Google released Gemini 3.5 Transcribe on August 26, a dedicated speech-to-text model split into two tracks: real-time streaming and pre-recorded audio.
According to measurements by Artificial Analysis, the non-streaming version has an average word error rate of 2.6%, versus 4.0% for the streaming version. On the multilingual FLEURS benchmark, the figures are 5.50% for streaming and 5.04% for non-streaming. Compared with the previous-generation Chirp 3, the time to produce a final transcript is cut by 70%. The model supports more than 85 languages, with automatic detection.
What It Delivers Is a Finished Draft
Google positions this model to turn raw audio directly into accurate, clean, already-formatted text. It does three specific things: it handles a speaker's self-corrections (the official example is "let's meet Tuesday — no, Wednesday," where the model keeps the final "Wednesday"); it strips filler words like "um" and "ah"; and it applies formatting automatically.
Traditional speech recognition models output a raw transcription stream, and this kind of cleanup has to happen downstream, handled by a separate model or a rules engine. Folding it into the transcription model itself saves an extra call and an extra round-trip of latency. For teams building real-time voice agents, that layer of latency is often what separates something that sounds like a real conversation from something that doesn't.
The model also supports function calling, letting it hand off complex tasks mid-transcription. It additionally offers custom vocabularies, speaker diarization (up to 3 speakers for pre-recorded audio, with anything beyond that experimental), word-level timestamps, and mid-call language switching.
Two APIs, Two Use Cases
The real-time model is called gemini-3.5-transcribe-live, accessed through the Live API and aimed at interactive voice applications. The pre-recorded model is gemini-3.5-transcribe, accessed through the Interactions API and aimed at meeting notes and call logs.
On the developer side, it enters public preview through the Gemini API (Google AI Studio) and Google Antigravity. On the enterprise side, it's in preview through the Gemini Enterprise Agent Platform, with integration into Gemini Enterprise for Customer Experience to follow. For everyday users, there are currently three entry points: the Gemini app on macOS (English only), Rambler on Android (in select countries and languages), with Chrome support coming later.
Two Open Questions Beyond the Numbers
Google hasn't published pricing. Speech transcription is a per-minute business, and if a 2.6% word error rate comes paired with an above-market price, adoption will slow down. This category has churned through generations quickly over the past two years, and customers are sensitive to switching costs — price often decides a purchase more than a few-tenths-of-a-percent gap in accuracy.
Speaker diarization is only guaranteed to be accurate for up to three people, and that limit shuts out a batch of high-frequency use cases: multi-person meetings, customer service quality checks, and podcast transcription routinely involve more than three speakers. By labeling anything above three as experimental, Google is effectively admitting this part isn't stable yet.
A rough sense of what that accuracy means in practice: a 2.6% word error rate works out to roughly 26 wrong words in a 1,000-word meeting recording (an estimate — actual distribution depends heavily on accent and background noise). For legal or medical scenarios that need verbatim quotes, that error rate still requires a human pass; for generating meeting summaries or feeding context to an agent, it's already good enough. The dividing line between the two kinds of needs isn't the model — it's how the downstream use consumes the text.
Voice has seen intense activity this year, from OpenAI shipping three speech models at once to companies across the board folding voice APIs into their agent stacks. The basis of competition is shifting from "how accurate is the recognition" to "how much effort does it save once it's wired into a workflow." Gemini 3.5 Transcribe, by pulling cleanup and formatting inside the model itself, is moving in the same direction.
Sources: Google's official blog, CocoLoop, Artificial Analysis, DeepMind model documentation; word error rate, latency improvement, and language count verified against the official release page — pricing not disclosed.