Google released Gemini 3.5 Transcribe on August 26, a speech-to-text model available through the Gemini API in Google AI Studio and through the Gemini Enterprise Agent Platform. The model targets developers building voice agents, live captioning tools, and post-call analytics pipelines rather than end consumers directly.
The pitch is not just lower error rates. Gemini 3.5 Transcribe is designed to output already-formatted text: it strips filler words, resolves self-corrections like “let’s meet Tuesday, no, Wednesday,” and applies formatting inline. That removes a step most transcription pipelines currently push downstream to a second LLM call or a rules-based cleanup layer. If the model output already reads as finished prose, builders can cut a stage out of their stack, which is the kind of consolidation Google has been pursuing across Gemini generally.
Google ships the model in two API forms. The Live API, using gemini-3.5-transcribe-live, covers two-way live streaming at latencies under a second, aimed at conversational voice products. The Interactions API, using gemini-3.5-transcribe, handles pre-recorded audio such as meetings and call logs, tagging who said what and timing each individual word, for as many as three voices (support beyond three is labeled experimental).
Google says the model, measured by Artificial Analysis, achieves a 4.0 percent word error rate on streaming transcription and 2.6 percent on non-streaming use. Those figures come from Google’s own release and were not independently verified in the announcement. Google also claims a 70 percent improvement in time-to-final-transcription over its prior transcription model, Chirp 3, and reports FLEURS scores of 5.50 percent word error rate when streaming and 5.04 percent when not, averaged over its best-supported languages.
This is a vendor claiming its own numbers against its own prior model. Google’s post does not include a side-by-side benchmark against OpenAI’s Whisper, Deepgram, AssemblyAI, or ElevenLabs Scribe, the incumbent transcription vendors that most production voice pipelines currently run on. Builders evaluating a switch will need to run their own comparison rather than take the FLEURS and Artificial Analysis figures as the full picture.
The model also supports function calling, letting it pass work such as generating an image or reading a document across to sibling Gemini models part-way through an exchange, a feature currently live in the Gemini app on macOS. Google lists 85-plus supported languages and says the model adapts to custom vocabulary, useful for teams transcribing domain jargon or alphanumeric strings like order IDs and postal codes.
Several developer platforms, including LiveKit, Pipecat, LangChain, and Vercel, already route to Gemini’s Live API for their voice infrastructure, according to Google’s post, which is the more concrete signal here: real integration partners building on the API before general availability. Google has not disclosed pricing for the transcription endpoints separately from its existing Gemini API rate card, and the release remains in public preview for both developer and enterprise access.
Teams currently paying per-minute for Whisper-based or Deepgram pipelines should benchmark Gemini 3.5 Transcribe’s formatted output against their existing cleanup layer before the next contract renewal, since removing that step is the actual cost argument here, not the WER delta.
Announced by Google on 26 August 2026 in a post on the Google blog titled “Intelligent transcription with Gemini 3.5 Transcribe.”
