Developers and enterprise teams have been waiting for Gemini 3.5 Pro, anticipating Google’s next flagship system to deliver substantial advances in deep reasoning, complex code architecture, and autonomous multimodal agent tasks. Instead of rolling out that complete Pro model, Google introduced another specialized variant today: Gemini 3.5 Transcribe.
Gemini 3.5 Transcribe: Google’s dedicated speech-to-text model for live and pre-recorded audio workflows
Listen to the official audio announcement:
A specialized speech model for live and recorded audio
Gemini 3.5 Transcribe focuses entirely on converting spoken audio into structured, formatted text. Standard speech recognition tools often struggle with background noise, specialized jargon, and messy speech habits. Google built 3.5 Transcribe to parse raw audio directly while handling disfluency cleanup and context mapping in the background.
The model runs across two distinct API interfaces:
gemini-3.5-transcribe-live on the Live API provides continuous bidirectional streaming with sub-second latency for real-time voice applications.
gemini-3.5-transcribe on the Interactions API processes pre-recorded files, meeting logs, and call records, including word-level timestamps and speaker attribution for up to three speakers (with experimental support for larger groups).
Gemini 3.5 Transcribe handles live language switches and continuous streaming transcription without losing context.
Side-by-side comparison showing Gemini 3.5 Transcribe filtering speech hesitations and correcting mid-sentence changes in real time.
Real-time interactive transcription with word-level timestamps and speaker attribution in a specialized discussion.
Accuracy and benchmark results
According to evaluations by Artificial Analysis, Gemini 3.5 Transcribe achieves a 4.0% Word Error Rate (WER) in live streaming mode and 2.6% in non-streaming workloads. On the multilingual FLEURS benchmark across top languages and locales, the model records a 5.50% WER for streaming and 5.04% for pre-recorded audio.
Compared to Google’s previous Chirp 3 speech model, time to final transcription drops by 70%. The model can detect and transcribe more than 85 languages automatically, accommodating regional accents, background noise, and alphanumeric strings such as postal codes and order IDs. Teams can also supply custom vocabulary to improve the recognition of internal company terms and technical jargon.
FLEURS benchmark evaluation: Word Error Rate reduction across global languages compared to Chirp 3
Metric / Benchmark
Gemini 3.5 Transcribe (Live)
Gemini 3.5 Transcribe (Non-Streaming)
Chirp 3 Comparison
Word Error Rate (Artificial Analysis)
4.0% WER
2.6% WER
Improved accuracy
FLEURS Multilingual Benchmark
5.50% WER
5.04% WER
Significant error reduction
Time to Final Transcription
Sub-second latency
Fast batch execution
70% latency improvement
Language Coverage
85+ auto-detected languages
85+ auto-detected languages
Broader dialect support
Speaker Diarization
Real-time turn tracking
Up to 3 speakers (3+ experimental)
Word-level timestamps
Integration across Google products
Google is integrating 3.5 Transcribe directly into several operating system surfaces and developer tools:
On Android, Gboard uses the model in its Rambler feature to transcribe speech, strip out filler words like “um” or “ah”, fix mid-sentence self-corrections, and edit text through voice commands.
In Google Antigravity, the model uses screen context and chat history, with user permission, to transcribe programming identifiers, file names, and technical terms accurately.
In Google AI Studio, developers can use voice input directly in Build mode when prototyping applications.
The Gemini macOS application pairs voice transcription with local screen context to run background agent tasks, such as file inspection or image generation.
Google Chrome is slated to receive talk-to-type functionality across standard web input fields and browser prompts.
Rambler on Android filters filler words and applies voice-directed text editing inside Gboard.
In Google Antigravity, 3.5 Transcribe pairs screen context with chat history to transcribe code symbols, file paths, and terminal commands.
Gemini desktop application on macOS executing background agent workflows through screen context and voice commands.
Availability and partner ecosystem
Several voice infrastructure and framework providers have added integration with the Live API, including LiveKit, LangChain, Vercel, Agora, Pipecat, Fishjam, and Vision Agents. These platforms manage real-time media streaming pipelines while developers build custom conversational logic. Early adopters testing the model include Vivo, Intellitek Health, and Lingopal.
Gemini 3.5 Transcribe is available in public preview:
For developers through the Gemini API in Google AI Studio and Google Antigravity.
For enterprise deployments via the Gemini Enterprise Agent Platform, with planned expansion to customer experience workflows.
For consumer testing through the Gemini macOS app (in English) and via Rambler on Android in select countries.
I’m really curious about the accuracy benchmarks mentioned in the article. It’ll be interesting to see how they compare to previous models.