Google launches Gemini 3.5 Transcribe for real-time speech-to-text with 2.6% word error rate

Google's new transcription model

Google has announced Gemini 3.5 Transcribe, a speech-to-text model designed to turn live audio into polished, ready-to-read text instead of a literal word-for-word transcript. In a company blog post dated 26 August 2026, Google calls it its most accurate transcription model yet and frames it as an upgrade over Chirp 3, with lower error rates and faster delivery of finished text.

Consumers have already been using the model through voice features such as Rambler on Android and in the Gemini app on macOS, according to Google. The announcement opens the same technology to developers in public preview through the Gemini API in Google AI Studio and on the Gemini Enterprise Agent Platform.

Two APIs cover live and recorded audio

The model ships through two separate interfaces. The first, gemini-3.5-transcribe-live, runs on the Live API and streams audio in both directions with latency below one second, aimed at interactive voice applications. The second, gemini-3.5-transcribe, works through the Interactions API and processes recorded material such as meetings and call logs, with speaker attribution and word-level timestamps.

Reported accuracy gains

Citing measurements from Artificial Analysis, Google reports an average word error rate of 4.0% for streaming use and 2.6% for non-streaming use. The company also says the model holds up in noisy real-world settings and captures alphanumeric details such as postal codes and order IDs accurately.

On the FLEURS benchmark, across a set of leading languages and locales, Google reports 5.50% WER in streaming mode and 5.04% in non-streaming mode, improvements over Chirp 3. The time until a final transcript is delivered improves by 70% compared with the older model, again per Artificial Analysis.

Transcription that cleans up after itself

The model is built to interpret how people actually speak. Google says it handles mid-sentence corrections, such as switching a meeting from Tuesday to Wednesday, strips filler words like "um" and "ah", and applies automatic formatting to the resulting text.

Developers can supply custom vocabulary — specialized jargon and particular spellings — that the model adapts to. Language detection is automatic across more than 85 languages, including regional accents and dialect variations. For pre-recorded audio, the model attributes speech to individual speakers with timestamps for up to three speakers, with support for more than three labeled experimental.

The model can also use function calls to hand off complex work, such as image generation or file analysis, to other Gemini models. Google notes this capability is currently available in the Gemini macOS app.

Rolling out across Google's surfaces

Beyond the APIs, Google is embedding the model into its own products. On Gboard for Android, the Rambler feature converts spoken thoughts into structured text and accepts voice edits, including spelling corrections and changes to writing style. In Google AI Studio's Build mode, developers can construct applications by voice while they are being built.

Google Antigravity pairs transcription with screen context and chat history, with user permission, to improve accuracy on items like file names and active documents. In the Gemini app on macOS, voice commands can combine with screen context to run multi-step workflows, such as summarizing local files or generating images at the cursor. A feature that lets users dictate into any text field on the web is planned for Chrome.

Google also lists partner platforms — including Agora, LangChain, LiveKit, Pipecat and Vercel — building on the Live API, and points to early feedback from vivo, Intellitek Health and Lingopal highlighting latency, accuracy and language coverage.

Why it matters

Speech-to-text systems have historically produced raw transcripts that need substantial cleanup before they are useful. By moving formatting, filler removal and correction handling into the model itself — and letting it call other models for follow-up tasks — Google is treating transcription as a language-understanding problem rather than a pure audio-processing one.

For teams building voice agents, captioning tools and call analytics pipelines, the reported error rates and latency gains over Chirp 3 lower a real barrier, while custom vocabulary and speaker attribution address long-standing enterprise pain points. The integration into Gboard, Chrome and the Gemini app also suggests Google wants voice input to become a default interaction mode across its products rather than a niche option.