Google DeepMind has released Gemini 3.5 Transcribe, a new AI model specifically designed for intelligent transcription tasks. The announcement, part of a broader update to the Gemini family in August 2026, positions the new model as a significant advancement in automatic speech recognition technology.
The release builds on DeepMind’s earlier work in speech processing and represents a focused effort to address longstanding challenges in transcription accuracy, particularly with accented speech, background noise, and multilingual content.
Key Features of Gemini 3.5 Transcribe
The new model brings several improvements over previous transcription systems:
- Enhanced multilingual support with particular improvements in non-Western language recognition
- Noise robustness maintaining accuracy in challenging audio environments
- Speaker diarization distinguishing between different speakers in multi-person conversations
- Contextual understanding improving accuracy for domain-specific terminology
“Transcription is foundational to making information accessible,” DeepMind noted in the announcement. “Gemini 3.5 Transcribe represents our commitment to building AI that understands the nuances of human speech across all its diversity.”
Market Position and Competition
The release enters a competitive transcription market dominated by specialized providers including Rev, Otter.ai, and various enterprise-focused solutions. Google’s entry leverages the company’s broader AI infrastructure, potentially allowing for tight integration with other Gemini-powered services.
The timing coincides with increased enterprise interest in transcription for meeting notes, customer service analysis, and accessibility compliance. Regulatory requirements in multiple jurisdictions have also driven demand for accurate, automated transcription services.
Availability and Pricing
Gemini 3.5 Transcribe is available through Google Cloud’s Speech-to-API, with tiered pricing based on usage volume. The company has also released an optimized version for on-premise deployment for organizations with strict data residency requirements.
Early benchmarks shared by Google show significant improvements in Word Error Rate (WER) compared to previous models, particularly in challenging acoustic conditions. Third-party evaluations are still underway as the community assesses the model’s performance in real-world deployments.