Gemini 3.5 Transcribe is a Google speech-to-text model based on Gemini’s audio understanding capabilities. It processes prerecorded audio files through the Interactions API.
The model provides automatic language identification, speaker diarization, word-level timestamps, custom vocabulary hints, and smart transcription features.
| Developer | |
|---|---|
| Model ID | google/gemini-3.5-transcribe |
| Release Date | October 14, 2026 |
| Modalities | Audio → Text |
Multilingual transcription
The model automatically detects speech across more than 85 languages and supports multilingual code-switching.
Speaker identification
It can distinguish individual speakers in prerecorded audio and attach speaker labels to transcript segments.
Word-level timing
The model can return timestamps for individual recognized words.
Custom vocabulary
Custom vocabulary biasing supports domain-specific terms, acronyms, proper names, and other specialized phrases.
Smart transcription
Smart transcription supports disfluency removal and formatting intended to produce cleaner transcripts.
Meetings and calls
The model can transcribe recorded meetings, support calls, and call logs with speaker attribution.
Subtitles and searchable audio
Word-level timestamps can support synchronized subtitles and navigable audio players.
Lectures and voice notes
The model can transform prerecorded lectures and voice notes into structured text.
Domain-specific recordings
Custom vocabulary can improve recognition of specialized jargon, company names, and proper nouns.
Sources: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-5-transcribe/ · https://ai.google.dev/gemini-api/docs/models/gemini-3.5-transcribe · https://aistudio.google.com/models/transcribe
Real-time captioning: The model is suited to near-real-time captioning by streaming microphone or live-audio input and returning transcription results incrementally.
Voice agents and interactive voice applications: Its low-latency streaming transcription supports voice agents and interactive voice apps.
Post-call analytics: The model is suited to post-call analytics pipelines using transcribed call audio.
Transcription of prerecorded recordings: It can transcribe recorded meetings, call logs, voice notes, support calls, and lectures into structured text with speaker attribution and timestamps.
OpenAI-compatible — swap the base URL and it works with your existing SDK.
| Type | Price |
|---|---|
| Output | |
A free tier is available for Gemini 3.5 Transcribe within Google AI Studio and Gemini API quota limits; free-tier content may be used to improve Google products.
| Benchmark | Score | What it measures | Source | Retrieved |
|---|---|---|---|---|
| FLEURS WER | 5.04% | WER on multilingual speech (102 languages) | Source | October 15, 2026 |
The primary published source is Google’s “Intelligent transcription with Gemini 3.5 Transcribe” announcement. Source: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-5-transcribe/
