Gemini 3.5 Transcribe: 2.6% WER and your call logs

Google shipped Gemini 3.5 Transcribe into public preview on August 26, claiming a 2.6% word error rate on pre-recorded audio and 4.0% on live streaming. If those numbers hold on your audio, the pile of customer calls you've never listened to just became a queryable dataset.

What actually happened

Per Google's announcement, there are two models: gemini-3.5-transcribe for recorded audio through the Interactions API, and gemini-3.5-transcribe-live for real-time streaming through the Live API. Google reports 2.6% WER non-streaming and 4.0% streaming, with a 70% improvement in time-to-final-transcription over its previous Chirp 3 model. On the multilingual FLEURS benchmark the numbers are 5.04% and 5.50%.

The feature list is where the practical value sits. It handles self-corrections and strips filler words, so you get formatted text instead of a raw dump of "um" and false starts. It supports over 85 languages with automatic detection, multi-speaker identification, custom vocabulary — which is how you teach it your SKUs and product names — and function calling. It already powers Gboard's Rambler on Android, voice control in the Gemini macOS app, and the mic in Google Antigravity, with Chrome integration announced as coming.

Developer access is through Google AI Studio and the Gemini API in public preview; enterprises get it via the Gemini Enterprise Agent Platform. 9to5Google confirms the same launch details. Google did not publish per-minute pricing in the announcement, so budget it from the API pricing page rather than from a number you read in a roundup.

Why accurate speech-to-text matters for your business

The old math on call transcription was bad. At 8-10% WER you spent more time correcting transcripts than you saved, so the honest answer was to skip it. Under 3%, the calculus flips: transcripts become an input other systems can read.

Concretely, for a shop doing 200 calls a month. Transcribe every call, run each transcript through a summarizer, and write the summary plus a next-action into the CRM record automatically. Nobody types notes. Then query the whole corpus monthly for the questions you keep answering by hand — those are your FAQ page, your quoting form, and your product description gaps, ranked by how often they cost you a phone call.

Two cautions from doing this work. Custom vocabulary is not optional — benchmark WER is measured on clean general speech, and your industry terms and brand names will drag real-world accuracy down until you feed them in. And check your two-party consent rules before you record anything; that's a state-law question, not an API question, and it decides your call-open script.

Public preview means no stability guarantee. Pin the model string, keep the audio files, and be able to re-run the pipeline when the API changes under you.

Sources: Google Blog, 9to5Google.