Google just rolled out its most polished speech-to-text system yet: Gemini 3.5 Transcribe, a dedicated model built to turn messy, real-world audio into clean, structured text with far fewer errors and a lot more context awareness. Announced this week and now in public preview for developers, it’s positioned as the successor to Google’s Chirp 3 and arrives as the company tightens its grip on the enterprise and creator workflows that depend on reliable transcription.

At its core, Gemini 3.5 Transcribe is a speech-to-text engine that lives inside Google’s Gemini API family, but it’s not just another audio mode bolted onto a chat model. It’s a purpose-built transcription model that ships in two flavors: one for live, real-time streaming with sub-second latency, and another for processing pre-recorded audio like meetings, interviews, or call logs.

• The live variant, gemini-3.5-transcribe-live, is meant for interactive voice apps, live broadcasts, or anything that needs continuous, bidirectional streaming.
• The pre-recorded variant, gemini-3.5-transcribe, handles batch-style transcription with speaker attribution and word-level timestamps, ideal for post-call analytics or meeting notes.

Google says the model is designed to “plug seamlessly” into developer workflows, whether you’re building voice agents, captioning tools, or analytics pipelines.

The headline number here is accuracy. On benchmark tests cited by Google, Gemini 3.5 Transcribe hits an average word error rate (WER) of 4.0% for streaming and 2.6% for non-streaming use cases. In plain terms, that means it’s getting roughly 97 out of every 100 words right on recorded audio, which is a meaningful jump over prior systems and puts it in the same ballpark as other top-tier transcription services.

But it’s not just about raw accuracy. Google is pushing “smart transcription” features that make the output feel more human-edited:

• It automatically handles self-corrections (like “let’s meet Tuesday—no, Wednesday”) without making the transcript look messy.
• It strips out filler words (“ums,” “ahs”) and auto-formats text for readability.
• It supports custom vocabulary, so industry jargon, product names, or unique spellings can be recognized more reliably.

On the speed front, Google claims a 70% improvement in time-to-final-transcript compared to Chirp 3, the model it’s replacing. For teams processing hours of audio daily, that’s a real productivity gain.

One of the more practical upgrades is language coverage. Gemini 3.5 Transcribe auto-detects and transcribes more than 85 languages, including regional accents and dialects, and it can handle mid-session code-switching (when speakers mix languages in the same conversation).

For multi-speaker scenarios, the pre-recorded mode supports speaker diarization—basically labeling who said what—with timestamps for up to three speakers reliably, and experimental support beyond that (up to eight in some configurations). That’s useful for podcasts, panel discussions, or customer support calls where tracking speakers matters.

Right now, Gemini 3.5 Transcribe is available to developers via:

• Google AI Studio, where you can prototype apps that use voice input and transcription on the fly.
• The Gemini Enterprise Agent Platform, for building more robust, production-grade voice agents and workflows.
• The Gemini macOS app, where it powers both clean dictation and voice commands that can pair with screen context for more complex tasks.

For everyday users, the most visible impact will likely be in tools that rely on transcription under the hood—think meeting summaries, call analytics, accessibility features, or even creator tools that auto-generate captions and show notes.

Google isn’t alone in this space. OpenAI, AssemblyAI, and others have been pushing hard on transcription quality and pricing. Early analysis suggests Gemini 3.5 Transcribe is cheaper than OpenAI’s GPT-Transcribe on streaming but pricier than AssemblyAI’s batch option, positioning it as a mid-to-high tier option with a focus on accuracy and integration with Google’s broader AI stack.

What sets it apart is the combination of low latency, strong multilingual support, and tight integration with Google’s ecosystem—especially for teams already using Google Cloud, Meet, or other Google AI services.

Gemini 3.5 Transcribe lands at a time when voice is becoming a more central part of how we interact with software. From customer service bots to creator workflows, the ability to convert speech to text accurately and quickly is a foundational capability. Google’s bet here is that developers and enterprises will choose the model that not only transcribes well but also understands context, cleans up the output, and plays nicely with other AI tools in the stack.

For now, it’s in public preview, which means pricing and limits could shift as Google gathers feedback. But if the early numbers hold, Gemini 3.5 Transcribe could quickly become a default choice for anyone building voice-first applications or needing reliable, scalable transcription.

If you’re a developer or product team, the practical next step is to test it in Google AI Studio or the Enterprise Agent Platform and see how it handles your specific audio—accents, background noise, domain-specific terms, and all. For the rest of us, expect to see its fingerprints on more of the tools we use every day, from meeting recaps to live captions, as Google pushes voice further into the mainstream.