Turn Audio into Action with Gemini 3.5 Transcribe

Summary

Katie Nguyen, Developer Relations Engineer, introduces Gemini 3.5 Transcribe as Google's next-generation speech transcription model delivered through the Agent Platform on Google Cloud. Organizations generate large volumes of customer calls, meetings, voice notes, and dictations that contain vertical-specific vocabulary. Gemini 3.5 Transcribe is designed to handle this variety by delivering transcription-grade accuracy, deep contextual comprehension, and specialized speech capabilities across multiple languages without requiring users to be speech recognition experts.

The system operates using two distinct model identifiers tailored to different workflows. Gemini-3.5-transcribe handles pre-existing recordings, such as meeting audio or call logs stored in Google Cloud Storage buckets. Gemini-3.5-transcribe-live handles streaming audio in real time for live captions and hands-free dictation. Both versions support automatic language detection, speaker diarization, word-level timestamps, and custom vocabulary lists.

The first demonstration models a clinical healthcare scenario where a physician dictates surgical and diagnostic notes hands-free. Medical terminology such as Nissen fundoplication and pyrosis can be difficult for general-purpose transcription systems. By adding these specialized phrases into the custom vocabulary configuration before starting the stream, Gemini 3.5 Transcribe Live captures and spells the terms accurately in real time. A comparative test using the same spoken audio without the custom vocabulary inputs shows the model misinterpreting the surgical procedure as Nissan fund application.

The second demonstration addresses enterprise call center logs by transcribing a recorded support conversation marked by background noise. Using batch processing with speaker diarization activated, the model separates the dialogue between Speaker 1 and Speaker 2, correctly transcribing account numbers and monetary amounts. The generated transcript is subsequently routed to Gemini 3.7 Flash to perform secondary analysis, producing a numerical sentiment score, highlighting specific causes of customer frustration, and drafting a structured summary of how the billing issue was resolved.