Google released Gemini 3.5 Transcribe on 26 August 2026, a speech-to-text model that its makers say applies reasoning to audio rather than merely converting sound to characters. The model is in public preview in the Gemini API and ships as two identifiers: gemini-3.5-transcribe for pre-recorded files and gemini-3.5-transcribe-live for real-time streaming.

The framing is deliberate. Google describes the model resolving a speaker’s self-corrections, stripping filler words such as “um” and “ah”, formatting the output, attributing up to three speakers, and — through function calling — handing follow-on work such as file analysis to other Gemini models. In Google’s telling, transcription stops being a standalone signal-processing task and becomes one more call on the same multimodal surface developers already use for text and images. Audio is billed in tokens, like any other Gemini input.

Why it matters

For buyers, speech-to-text has long been a separate procurement: a dedicated vendor, a dedicated contract, its own accuracy claims. Folding it onto the Gemini API removes that seam for teams already on Google’s stack, and it changes the comparison from “which transcription engine” to “which model, at which price, on the same bill”. That is the same consolidation logic now visible across Google’s line, and it is the axis on which buyers should judge it — not the headline accuracy number.

It also lands in a market where the raw word-error rate has largely stopped being the differentiator. Google is explicitly competing on what the model does after it has the words: cleanup, speaker labels, and the ability to trigger downstream tasks.

What the accuracy numbers actually measure

Google’s headline claim is a 2.6 percent word-error rate on pre-recorded audio and 4 percent on streaming audio, and it attributes both figures to the third-party evaluator Artificial Analysis, measured across what Google calls “diverse real-world conditions” including background noise. Those are the numbers most coverage has led with. They are not the whole picture.

On FLEURS — the standard multilingual benchmark, run across a set of top languages and locales — Google reports a 5.5 percent word-error rate in streaming mode and 5.04 percent in non-streaming. In other words, the sub-3-percent figure and the ~5-percent figure describe different test sets: the low number reflects a real-world English-leaning evaluation, while the multilingual benchmark, spread across many languages, roughly doubles the error rate. Both are Google’s own reported results, and both should be quoted with their conditions attached. A buyer transcribing a single clean language will see something closer to the low number; a buyer transcribing many languages should plan against the FLEURS figure.

The model supports more than 85 languages and can attribute up to three speakers; Google labels support for more than three speakers experimental. None of these are independently verified accuracy claims — they are the vendor’s own specifications, and the word-error rates are the vendor’s own reported metrics, one set of them measured by a third party the vendor selected.

The distribution is the strategy

Beyond the API, Google says the model is reaching users through Gboard’s Rambler dictation feature, the Gemini app including on macOS, Google AI Studio, Workspace, and Google’s enterprise agent platform. That breadth is the point: the same transcription model powers a consumer keyboard and an enterprise API at once, and each deployment feeds the same model line rather than a separate speech product.

This continues a pattern 18BYTE has tracked in Google’s model organisation, where the recent leadership split put Gemini’s roadmap under a single owner. A transcription model that is an endpoint rather than a product is exactly what that consolidation produces. For readers weighing Google against dedicated speech vendors, the relevant profiles remain the Gemini 3.6 Flash line and the Gemini app, since the transcription model draws on the same family and billing.

What this means for buyers

If you already run on the Gemini API, the migration cost is low: the same authentication, the same token billing, two new model identifiers. Test against your own audio in your own languages before trusting the 2.6 percent figure — that number reflects a curated real-world set, and your mileage on accented, multilingual, or many-speaker audio will sit closer to the FLEURS ~5 percent range, with three-plus-speaker attribution still flagged experimental.

If you buy transcription from a dedicated vendor, the question is no longer accuracy alone but whether the downstream features — filler removal, speaker labels, function-calling into other models — are worth consolidating onto one supplier, against the concentration risk of doing so. Because the model is in public preview, treat both pricing and behaviour as provisional; Google’s own pages did not publish a fixed per-minute rate, and preview terms can change before general availability.

What would change our reading

We would revise upward if independent evaluators, not only Google’s chosen third party, reproduce the sub-3-percent word-error rate on comparable audio, and if Google publishes stable general-availability pricing and rate limits. We would revise downward if the multilingual gap proves wider than FLEURS suggests in production, if speaker attribution degrades on real meetings, or if the preview’s per-token billing turns out materially more expensive per minute than incumbent speech-to-text vendors once list pricing is confirmed.

Sources

  • Google, The Keyword blog — official announcement of Gemini 3.5 Transcribe — blog.google
  • Google DeepMind, Gemini Audio model page — capabilities, benchmarks and availability — deepmind.google