Google's new Gemini 3.5 Transcribe raises the bar for voice AI, and the risks that come with it
Google DeepMind's latest speech-to-text model is faster, sharper, and handles 85 languages. It also hands scammers a cleaner tool.
Key points
- Google DeepMind launched Gemini 3.5 Transcribe, calling it its most accurate speech-to-text model yet.
- The model scores a 4.0% Word Error Rate in streaming mode and 2.6% for pre-recorded audio, per Artificial Analysis.
- Time to final transcription is 70% faster than Google's previous Chirp 3 model, again per Artificial Analysis.
- The system supports over 85 languages, identifies up to three speakers, and can be tuned to custom vocabulary.
- It is in public preview through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform.
Google DeepMind has released a new speech-to-text engine called Gemini 3.5 Transcribe. Speech-to-text turns spoken audio into written words, the thing your phone does when you dictate a message.
The pitch is simple: it listens more accurately, in more languages, with less delay than its predecessor. For a nurse dictating notes or a small business running a call centre, that sounds like pure upside. My job on the security beat is to tell you where the sharp edges are.
What is actually new here?
Gemini 3.5 Transcribe comes in two versions: a live streaming mode for real-time apps and a batch mode for recorded audio like meetings and call logs.
The numbers matter. On a benchmark run by the independent firm Artificial Analysis, the model hits a 4.0% Word Error Rate in streaming and 2.6% for pre-recorded audio. Word Error Rate is the share of words the system gets wrong, so lower is better. Our August story on AI speech models cheating on their own benchmarks is worth keeping in mind when reading those figures: high scores don't always mean real-world ability.
Speed also jumped. Time to final transcription improved by 70% over Chirp 3, Google's earlier model.
| Feature | Streaming | Pre-recorded |
|---|---|---|
| Word Error Rate (Artificial Analysis) | 4.0% | 2.6% |
| FLEURS multilingual WER | 5.50% | 5.04% |
| Languages supported | 85+ | 85+ |
| Speaker attribution | No | Up to 3 |
The model also does "smart" cleanup. If you say "let's meet Tuesday, no, Wednesday," it writes Wednesday. Filler words vanish. Text is formatted on the fly.
Where does this show up for ordinary people?
In a few places you probably already use. On Android, the Gboard keyboard gets a feature called Rambler that turns rambling speech into tidy text. The Gemini app on macOS gets voice commands that can pull in what's on your screen. Chrome will soon let you talk to type in any web field.
Developers can plug it into their own products through the Gemini API in Google AI Studio, and businesses can access it through the Gemini Enterprise Agent Platform.
Should users be worried?
Only in a specific way. Better transcription is genuinely useful. It also lowers the cost of scams that rely on clean, believable audio.
Voice phishing, sometimes called vishing, already uses AI-generated voices to impersonate a boss or a relative in trouble. A transcription model that handles noisy calls and dozens of languages makes it easier to build fully automated scam call centres that react to what a victim says in real time, in the victim's own language. Automated attack pipelines are exactly what we covered when Ringg expanded its voice AI capabilities last month.
What matters most here is the ecosystem, not the model. Google's guardrails apply inside its own products. Any developer with API access builds their own experience, with their own rules.
A second concern: the macOS version can read your screen to improve accuracy. That's a permission you grant, not a default, so treat it the way you'd treat screen-sharing on a work call.
What readers should watch for
- A caller who sounds smooth and responsive but pushes urgency. Real people pause; automated voice agents rarely do unless trained to.
- Unexpected calls in your own language from services that normally use another. Cheap multilingual AI is why.
- Any app requesting microphone plus screen access together. Grant one, not both, unless you trust the vendor completely.
- Voicemails that sound like a family member in trouble. Hang up, call the person back on a known number.
Credit where it's due: Artificial Analysis ran the benchmarks Google is citing, which is worth knowing when you read the marketing.
