Google Releases Gemini 3.5 Transcribe for Ultra-Low-Latency Speech-to-Text

Google has announced the general availability of Gemini 3.5 Transcribe, a specialized audio intelligence model capable of real-time speech recognition, multilingual translation, and acoustic background filtering across 100+ languages. Designed for enterprise developers, the model achieves state-of-the-art word error rate (WER) scores.

Google AI Research has introduced Gemini 3.5 Transcribe, a highly optimized speech recognition model designed to process live audio streams with minimal latency. Available immediately via Google Cloud Vertex AI, the model handles complex multi-speaker conversations and accented audio with remarkable precision.

Benchmark evaluations demonstrate that Gemini 3.5 Transcribe achieves industry-leading word error rates (WER), outperforming existing commercial speech APIs across noisy environments. The architecture incorporates advanced acoustic filtering that isolates background chatter and environmental noise.

In addition to raw transcription, Gemini 3.5 Transcribe performs simultaneous translation, emotion analysis, and automated timestamp alignment. Developers can integrate the API into customer support call centers, video conferencing platforms, and accessibility software.

Google's release underscores the rapid evolution of specialized audio LLMs. By combining speech recognition directly with semantic reasoning, Gemini 3.5 Transcribe simplifies the pipeline for building real-time conversational voice assistants.