Gemini 3.5 Transcribe Splits Speed From Precision

Google has a new transcription model. Gemini 3.5 Transcribe targets real-time voice interfaces and recorded audio, combining faster final results with tools for cleaning and structuring spoken language.

Google announced Gemini 3.5 Transcribe on August 26, 2026, with availability beginning that day. The model is rolling out in English for macOS Gemini app users and Rambler on Android in select countries and languages, while Google reported the release on August 27.

The headline numbers are strong: Artificial Analysis measured an average word error rate of 4.0% for streaming and 2.6% for non-streaming. On the multilingual FLEURS benchmark, Gemini 3.5 Transcribe reported 5.50% streaming and 5.04% non-streaming.

Google also reports a 70% improvement in time to final transcription compared with Chirp 3, the previous model. That makes the upgrade less about a new decimal point in a benchmark table and more about reducing the wait between speech and usable text.

Two APIs, Two Different Trade-offs

The split between the two endpoints is the part worth planning around. The Live API delivers sub-second, continuous transcription and emits interim_input_transcription and input_transcription, with automatic, hybrid, and manual voice-activity detection available.

Live API sessions cap at 10 minutes of continuous streaming. The endpoint does not support speaker diarization or word-level timestamps, so developers building transcripts that identify speakers or track exact word positions will need the other route.

The Interactions API provides speaker diarization, word-level start and end offsets, and custom vocabulary biasing. Standard requests accept up to one hour of audio, but that limit falls to 30 minutes when diarization or word timestamps are used.

Neither endpoint shares the same feature set, limits, or price. That is not a footnote; it is the product decision. A short live voice interface and a long recorded meeting may both need transcription, but Gemini 3.5 Transcribe does not treat them as the same job.

Cleaner Text, Broader Language Coverage

Gemini 3.5 Transcribe supports automatic detection across more than 85 languages, including code-switching in the middle of a sentence. The coverage gives the model a wider operating range than a system designed around one language per recording.

Its output modes also separate raw capture from polished text. Verbatim mode returns every disfluency, while smart mode removes disfluencies and applies structured formatting.

Smart mode cannot be combined with word timestamps or diarization. Users must choose between cleaner output and the metadata needed for detailed transcript analysis — because apparently transcription still requires picking a lane.

Custom vocabulary support allows up to 1,000 terms, with the best results below 100. That feature gives specialized names and phrases a better chance of surviving transcription, provided the vocabulary list stays focused instead of becoming a small dictionary with delusions of grandeur.

Google says Gemini 3.5 Transcribe allows users to “edit naturally with just your voice,” while automatic formatting and removal of filler words such as “um” and “uh” turn spoken input into more usable text. The model is designed to handle both the speech itself and the untidy habits that arrive with it.

Early Integrations And Access

The Live API is wired into LiveKit, Pipecat, Agora, Fishjam, Vercel, and Vision Agents. Gemini 3.5 Transcribe also powers Rambler on Android, the Gemini app on macOS, and Google Antigravity.

Developers can access the model through the Gemini API using AI Studio and Antigravity. Chrome support is coming soon, while Gemini 3.5 Live and 3.5 Live Experimental updates remain planned but have not launched.

For now, the model’s appeal rests on a clear combination: 2.6% average word error for non-streaming audio, 4.0% for streaming, a 70% faster path to final transcription than Chirp 3, and coverage spanning more than 85 languages. The catch is equally clear — choosing the wrong endpoint can cost access to the exact feature your transcript needs.

Stay connected via Google News

Follow us for the latest AI News