Intelligent transcription with Gemini 3.5 Transcribe

2026-08-27 · Google DeepMind

Gemini 3.5 Transcribe is now introduced

Google has announced Gemini 3.5 Transcribe, describing it as its most accurate speech-to-text model so far, built for intelligent voice interactions. The focus is not only recognition, but turning raw speech into clean, polished, and formatted text.

According to Google, the model is designed to handle common pain points in conventional ASR systems, including background noise, jargon-heavy speech, and disfluencies.

Two API paths for different workflows

Google says 3.5 Transcribe is built to fit developer pipelines such as voice agents, real-time captioning, and post-call analytics. It is exposed through two APIs:

  • Real-time streaming (Live API)
  • Model: `gemini-3.5-transcribe-live`
  • Continuous bidirectional streaming with sub-second latency for interactive voice apps.
  • Pre-recorded processing (Interactions API)
  • Model: `gemini-3.5-transcribe`
  • For meetings, call logs, and recorded audio, with speaker attribution and word-level timestamps.

Core capabilities highlighted by Google

Google presents 3.5 Transcribe as a more “intelligent” transcription layer with features including:

  • Smart transcription: handles self-corrections (e.g., “Tuesday—no, Wednesday”), removes filler words, and auto-formats output.
  • Function calling: can delegate complex tasks (such as image generation and file analysis) to other Gemini models; currently available in the Gemini macOS app.
  • Higher accuracy: Artificial Analysis reports average WER of 4.0% (streaming) and 2.6% (non-streaming).
  • Custom vocabulary: adapts to user-provided terminology and unique spellings.
  • Global language support: automatic detection and transcription across 85+ languages, including regional accents and dialects.
  • Multi-speaker identification: pre-recorded attribution for up to three speakers with timestamps (support for more than three is experimental).

Claimed gains over Chirp 3

Google frames this release as a major step beyond Chirp 3, citing improvements in capability breadth, error rates, and latency.

Based on Artificial Analysis measurements, Google says time to final transcription improves by 70%. On the FLEURS benchmark across major languages/locales, it reports:

  • 5.50% WER in streaming mode
  • 5.04% WER in non-streaming mode

Product integration across Google surfaces

Beyond API availability in Google AI Studio and Gemini Enterprise Agent Platform, Google is rolling 3.5 Transcribe into user-facing products:

  • Gboard on Android (Rambler): converts spoken thoughts into formatted text, removes filler words, and supports voice-based edits, corrections, and style changes.
  • Google Antigravity: with user permission, combines screen context and chat history to improve transcription precision for file names, active documents, and related context.
  • Google AI Studio (Build mode): enables voice-driven app building and coding workflows.
  • Gemini app on macOS: supports natural-speech transcription plus voice commands tied to screen context; can invoke other Gemini models for tasks like local file summarization, cross-app text repurposing, and image generation at the cursor.
  • Chrome (coming soon): voice typing in web fields for replies, post drafting, and prompting Gemini more naturally.

Bottom line

Gemini 3.5 Transcribe is positioned as more than a standard speech-to-text engine. Google is presenting it as a context-aware voice interface layer that combines transcription quality, formatting intelligence, and task orchestration. With both developer APIs and broad product integration, the release signals a push toward voice-first workflows across Google’s ecosystem.

Source