Gemini 3.5 Transcribe brings emotion detection and speaker ID to speech-to-text

Google has quietly raised the bar for AI-powered transcription. Unveiled at Google I/O on May 19, 2026, and detailed further with a dedicated model card published on August 26, 2026, Gemini 3.5 Transcribe is the company’s most capable audio processing model to date, and it does considerably more than convert spoken words into text.

The model supports timestamps in MM:SS format, speaker identification, translation, summarization, and emotion detection.

What the model actually does

At 96,000 tokens, Gemini 3.5 Transcribe can process extended audio sessions without losing track of what was said earlier in the recording.

Users can upload common audio file formats, including MP3, directly through the Gemini API, Google AI Studio, or the Gemini macOS application. From there, the model can clean up filler words, identify individual speakers, translate content, or produce a summary, depending on what the user requests.

Google does draw a line on one use case. For real-time transcription, the company points users toward its Cloud Speech-to-Text API and a separate Live API, rather than routing live audio through the general Gemini model. That distinction matters for developers building applications where latency is a constraint.

The macOS Gemini app began receiving the voice dictation features around mid-2026. Rather than producing a raw transcript, the app is designed to output polished text directly from natural speech.

Why this is a significant product moment for Google

Emotion detection is a capability that differentiates this release. Most transcription products focus on accuracy at the word level. Layering in emotional context changes the nature of the output entirely. A call center summary that notes a customer sounded frustrated at minute four is a different product from one that simply logs what was said.

The 96K token context window is meaningful in practical terms. A standard business meeting runs roughly 60 minutes. A legal deposition or a lengthy earnings call can run considerably longer. Supporting extended audio without requiring the user to split files into chunks removes friction that previously sent users toward more specialized, purpose-built tools.

Who this affects and what to watch

The most immediate beneficiaries are professionals whose work generates significant volumes of spoken content: journalists, lawyers, academics, medical practitioners, and enterprise teams running frequent recorded meetings.

The competitive implications are real for companies like OpenAI, which offers Whisper and related transcription tools, as well as for dedicated transcription platforms that have built businesses on capabilities Gemini 3.5 Transcribe now packages into a general-purpose API.

Developers building on the Gemini API should note the clear guidance on real-time versus batch use cases. Google’s decision to route live audio toward dedicated infrastructure rather than the general model suggests the company is managing capacity and latency expectations carefully, which is worth factoring into any integration planning.

The model card published on August 26, 2026 gives technical users a structured reference point for understanding capability boundaries.