Izood MAG

AI — explainer

Gemini 3.5 Transcribe: Google's newest speech-to-text model explained

5 min read
Three photographed people at different scales work around a large flat black-and-white photograph of an open laptop that lies flat on the pale ground.

Gemini 3.5 Transcribe: Google's most precise speech-to-text model, with an API that splits cleanly between live and recorded audio — and one macOS-only feature worth knowing about before you integrate.

What is Gemini 3.5 Transcribe?

Gemini 3.5 Transcribe is Google's most precise speech-to-text model yet. It's built on Gemini's audio understanding capabilities, not a separate architecture bolted onto the side.

What it does is straightforward: it converts raw audio directly into accurate, polished, formatted text. Not a rough first draft you then spend an hour cleaning up. The formatting is part of the output.

Skip this if you're happy with whatever transcriber you already use. This matters when accuracy on messy audio — overlapping speakers, accents, background noise — is the thing standing between you and shipping.

Key features: smart transcription, function calling, and custom vocabulary

Gemini 3.5 Transcribe handles self-corrections — "let's meet Tuesday—no, Wednesday" comes out as Wednesday. It strips filler words, those "ums" and "ahs" that pollute raw transcripts. And it auto-formats text into readable paragraphs instead of dumping a wall of words. That's three cleanup passes a conventional ASR pipeline needs separate tools for.

The function calling piece changes what a transcript can do. Transcribe can delegate complex tasks to other Gemini models via function calls. You're not just getting text back — you're getting an agent that can act on what it heard. A meeting transcript could trigger a calendar event, a summary, a follow-up email, without leaving the model.

One catch: function calling is currently available in the Gemini macOS app. If you're on Windows or Linux, or building against the API directly, this feature isn't in your build yet. The smart transcription and formatting are the parts you can actually use today.

Accuracy and speed benchmarks

Gemini 3.5 Transcribe hits a 4.0% average Word Error Rate streaming and 2.6% non-streaming, as measured by Artificial Analysis. The FLEURS benchmark lands at 5.50% WER streaming and 5.04% non-streaming. Different tests, different numbers — that's normal. WER varies with audio quality, accent, and background noise.

Speed is the bigger story. Time to final transcription drops 70% compared to Chirp 3. That's the figure that changes what you can build. A captioning pipeline that waited seconds now waits under one.

Skip the FLEURS numbers if you're testing on clean, studio-grade audio — the average WER is the one that tracks your reality. I'd still run your own eval set before trusting any published benchmark.

APIs: Live streaming vs. Interactions

Gemini 3.5 Transcribe is available via two separate APIs. Use the Live API for real-time streaming. Use the Interactions API for pre-recorded audio. They are not interchangeable — the features you get depend on which one you call.

The Live API delivers continuous, bidirectional streaming with sub-second latency. That's what you want for interactive voice apps, and you access it through the model identifier gemini-3.5-transcribe-live. Word-level timestamps are not supported in live streaming. If your app needs to highlight each word as it's spoken, live mode won't give you that.

The Interactions API handles recorded audio — meetings, call logs, anything that's already on disk. Call it with gemini-3.5-transcribe and you get two things the Live API doesn't offer: speaker attribution and word-level timestamps. Those matter for post-meeting search and transcript navigation.

If you're building anything interactive, start with Live. The Interaction API's timestamps are useful, but they're for batch processing — not for something a user is talking to right now.

Language support and multi-speaker identification

Gemini 3.5 Transcribe automatically detects and transcribes over 85 languages. You don't specify a language code — the model figures it out from the audio. That's convenient until you feed it code-switched audio, which is where I've seen it hesitate.

For speaker attribution, the production feature handles up to eight speakers in audio file processing. That's the ceiling for the file API as documented. Blog coverage narrows it further: the model accurately attributes speech with timestamps for up to three speakers. Anything above three — five people in a podcast, a full meeting room — is marked experimental. Don't build a production meeting-notes product on the experimental tier and expect stable names. I got this wrong the first time: I assumed eight was the general limit, not the file-processing ceiling with a three-speaker reliability threshold underneath it.

API limits and custom vocabulary details

You get up to 1 hour of audio per request on Gemini 3.5 Transcribe — but that ceiling drops the moment you turn on anything that needs finer-grained analysis. Enable speaker diarization or word-level timestamps and the limit is 30 minutes per request. If your pipeline needs either feature, plan your chunking around the smaller number, not the headline figure.

Custom vocabulary biasing accepts up to 1,000 terms, which sits somewhere between generous and a trap. The docs report that customers typically see best results with up to 100 terms. I'd treat 100 as the working limit and the extra 900 as headroom for experiments, not something to fill on day one.

Skip this if your audio averages under five minutes and you don't need per-speaker labels — none of these constraints will touch you. This reflects the API as of the August 2026 update.

Availability and integrations

Skip this if you want a waitlist drama or a private beta you can't touch. Gemini 3.5 Transcribe is in public preview in the Gemini API right now, via Google AI Studio and Google Antigravity. That's the developer route.

Enterprises get it through the Gemini Enterprise Agent Platform, with Gemini Enterprise for Customer Experience listed as coming soon. Everyone else can try it today in the Gemini app on macOS in English, or on Rambler for Android in select countries and languages. Chrome is next — talk to type in any web field is on the roadmap.

The integrations are where this gets interesting for builders. Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel, and Vision Agents are all using the Gemini Live API to help developers ship voice-driven interfaces. In Google AI Studio, Build mode lets you vibe code apps with your voice — spoken prompts compile into working projects. On Antigravity, the model pairs screen context and chat history, with permission, to pin transcription to what you're actually looking at.

Early adopters are real companies, not demos. vivo, Intellitek Health, and Lingopal have shared positive feedback. Gboard on Android, through Rambler, turns spoken thoughts into formatted text, filtering filler words.

The macOS app is the fastest way to test whether this model fits your workflow — no API key, no setup, just talk. The catch is language coverage. English on desktop, select countries on Android. If you need French or Japanese in a browser today, you're waiting.