Gemini 3.5 Transcribe Live

Google's launch blog post from August 26, 2026 introduces Gemini 3.5 Transcribe and splits availability across two distinct APIs: gemini-3.5-transcribe-live serves continuous, bidirectional, sub-second-latency streaming via the Live API, while gemini-3.5-transcribe handles pre-recorded audio (meetings, call logs) via t Function calling is highlighted as a way to delegate complex tasks during voice interactions, and the post notes consumer-facing use already shipping in Rambler on Android and the Gemini app on macOS before the developer API opened up in Google AI Studio and the Gemini Enterprise Agent Platform. This is the primary ann

A third-party recap confirms the August 26, 2026 launch and adds Artificial Analysis accuracy figures: average Word Error Rate of 4.0% for streaming and 2.6% for non-streaming, with FLEURS at 5.50% WER streaming and 5.04% non-streaming. It reiterates that the pre-recorded gemini-3.5-transcribe path carries meeting feat The recap enumerates four smart-transcription capabilities (disfluency cleanup, filler-word removal, auto-formatting, and custom-vocabulary handling) and contrasts them with conventional speech recognition that struggles with background noise and jargon. It is derivative of Google's own announcement but useful for cons

MarkTechPost's 27 Aug 2026 write-up details Google's launch of Gemini 3.5 Transcribe as two distinct endpoints: gemini-3.5-transcribe accessed via the Interactions API for pre-recorded files, and gemini-3.5-transcribe-live accessed via the Live API for bidirectional streaming. It cites Artificial Analysis word error ra The piece also highlights product features relevant to voice-agent builders, including speaker diarization for up to three speakers, filler-word removal, custom-vocabulary support, and roughly a 70% improvement in time-to-final-transcription versus Google's prior Chirp 3 model, giving developers concrete reasons to eva

Times of India coverage from 27 Aug 2026 describes Gemini 3.5 Transcribe as a speech-to-text model that tidies filler words in real time, formats output as it goes, and adapts to a custom vocabulary list for jargon-heavy workflows. It reports that Chrome voice typing is slated to receive the model next, and that the ne Quantitative claims attributed to Artificial Analysis include a 70% improvement in time-to-final-transcription, a 4% word error rate on streaming audio and 2.6% on pre-recorded files, and FLEURS multilingual scores of 5.50% streaming and 5.04% non-streaming versus 7.32% for Chirp 3 on live speech. Language coverage str

Coverage dated 26 Aug 2026 reports Artificial Analysis benchmark results for Google's two new transcription models: Gemini 3.5 Transcribe ranks #5 on the AA-WER leaderboard at 2.6% word error rate, while Gemini 3.5 Transcribe Live posts a 4.0% AA-WER Streaming score with latency of just 0.40 seconds after speech ends. The article frames the release as Google stepping into the same speech-API lane where OpenAI, ElevenLabs, and AWS already compete for voice-agent contracts, arguing that the 0.40-second streaming latency is the kind of figure that decides whether a voice-agent product feels responsive in production. Both models are not

A Vercel changelog mirror entry dated 26 Aug 2026 announces 'Gemini 3.5 Transcribe now available on AI Gateway,' noting that google/gemini-3.5-transcribe handles complete audio files while a Live variant supports streaming. The same changelog covers other recent Gateway additions such as Ling 3.0 Flash Fin (free throug Adjacent entries describe Cursor being added to the AI SDK harness layer via an @ai-sdk/harness-cursor adapter on 27 Aug 2026, Claude Managed Agents becoming runnable with Chat SDK on 28 Aug 2026, and redesigned deployment filters on 27 Aug 2026. Together these confirm the timeframe and platform context in which the Tr

Vercel's AI Gateway Speech-to-Text modality docs describe how to transcribe recorded audio using models such as openai/whisper-1, openai/gpt-4o-transcribe, and google/gemini-3.5-transcribe, and how to handle live audio with streaming transcription. The page specifically names google/gemini-3.5-transcribe-live as one of Integration guidance includes the experimental streamTranscribe API that streams results back as the provider produces them, AI SDK version requirements of ai 7.0.31 and @ai-sdk/gateway 4.0.23 or later, and a browser-side pattern using gateway.experimental.transcription.getToken to mint a single-use client secret that

An aggregator page dedicated to Gemini 3.5 Transcribe Live records the FLEURS streaming WER at 5.50% (reported as 1-WER accuracy 0.94) and attributes it to blog.google as a self-reported, not independently verified score. It also lists the model's ranking against older Gemini variants on the FLEURS speech-to-text leade The same page lists provider pricing of $3.50 per million input tokens and $21.00 per million output tokens attributed to Google (not the Vercel AI Gateway), with no cached-input or TTFT figures available. Per policy this is upstream Google pricing rather than gateway-specific news, so it should not be cited as a Verce