Google ships Gemini 3.5 Transcribe with batch and live speech-to-text models

Google's gemini-3.5-transcribe and gemini-3.5-transcribe-live are in public preview; Google relays an Artificial Analysis WER of 2.6% for the batch model, lists 85+ languages, and its pricing page pegs batch transcription at roughly half a cent per minute.

Key points

• Two model IDs: gemini-3.5-transcribe takes pre-recorded files through the Interactions API and returns speaker labels plus word-level timestamps; gemini-3.5-transcribe-live handles bidirectional streaming over the Live API with sub-second latency. Both are public preview.
• Figures Google credits to Artificial Analysis: 2.6% average WER non-streaming, 4.0% streaming, and a 70% improvement in time to final transcript over Chirp 3. Google also reports FLEURS multilingual WER on a selection of major languages, 5.04% batch and 5.50% streaming, without naming who measured it.
• Pricing is on the Gemini API pricing page, not in the launch post: batch costs $2 per million audio-input tokens and $12 per million output tokens (thinking tokens included), which Google estimates at about $0.005/min; live is $3.50 in and $21 out, about $0.009/min. Both models have a free tier, and the page notes free-tier traffic can be used to improve Google's products.
• Auto-detects 85+ languages and copes with a speaker changing language mid-stream; diarizes up to three speakers in pre-recorded audio, with more than three flagged experimental; strips filler words and self-corrections, formats the output, and accepts a custom vocabulary for jargon.
• Developers get it as a public preview through the Gemini API in AI Studio and inside Google Antigravity, plus the Gemini Enterprise Agent Platform; Gemini Enterprise for Customer Experience is listed as coming soon. Consumer surfaces already running it are the Gemini macOS app (English) and Gboard's Rambler on Android, with Chrome talk-to-type coming later. It is a hosted API model; the post says nothing about weights.
• Voice-stack partners named at launch are LiveKit, Pipecat, Vercel, LangChain, Agora, Fishjam and Vision Agents, all building on the Gemini Live API.

“our most precise speech-to-text model yet.” — Google's launch post, by Diego Melendo Casado and Luke Leonhard

Why it matters

For a builder adding voice to an app, the batch model's roughly half-cent-per-minute estimate plus a free tier makes it worth pricing against a self-hosted Whisper box, and the live model slots into LiveKit or Pipecat pipelines. The caveats: it is preview-only and hosted-only, the headline 2.6% is an average Google is passing along from Artificial Analysis, and the FLEURS numbers carry no outside attribution. Run it on your own accented, noisy or multi-speaker audio before swapping pipelines, and remember the per-minute figures are token-based estimates from Google, so check real invoices on long recordings.

This is an AI-written summary of the reporting credited above and the other sources linked in the text, read and edited by Nicholas before publishing. The facts and any quote belong to those sources; the wording is ours. Read the original.