A 2.6% word error rate on pre-recorded audio is a number worth stopping at. That’s what Google is claiming for Gemini 3.5 Transcribe, the company’s latest speech-to-text model, which was announced on August 26, 2026. For context, that’s competitive with the best numbers Deepgram and AssemblyAI publish, and it puts pressure on OpenAI’s Whisper-based offerings, which have historically struggled in noisy, real-world conditions.
The model comes in two forms. One handles live, real-time streaming via the Live API with sub-second latency. The other processes pre-recorded audio through the Interactions API, with speaker attribution and word-level timestamps for up to three speakers. Experimental support for more than three speakers is included too, though Google is careful to flag that as a work in progress.
What the model actually does differently
Most speech-to-text models convert audio to text and stop there. Gemini 3.5 Transcribe goes further by cleaning up the output intelligently. It strips filler words, handles self-corrections in natural speech, and auto-formats the transcript. So when someone says “let’s meet Tuesday, no, Wednesday,” the model outputs the corrected intent, not the mess.
That’s not just a nice-to-have. For post-call analytics, meeting notes, or voice agent transcripts, raw output is often unusable without a cleanup pass. Baking that into the model saves a step and reduces error compounding.
The feature list is worth laying out clearly:
- Smart transcription with disfluency removal and auto-formatting
- Function calling to delegate tasks like image generation or file analysis to other Gemini models
- Custom vocabulary support for domain-specific jargon and unusual spellings
- Automatic language detection across 85-plus languages with dialect handling
- Multi-speaker identification with timestamps for pre-recorded audio
- Word error rates of 4.0% for streaming and 2.6% for non-streaming, as measured by Artificial Analysis
The benchmark comparison against Chirp 3, Google’s previous transcription model, also matters here. Time to final transcription improves by 70%. On the FLEURS benchmark, the model hits 5.50% WER in streaming mode and 5.04% in non-streaming, both better than Chirp 3 across the tested language set.
Where the model shows up for end users
Beyond the API, Google is pushing 3.5 Transcribe into its own products. Rambler on Android uses the model to turn spoken thoughts into clean, formatted text with voice-based editing. On Google Antigravity, it pairs with screen context and chat history to sharpen accuracy across file names and active documents. In the Gemini app on macOS, it goes further still, allowing users to summarize files, generate images, and run complex workflows by voice, with other Gemini models handling the heavy lifting in the background.
Chrome is next. Google says users will soon be able to dictate into any web field using 3.5 Transcribe, which is a meaningful surface area given how many people still type replies, draft posts, and prompt AI tools manually.
Developer access and availability
For developers, 3.5 Transcribe is in public preview through Google AI Studio and the Gemini API. Enterprises can access it via the Gemini Enterprise Agent Platform, with a Gemini Enterprise for Customer Experience integration coming soon. Platforms like LangChain, LiveKit, Pipecat, and Vercel already have integrations built on the Live API.
So is this noise or something real? It’s real. The accuracy numbers are independently benchmarked, the latency improvement is significant, and the smart cleanup features address a genuine gap in existing tools. If you’re building voice agents or processing call audio, this is worth testing against your current stack.
