--- title: Google launches Gemini 3.5 Transcribe with sub-second... description: Google has released Gemini 3.5 Transcribe, a new speech-to-text model in public preview via the Gemini API and Google AI Studio, with two variants: a streaming... canonical: https://daily.dev/posts/google-launches-gemini-3-5-transcribe-with-sub-second-streaming-and-85-language-support-cef0b9qnr twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Google launches Gemini 3.5 Transcribe with sub-second streaming and 85+ language support | daily.dev og:description: Google has released Gemini 3.5 Transcribe, a new speech-to-text model in public preview via the Gemini API and Google AI Studio, with two variants: a streaming... og:url: https://daily.dev/posts/google-launches-gemini-3-5-transcribe-with-sub-second-streaming-and-85-language-support-cef0b9qnr og:image: https://api.daily.dev/og/posts/CEF0B9qNr.png og:image:alt: Google launches Gemini 3.5 Transcribe with sub-second streaming and 85+ language support og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Google launches Gemini 3.5 Transcribe with sub-second streaming and 85+ language support **[Collections](https://daily.dev/sources/collections)** · 2 min read · 0 upvotes · 0 comments ## Summary Google has released Gemini 3.5 Transcribe, a new speech-to-text model in public preview via the Gemini API and Google AI Studio, with two variants: a streaming model (gemini-3.5-transcribe-live) offering sub-second bidirectional transcription, and a batch model (gemini-3.5-transcribe) with word-level timestamps and speaker attribution for up to three speakers. Compared to Chirp 3, it achieves 2.6% word error rate on non-streaming audio, 4.0% on streaming, a 70% cut in time-to-final-transcription, and auto-detects 85+ languages. Notable post-processing features include filler-word removal, self-correction handling, and contextual alphanumeric interpretation (e.g., recognizing spoken ## Content Google has released Gemini 3.5 Transcribe, its first Gemini-based speech-to-text model, now in public preview via Google AI Studio and the Gemini API. ## What it does The model comes in two variants: - **gemini-3.5-transcribe** — processes pre-recorded audio up to one hour long, with word-level timestamps, speaker attribution for up to three speakers, and custom vocabulary support - **gemini-3.5-transcribe-live** — real-time WebSocket streaming with sub-second latency; individual sessions cap at 10 minutes Being LLM-based rather than a traditional acoustic model, it handles the stuff that trips up older transcription systems: alphanumeric strings, phone numbers, postal codes, email addresses. One engineer noted it correctly writes `.json` rather than transcribing it as the name "Jason." It also strips filler words like "um" and "ah," handles mid-sentence self-corrections, and auto-detects over 85 languages — even when language hints are set to English only. ## Performance Compared to Chirp 3, Google's previous transcription model, Gemini 3.5 Transcribe achieves: - 2.6% word error rate (non-streaming) - 4.0% word error rate (streaming) - 70% reduction in time-to-final-transcription It also outperforms prior models on the FLEURS benchmark in streaming mode. ## Where it's showing up The model already powers Gboard's Rambler feature and is integrated into the Gemini app on macOS. It's coming to Chrome, and Google Antigravity uses it for voice workflows across devices. Third-party platforms including LiveKit, Pipecat, and Vercel have already integrated it — Vercel's AI Gateway supports it with automatic language detection. ## Building with it Developers can access both endpoints through the Gemini API today. Google has published a demo showing a live speech translation app built with the Live API, LiveKit for WebRTC streaming, and Cloud Run for deployment. The demo handles roughly 15-20 simultaneous languages on a single instance before hitting CPU limits; moving session state to a database would allow multi-instance scaling. Custom vocabulary and language hints are supported to improve accuracy for domain-specific terms and names. ## Questions this post answers ### What is the word error rate of Gemini 3.5 Transcribe compared to Chirp 3? Gemini 3.5 Transcribe achieves a 2.6% word error rate in non-streaming mode and 4.0% in streaming mode, both improvements over Google's previous Chirp 3 transcription model. It also delivers a 70% reduction in time-to-final-transcription and outperforms prior models on the FLEURS benchmark in streaming mode. _Teams choosing a transcription API can track model benchmarks like these on daily.dev._ ### What's the difference between gemini-3.5-transcribe and gemini-3.5-transcribe-live? gemini-3.5-transcribe processes pre-recorded audio up to one hour long with word-level timestamps, speaker attribution for up to three speakers, and custom vocabulary support. gemini-3.5-transcribe-live provides real-time WebSocket streaming with sub-second latency, though individual sessions are capped at 10 minutes. _Developers picking between batch and streaming transcription APIs can follow updates like this on daily.dev._ ### How many languages can a single Cloud Run instance handle for live speech translation with Gemini 3.5 Transcribe? A demo built with the Live API, LiveKit for WebRTC streaming, and Cloud Run for deployment handles roughly 15-20 simultaneous languages on a single instance before hitting CPU limits. Moving session state to a database would allow scaling across multiple instances to support more languages concurrently. _Engineers architecting multilingual voice apps can find scaling details like this on daily.dev._ ## Community take How the wider developer community reacted, aggregated from 3 discussions and 10 comments across x (as of 2026-09-13). **TL;DR:** Reactions focus on the automatic cleanup of filler words and false starts as the standout feature, while others raise privacy and data-fidelity concerns, and some question whether this is really a leap over existing tools. **Sentiment:** 30% positive · 30% mixed · 40% skeptical **The case for** - The cleanup of false