Google's newest speech model ships with a documented tradeoff most teams will switch on without reading: turning on word level timestamps degrades transcription accuracy. Here is what the specification actually allows, and how to configure it so you do not pay for precision you did not need.
Gemini 3.5 Transcribe landed on August 26 with a 2.6 percent word error rate on prerecorded audio, and the first thing most engineering teams will do is enable every feature in the request. That is the mistake. Google's own documentation states plainly that enabling word level timestamps may degrade overall transcription accuracy. The headline number and the feature you want are, in this model, in tension.
That tradeoff is worth thinking about before it becomes a default in your code. Word level timestamps are genuinely useful for a small number of jobs: aligning captions to video, letting a reviewer click a phrase and jump to that moment in the recording, or building a search index that lands the user on the exact second. They are useless for the majority of transcription work, which is summarization, note taking, ticket creation and anything that feeds text to another model. If your pipeline consumes the transcript as prose, timestamps buy you nothing and cost you accuracy on every request.
The rest of the specification has similar sharp edges. Speaker diarization is supported for up to eight speakers on audio files, and attribution for three or more speakers is marked experimental, which means a four person meeting is not a supported configuration so much as a hopeful one. Diarization is not supported at all in live streaming, so a real time pipeline that needs to know who is talking has to solve that somewhere else. Custom vocabulary biasing accepts up to one thousand terms, which is a real budget and worth spending deliberately on product names, drug names, ticker symbols and internal acronyms rather than filling it with general vocabulary the model already knows.
Two length limits shape the architecture. The file endpoint, gemini-3.5-transcribe, handles up to one hour per request. The streaming endpoint, gemini-3.5-transcribe-live, runs up to ten minutes per session. A ninety minute board meeting needs chunking on the file path and session rotation on the streaming path, and the seam between chunks is exactly where speaker labels and custom vocabulary context get lost. Plan the seam before you build the loop.
One more default deserves a decision rather than an assumption. Smart dictation includes filler word removal and intent aware alphanumeric formatting, which means the model is editing the speaker's words on the way to your database. For a meeting recap that is an improvement. For anything that could become evidence, a record of what was said should be what was said. Decide which of your pipelines needs a verbatim transcript, and make that an explicit configuration rather than something you discover during a dispute.
Try it today
- Audit every place your code requests word level timestamps and turn it off wherever the transcript is consumed as prose rather than aligned to media. Google documents that the feature may degrade overall accuracy.
- Decide diarization per pipeline, not globally. Up to eight speakers is supported on audio files with three or more marked experimental, and diarization is not supported in live streaming at all.
- Spend the custom vocabulary budget deliberately. You get up to one thousand terms, so load product names, internal acronyms, drug names and ticker symbols rather than ordinary words.
- Design for the length limits before load testing. Files are capped at one hour per request and live sessions at ten minutes, so build chunking and session rotation and test that speaker labels survive the seam.
- Classify which pipelines need verbatim output. Smart dictation removes filler words and reformats alphanumerics by default, which is wrong for legal, clinical and compliance records.
- Rerun your accuracy benchmark with your real production settings. The published 2.6 percent figure is a clean configuration number and will not match a request with timestamps and diarization enabled.