Google Launches Gemini 3.5 Transcribe in Preview
Google’s new speech model handles live multilingual audio, cleans disfluencies and connects dictation directly to tool-using workflows.
Two models for live and recorded speech
Google introduced Gemini 3.5 Transcribe, a speech-to-text model offered in separate preview variants for real-time streaming and prerecorded or batch audio. It is available through the Gemini API and Google AI Studio, as well as products and platforms including Gemini for macOS, Gboard, Workspace and Google’s enterprise agent environment.
The batch model supports more than 85 languages, a 96,000-token input window and speaker attribution with timestamps for up to three speakers. Google reports word error rates of 5.04% on a selected group of FLEURS languages in non-streaming evaluation and 5.50% in streaming mode. Those results are Google’s own measurements and do not establish performance across every accent, acoustic setting or domain vocabulary.
Transcription becomes an interface
The model is designed to produce usable prose rather than a strictly literal transcript. It can remove filler words, add punctuation and formatting, interpret spoken corrections, and improve recognition of phone numbers, postal codes and order identifiers. It also detects language changes during a conversation.
More consequentially, the streaming version supports function calling. Spoken requests can be routed to other Gemini capabilities for web search, image generation, summarization or file analysis. That turns transcription into an input layer for agents: a user can dictate, revise the wording verbally and initiate a task without switching interfaces. The same flexibility raises accuracy and governance questions because removing disfluencies or interpreting intent can alter the evidentiary value of a transcript.
Why it matters
Speech recognition is becoming infrastructure for agentic systems rather than a standalone conversion utility. Google can distribute this model across consumer keyboards, desktop software, enterprise tools and a developer API, giving it unusually broad reach at launch. Its handling of alphanumeric strings addresses a persistent weakness in voice interfaces, while function calling connects speech directly to actions. Buyers will still need to distinguish polished dictation from faithful transcription—especially in medical, legal and compliance settings where an omitted hesitation or normalized phrase may carry meaning.