starts, filler words, and formatting is seen as genuinely useful and well executed. - Accurate multilingual transcription is viewed as strong enough to pressure dedicated speech-to-text startups. **The pushback** - Cleaning up speech means the model infers intended meaning rather than preserving an auditable record of what was actually said. - Concerns that offering the service lets a large company collect voice prints and richer personal data profiles. - Some argue similar capability already existed elsewhere years ago, questioning the novelty. **By community** - x (mixed): A handful of substantive replies mix praise for the automatic transcript cleanup with concerns about data privacy, auditability, and whether it's truly novel. **Hottest debate:** Whether the model's automatic cleanup of speech is a helpful feature or a loss of an accurate, auditable record. **Open questions** - Will the model run locally rather than requiring cloud processing? - How does its accuracy and cost actually compare to established speech-to-text competitors? **Highlights** > @rohanpaul_ai It's not a transcriber, it's a compiler from intent to text. Every cleanup - filler, number format, vocab - is an inference you can't audit. You get what the model thinks you meant, not proof of what you said. The smoother the output, the less the record is really yours. > — [0xnxzt\_ on x](https://x.com/0xnxzt_/status/2092736881326694711) > @rohanpaul_ai Offering services like this helps "Govt Google" because they get both our data and a verified voice print. This provides a complete profile with a photo, voice, and psych profile from our harvested data. > — [HowardAulsbrook on x · 1 points](https://x.com/HowardAulsbrook/status/2092773901310824671) > @rohanpaul_ai Multilingual transcription getting this accurate could pressure dedicated speech startups hard > — [Awesome\_O\_AI on x · 1 points](https://x.com/Awesome_O_AI/status/2092746048611692969) > @rohanpaul_ai so Gemini now can do what whsiper in chatgpt could do 3 years ago. > — [ereoeoeo on x](https://x.com/ereoeoeo/status/2092853839225319513) > @rohanpaul_ai the part that gets me is the cleanup 🥲 false starts, filler words, dates, numbers... all handled without making you sound like you dictated a ransom note > — [HiddenSignalsHQ on x](https://x.com/HiddenSignalsHQ/status/2092729197785690170) **Source threads** - [x](https://x.com/business_today/status/2092867472235888828) · 2 points · 0 comments - [x](https://x.com/zvawda/status/2092660990533836880) · 0 points · 0 comments - [x](https://x.com/rohanpaul_ai/status/2092723656388677752) · 0 points · 10 comments --- Tags: [#google](https://daily.dev/tags/google), [#google-gemini](https://daily.dev/tags/google-gemini), [#speech-recognition](https://daily.dev/tags/speech-recognition) [View this post on daily.dev](https://daily.dev/posts/google-launches-gemini-3-5-transcribe-with-sub-second-streaming-and-85-language-support-cef0b9qnr) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Google launches Gemini 3.5 Transcribe with sub-second streaming and 85+ language support","url":"https://daily.dev/posts/google-launches-gemini-3-5-transcribe-with-sub-second-streaming-and-85-language-support-cef0b9qnr","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/google-launches-gemini-3-5-transcribe-with-sub-second-streaming-and-85-language-support-cef0b9qnr"},"datePublished":"2026-08-26T17:21:58.109Z","dateModified":"2026-09-13T19:30:39.721Z","description":"Google has released Gemini 3.5 Transcribe, a new speech-to-text model in public preview via the Gemini API and Google AI Studio, with two variants: a streaming...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/57ca82980b7b1524508eaba413ca9a16?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/57ca82980b7b1524508eaba413ca9a16?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/google-launches-gemini-3-5-transcribe-with-sub-second-streaming-and-85-language-support-cef0b9qnr","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"google,google-gemini,speech-recognition","timeRequired":"PT2M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"Google launches Gemini 3.5 Transcribe with sub-second streaming and 85+ language support"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/google-launches-gemini-3-5-transcribe-with-sub-second-streaming-and-85-language-support-cef0b9qnr#faq","mainEntity":[{"@type":"Question","name":"What is the word error rate of Gemini 3.5 Transcribe compared to Chirp 3?","acceptedAnswer":{"@type":"Answer","text":"Gemini 3.5 Transcribe achieves a 2.6% word error rate in non-streaming mode and 4.0% in streaming mode, both improvements over Google's previous Chirp 3 transcription model. It also delivers a 70% reduction in time-to-final-transcription and outperforms prior models on the FLEURS benchmark in streaming mode. Teams choosing a transcription API can track model benchmarks like these on daily.dev."}},{"@type":"Question","name":"What's the difference between gemini-3.5-transcribe and gemini-3.5-transcribe-live?","acceptedAnswer":{"@type":"Answer","text":"gemini-3.5-transcribe processes pre-recorded audio up to one hour long with word-level timestamps, speaker attribution for up to three speakers, and custom vocabulary support. gemini-3.5-transcribe-live provides real-time WebSocket streaming with sub-second latency, though individual sessions are capped at 10 minutes. Developers picking between batch and streaming transcription APIs can follow updates like this on daily.dev."}},{"@type":"Question","name":"How many languages can a single Cloud Run instance handle for live speech translation with Gemini 3.5 Transcribe?","acceptedAnswer":{"@type":"Answer","text":"A demo built with the Live API, LiveKit for WebRTC streaming, and Cloud Run for deployment handles roughly 15-20 simultaneous languages on a single instance before hitting CPU limits. Moving session state to a database would allow scaling across multiple instances to support more languages concurrently. Engineers architecting multilingual voice apps can find scaling details like this on daily.dev."}}]} ```