Google DeepMind launches Gemini 3.5 Transcribe, a new speech-to-text model with real-time and pre-recorded transcription APIs

Google DeepMind released Gemini 3.5 Transcribe, a speech-to-text model available through two new API endpoints: one for real-time streaming and one for pre-recorded audio with speaker attribution and word-level timestamps. It is in public preview across the Gemini API, Google AI Studio and Gemini Enterprise Agent Platform, and is already integrated into consumer products including Rambler on Android, the Gemini macOS app and Google Antigravity. It replaces Chirp 3 as Google's transcription model.

Developers building voice or transcription features now have a new default Google option to evaluate against Whisper, Deepgram, AssemblyAI and ElevenLabs, with third-party-measured error rates (4.0% streaming, 2.6% non-streaming) that are genuinely competitive if they hold up in practice. Enterprises using Google's stack get an immediate upgrade path since the model is already live in several products, but anyone planning production deployment should note this is preview software with no published pricing or rate limits yet. The broader significance is limited: this is an incremental capability upgrade within Google's existing ecosystem, not a shift in who controls voice AI infrastructure.

The "most precise speech-to-text model yet" framing and the 70% latency improvement are Google's own claims, sourced partly from Artificial Analysis but not independently replicated here, and preview-stage performance often changes before general availability. The long list of partner and integration names (Agora, LangChain, LiveKit, Pipecat, Vercel, vivo, etc.) reads as ecosystem packaging to signal momentum rather than evidence of proven real-world adoption at scale.

1. 01Independent WER and latency benchmarks (e.g. from Artificial Analysis or third-party developers) replicating the claimed 4.0%/2.6% error rates and 70% latency improvement outside Google's own post
2. 02Pricing and rate limits once the API exits public preview, since cost will determine whether it actually displaces Whisper, Deepgram, AssemblyAI or ElevenLabs in production pipelines
3. 03Whether multi-speaker attribution beyond three speakers and the promised Chrome and Gemini Enterprise for Customer Experience integrations ship, or remain stuck as experimental/coming soon