Google Launches Gemini 3.5 Transcribe with Smart Disfluency Cleanup

On August 26, 2026, Google introduced Gemini 3.5 Transcribe, describing it as its most precise speech-to-text model yet for intelligent voice interactions. [4] Unlike conventional speech recognition models that struggle with background noise, complex jargon, and disfluency cleanup, Gemini 3.5 Transcribe converts raw audio directly into accurate, polished, formatted text. [4]

The smart transcription feature handles self-corrections, auto-formats text, and edits out filler words like 'ums' and 'ahs'. [1][4] Furthermore, the model can delegate complex tasks, such as image generation and file analysis, to other Gemini models via function calls. [4]

The model is currently in public preview for developers and enterprises. [3] Developers can build voice agents, real-time captioning tools, or post-call analytics pipelines with Gemini 3.5 Transcribe in the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform. [4]

Google made the model available across two separate APIs to suit different developer workflows. [4] The real-time streaming API utilizes `gemini-3.5-transcribe-live` to deliver continuous, bidirectional streaming with sub-second latency for interactive voice apps. [4] Meanwhile, the Interactions API uses `gemini-3.5-transcribe` for pre-recorded audio processing, transcribing meetings and call logs with speaker attribution and word-level timestamps. [4]

Consumers are already benefiting from this transcription model with new voice capabilities across products like the Gemini app and on Android. [4] Specifically, the speech-to-text model powers Gboard Rambler and is slated to come to Chrome. [3]

The new voice capabilities are also active via Rambler on Android and in the Gemini app on macOS. [4] According to Google, the function calling capability that delegates tasks to other models is currently available in the Gemini macOS app. [4] The system is designed to capture a user's natural speaking style to better understand intent and recognize custom vocabulary, enabling users to execute tasks with their voice. [4]

By integrating disfluency cleanup and function calling natively into the transcription model, Google is moving beyond conventional speech recognition systems that struggle with jargon and background noise. Packaging these features into a streaming API for sub-second latency suggests a strong push to enable interactive voice agents. The rollout spans multiple consumer entry points, including Gboard Rambler, Android, macOS, and Chrome, establishing a wide deployment footprint for the new architecture. What the sources don't address: How the operational costs of running these advanced transcription models will impact pricing tiers for developers using the API.