Google launched Gemini 3.5 Transcribe on August 26, 2026. It is in public preview in the Gemini API: one model for live streaming, one for recorded audio. You can call it from AI Studio today. That is not the same thing as shipping a voice product your staff will trust on Monday.
This post is for one search: gemini 3.5 transcribe. The only product link is Build Your App. If you have a call queue, a field notebook, or a meeting pile that still gets typed by a person, that form is how we scope the workflow around the new model.
What actually shipped on August 26
Google’s Gemini Audio team posted the launch. There are two entry points, and they are not interchangeable.
- Live streaming:
gemini-3.5-transcribe-liveon the Live API. Continuous, bidirectional, built for voice agents and captions that cannot wait for the file to finish. - Recorded audio:
gemini-3.5-transcribeon the Interactions API. Meetings, call logs, and uploads. Speaker attribution and word-level timestamps. Official support is three speakers; more than that is still experimental.
The model is meant to clean speech, not just decode it. It handles self-corrections (“Tuesday—no, Wednesday”), drops filler words, formats the text, and accepts a custom vocabulary so SKUs, site names, and order codes survive. Google says it covers more than 85 languages and locales. Independent numbers from Artificial Analysis: average word error rate of 4.0% streaming and 2.6% on files, and time-to-final about 70% faster than Chirp 3. On FLEURS, Google published 5.50% streaming and 5.04% non-streaming against Chirp 3.
Availability is split. Developers get a public preview in the Gemini API through AI Studio and Antigravity. Enterprises get the same preview on Gemini Enterprise Agent Platform. Consumers already see pieces of it: Rambler dictation on Android in some countries, the Gemini app on macOS in English, Chrome later. None of those consumer surfaces are your product. They are proof the model exists.
A paste-the-audio demo is not the job
The demo writes a paragraph. Your operation needs a record. Those are different objects.
A paragraph has no case id, no speaker that matches a CRM contact, no retention rule, and no person who owns the 3% that is wrong. A record has all of that. If you ship the paragraph into Slack and call it done, someone will still re-type the same call into the system of record. You paid for a model and kept the labor.
That is the named problem this week: audio already exists, the system of record still waits for a human to type it. Support calls. Site walkthroughs. Shift handoffs. Sales recaps. The new model is good enough that the bottleneck is no longer “can we transcribe.” The bottleneck is “can we land the transcript where work happens, with a name on it, and a way to fix it.”
What a production voice workflow actually includes
Start from the file or the stream, not from the model card.
Retention you can explain. How long the audio sits, who can open it, and whether the transcript inherits the same policy. If that is unclear, do not turn the model on. “We will figure out storage later” is how files end up in the wrong bucket.
A speaker map that is not Speaker 1. Three-speaker attribution is useful. It is not a roster. You still need a way to attach “agent,” “customer,” or “site supervisor” to the people your system already knows. We usually do that with the call metadata you already have (agent id, ticket id), not with a miracle diarization pass.
Your words, not Google’s. Custom vocabulary is the feature that decides whether the model is usable in your vertical. Load the SKU list, the site names, the product codes. Then keep a living eval set of ten real calls and score those terms every time you change a prompt or a model id. If “gemini-3.5-transcribe” becomes “gemini-3.6-something” in October, you want a number, not a vibe.
A queue, not a button. Live audio dies on a laptop refresh. Files die on a 413. Production means retries, a dead-letter pile, and a max runtime you can explain to finance. Streaming for the agent in the ear. Batch for the archive. Same product, two paths, because Google split the APIs that way on purpose.
A review screen for the misses. A 2.6% word error rate on a clean benchmark is not zero errors on a noisy jobsite or a two-year-old conference mic. Someone has to correct the account number. That UI is the product. Hide it and your staff will stop trusting the feed in a week.
A place it writes. CRM note, ticket comment, work-order description, or a search index. If the transcript only lives in the model vendor’s playground, you built a tour.
What we would actually scope this week
A first slice we will take: one audio source, one destination system, one review role. Example: inbound support calls land in storage, the batch model writes a transcript plus speaker turns, a supervisor fixes the account number, and the ticket gets the cleaned note. No voice agent. No “talk to your whole stack.” One queue that already hurts.
We pick the API from the job. Live captions and an interruptible agent use gemini-3.5-transcribe-live. Yesterday’s calls and the weekly meeting dump use gemini-3.5-transcribe. Mixing them in one weekend spike is how you get two half-working UIs.
We will tell you if the audio is too dirty, if policy is missing, or if the destination system cannot accept a write. That answer is cheaper than a prototype that never leaves a staging bucket.
This is not a remake of a generic AI services pitch. The model, the date, and the APIs are specific. The page you use next is specific too.
How to start on Build Your App
Open the Build Your App form. In the solution write-up, name the audio source, the system the note has to hit, and who reviews the misses. On the platform step, choose AI/ML Solution. When you get to category, pick Voice Recognition & Speech AI. Budget and timeline can be a range. We would rather see the workflow than a stack wishlist.
Bring one hour of real audio you are allowed to share, or a written description of the call types if you cannot. We reply with a yes, a not-yet, or a smaller first slice. If the honest work is “set retention and stop emailing wavs” for 30 days, that is the deliverable. The model will still be there in September.
If you only wanted a playground wrapper, you do not need us. If you want the transcript to become a record your team already trusts, start on Build Your App today.