI Studio: `gemini-3.5-transcribe` for pre-recorded audio and `gemini-3.5-transcribe-live` for real-time. Voice infrastructure platforms including LiveKit, Pipecat, LangChain, and Vercel already support the Live API integration.

Google launched Gemini 3.5 Transcribe, its latest speech-to-text model, on Aug. 26
2.6% non-streaming / 4.0% streaming WER (Artificial Analysis); auto-detects 85+ languages
Split into real-time (`-live`) and pre-recorded APIs; pre-recorded adds up to 3-speaker attribution with timestamps
70% faster time to final transcription than Chirp 3; function calling delegates work to other Gemini models
Public preview for developers and enterprises; ships in Gboard Rambler, macOS app, and Antigravity, with Chrome coming