TL;DR: Google on Aug. 26, 2026 introduced Gemini 3.5 Transcribe, its most precise speech-to-text model, now in public preview for developers and already powering Rambler on Pixel 11 and voice input in the Gemini app on macOS.

What Google announced

Google introduced Gemini 3.5 Transcribe on Aug. 26, 2026 as a speech-to-text model for intelligent voice interactions and real-time transcription, according to the official Google blog post by Diego Melendo Casado and Luke Leonhard of Gemini Audio. The company described it as its most precise speech-to-text model yet, designed to convert raw audio directly into accurate, polished, formatted text. The announcement positioned the system against conventional speech recognition models that struggle with background noise, complex jargon, and disfluency cleanup. Google said consumers were already benefiting from the model through new voice capabilities such as Rambler on Android and voice features in the Gemini app on macOS. The company made the model available to developers through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform. Ars Technica, reporting on Aug. 26, 2026, described the release as a different branch from the awaited Gemini 3.5 Pro, focused on streamlining voice input by editing out hesitations and corrections while outputting cleaned text. The Ars Technica report noted that the model already powered the Gboard Rambler feature on Pixel 11 and was set to appear across more Google products including Chrome.

How the two transcription modes work

Google built Gemini 3.5 Transcribe to support both live and recorded workflows through two separate APIs, according to the Google blog. Real-time streaming uses the Live API with the model code gemini-3.5-transcribe-live and delivers continuous, bidirectional streaming with sub-second latency for interactive voice apps. Pre-recorded audio processing uses the Interactions API with the model code gemini-3.5-transcribe for recorded audio, meetings, and call logs, with speaker attribution and word-level timestamps. The Google developer documentation for Gemini 3.5 Transcribe lists the same two model codes and distinguishes unary audio file processing from Live API streaming. For file-based requests, the documentation states an audio limit of up to one hour per request, reduced to up to 30 minutes when speaker diarization or word-level timestamps are enabled. For live streaming, the documentation states a maximum of 10 minutes per session. The blog said developers could use the model for voice agents, real-time captioning tools, and post-call analytics pipelines. The developer documentation adds that the model provides utterance-based language detection, speaker diarization, word-level timestamps, smart transcription, and custom vocabulary speech biasing.

Accuracy and speed numbers Google cited

Google cited third-party measurement by Artificial Analysis for word error rate and latency, according to the Google blog and 9to5Google. The blog stated an average Word Error Rate of 4.0% for streaming use cases and 2.6% for non-streaming use cases. It said the model performed strongly in noisy, real-world environments and accurately captured alphanumeric entities such as postal codes and order IDs. For comparison with the prior Chirp 3 transcription model from 2025, Google cited a 70% improvement in time to final transcription, as measured by Artificial Analysis. On the FLEURS benchmark across top languages and locales, Google reported 5.50% WER in streaming mode and 5.04% WER in non-streaming use cases. Ars Technica restated the comparison as about 70 percent faster from voice to final transcribed text, with a live-speech error rate of 5.5 percent versus 7.32 percent for Chirp 3 cited by Google. Ars Technica characterized the live accuracy gain as modest but practically useful because correcting typos from voice input is cumbersome. 9to5Google similarly described the Chirp 3 comparison as a major advancement across capabilities, word error rates, and latency.

Smart editing, languages, and speakers

The model is designed to preserve natural speaking style while cleaning the output to reflect intent, according to the Google blog. Smart transcription handles self-corrections such as “let’s meet Tuesday — no, Wednesday,” removes filler words including “ums” and “ahs,” and auto-formats text. Ars Technica said the system can remove “ums” and “uhs” during speech and edit text when a speaker corrects course, while referring to provided custom vocabulary for specialized jargon. The custom vocabulary feature lets the model recognize specialized jargon and unique spellings by adapting transcription to terms supplied by the user or developer. The developer documentation states custom vocabulary biasing supports up to 1,000 terms on both live and file endpoints, with best results typically seen with up to 100 terms. Global language support covers automatic detection and transcription of over 85 languages, including regional accents and diverse dialects, according to the Google blog and developer documentation. The documentation adds support for mid-session code-mixing and live language switches. For multi-speaker audio, the blog states accurate attribution with timestamps for up to three speakers in pre-recorded audio, with support beyond three speakers labeled experimental. The developer documentation notes that word-level timestamps can degrade transcription accuracy and that custom vocabulary is incompatible with diarization and timestamps.

Voice tasks beyond dictation

Google positioned Gemini 3.5 Transcribe as a way to execute tasks by voice, not only to dictate, according to the Google blog and 9to5Google. Function calling allows the model to delegate complex tasks such as image generation and file analysis to other Gemini models via function calls. Google said that capability was currently available in the Gemini app on macOS, where voice input can pair with screen context to support workflows. In that macOS app, the model transcribes free natural speech into clean formatted text and enables voice commands that can summarize local files, repurpose text across apps, or generate images at the cursor. 9to5Google described the macOS capability as Speak to Window, in which transcription pairs with screen context to handle heavier work in the background. Ars Technica noted the same trade-off: the model aims to capture intended meaning rather than only literal words, which helps short passages but technically changes wording and may not suit every setting. Google also described smart transcription in Google AI Studio Build mode for building apps by voice. The developer documentation states that function calling, caching, code execution, file search, image generation, and thinking are not supported directly by the gemini-3.5-transcribe file model itself, indicating that broader task execution happens through pairing with other Gemini models.

Where consumers will see it

The model was already live in several first-party surfaces at announcement, according to the Google blog. On Gboard on Android, the Rambler feature transforms spoken thoughts into well-formatted text, filters filler words, and supports voice-based edits, spelling corrections, and style changes. Ars Technica reported that Rambler remained limited to Pixel 11 phones at launch, with Google planning expansion to more Gemini Intelligence devices later this year. On Google Antigravity, transcription pairs the prompt-box microphone with screen context and chat history, with user permission, to improve accuracy on file names, agent thoughts, and active documents. In Google AI Studio, Build mode provides access for voice-driven app creation. In the Gemini app on macOS, voice input benefits from Gemini 3.5 Transcribe starting at launch. Google said a Chrome integration is coming soon that will allow talking to type in any web field for replies, posts, or Gemini prompts, according to the Google blog and 9to5Google. Ars Technica likewise said Chrome support was planned for users outside the initial apps and devices. For developers, Google offered public preview in the Gemini API via Google AI Studio and Google Antigravity, while enterprises were offered public preview via the Gemini Enterprise Agent Platform, with availability in Gemini Enterprise for Customer Experience described as coming soon.

What developers and partners get

For developers, Google framed Gemini 3.5 Transcribe as a drop-in component for interactive voice apps and audio pipelines, according to the Google blog. The Live API path targets sub-second latency and continuous streaming, while the Interactions API path targets meetings, call logs, and other recorded material with timestamps and speaker labels. The Google blog said early reviewers using the Live API included developer platforms such as Agora, Fishjam, LangChain, LiveKit, Pipecat, and Vercel. The developer documentation lists August 2026 as the latest model update and points developers to separate audio transcription and live transcription guides plus a pricing page. It also states that batch API, flex inference, and priority inference consumption options are not supported for the transcription model. 9to5Google summarized availability as public preview for developers in AI Studio and Antigravity, and public preview for enterprises through the Gemini Enterprise Agent Platform. The announcement did not list consumer pricing because access reaches consumers through Gboard, Antigravity, the Gemini app, AI Studio, and Chrome rather than as a standalone app.

Sources referenced