# Google ships Gemini 3.5 Transcribe for real-time speech apps > Source: > Published: 2026-08-26 18:10:14+00:00 # Google ships Gemini 3.5 Transcribe for real-time speech apps **The public-preview model handles 85+ languages, custom jargon and up to three speakers across live and recorded audio.** By [Ryan Merket](/author/ryan-merket) ยท Published Primary source: [X](https://x.com/sundarpichai/status/2092659467284517088) ## Why it matters Google is packaging speech recognition as a native input layer for Gemini agents, reducing the plumbing needed to connect live audio, transcription and model-driven actions. Google released Gemini 3.5 Transcribe on August 26th, giving developers a dedicated [speech-to-text](/models/fal/speech-to-text-stream) model for live voice interfaces and recorded audio inside its Gemini developer stack. [https://x.com/sundarpichai/status/2092659467284517088](https://x.com/sundarpichai/status/2092659467284517088) [Google and Alphabet CEO Sundar Pichai (@sundarpichai)](https://x.com/sundarpichai) announced the release [in a post on X](https://x.com/sundarpichai/status/2092659467284517088), pitching the model as a way to build applications that recognize a speaker's words and intended edits, including conversations involving multiple people. The API is available in public preview through Google AI Studio and Google Antigravity, with an enterprise preview through the Gemini Enterprise Agent Platform. The release turns transcription into a named Gemini product with separate endpoints for streaming and prerecorded audio. Developers can use `gemini-3.5-transcribe-live` through Google's Live API for continuous, bidirectional streaming with what Google describes as sub-second latency. The `gemini-3.5-transcribe` endpoint runs through the Interactions API for meetings, call recordings and other completed audio, adding speaker attribution and word-level timestamps. That division matters for developers building voice agents. A customer-service bot needs partial text quickly enough to respond during a conversation, while a meeting or post-call analysis product can wait for a more complete transcript with speaker labels and precise timing. Google is putting both workloads behind the same Gemini platform rather than asking developers to assemble separate recognition and reasoning systems. ### Transcription that edits while it listens Gemini 3.5 Transcribe is designed to produce cleaned-up text rather than a literal record of every sound. Google says the model can remove filler words, format sentences and resolve spoken corrections such as changing a meeting from Tuesday to Wednesday mid-sentence. It also accepts custom vocabulary intended to improve recognition of product names, proper nouns and specialized jargon. The model automatically detects more than 85 languages, according to Google, and can handle language changes within a live stream. For prerecorded audio, it can identify and timestamp as many as three speakers. Support for larger groups remains experimental, limiting the initial diarization pitch for crowded meetings, conferences and overlapping group conversations. Google says measurements by Artificial Analysis produced an average word error rate of 4.0% for streaming transcription and 2.6% for non-streaming use. Lower word error rates indicate fewer substitutions, omissions and insertions, though results can vary sharply with accents, microphone quality, background noise and specialist terminology. On Google's multilingual FLEURS evaluation, the model recorded a 5.50% word error rate in streaming mode and 5.04% for non-streaming transcription. Google also says the time required to finalize a transcript improved by 70% compared with Chirp 3, its earlier speech-recognition model. Those performance figures come from Google's [launch announcement](https://goo.gle/3UlQoob); developers will still need to test the model against their own audio and latency requirements. ### Google folds speech recognition into the Gemini stack The strategic piece is the connection between transcription and the rest of Gemini. Google is positioning speech recognition as an input layer for agents, coding tools and multimodal workflows rather than a standalone conversion service. Inside the Gemini app for macOS, the model can pass a spoken request to other Gemini models through function calls. Google lists image generation, file analysis, document summarization and search as examples. That function-calling feature is currently limited to the macOS app, so developers using the transcription endpoints should not assume the same orchestration arrives automatically with API access. Google has already deployed the underlying model in consumer products. The Gemini app for macOS uses it to turn free-form dictation into formatted text, while Rambler on Android can remove filler words and apply spoken edits. Google says Gemini 3.5 Transcribe will also come to Chrome for dictation in web fields. The model also draws on context in Google's own development tools. In Antigravity, Google says transcription can use the active document, file names and chat history, with user permission, to improve recognition. In AI Studio's Build mode, developers can use voice while creating applications. Google is supplying integrations through platforms including Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel and Vision Agents. Those services handle parts of the streaming and communications infrastructure around the model, giving Google distribution among developers already assembling voice-agent systems. Gemini 3.5 Transcribe enters public preview with the technical split that voice developers need: a low-latency stream for interactive systems and a richer processing path for completed recordings. Google's larger bet is that transcription becomes another native Gemini capability, feeding directly into agents that interpret a request and act on it before the user returns to a keyboard.