Google has started pushing its Gemini 3.5 Transcribe speech-to-text model out of the lab and into the apps hundreds of millions of people already use every day. The model, which replaced Google’s older Chirp 3 transcription engine, is landing in Gmail, Docs, Keep, Search Live, and Chrome’s voice typing feature, according to Google’s own product announcements. It already powers dictation in the Gemini app and on Pixel phones through a feature Google calls Rambler.
The technical specs (a 2.6% word error rate on non-streaming audio, support for more than 85 languages) already made headlines when the model shipped. What matters now, on August 30, 2026, is distribution. Google is betting that dictation only becomes useful once it works everywhere a person types, not just in a dedicated transcription app. That bet puts Gemini 3.5 Transcribe on a collision course with OpenAI’s Whisper, Otter.ai, and Nuance’s Dragon line, three products that have defined voice-to-text for the past decade. This piece looks at what actually changed, how the rollout compares to rivals on price and reach, and what it signals about where dictation software goes next.
What Gemini 3.5 Transcribe Fixes That Chirp 3 Didn’t
Dictation software has had one persistent, annoying flaw since the earliest versions of Dragon shipped in the 1990s: it writes down exactly what you said, filler words, false starts, and all. Say “let’s meet Tuesday, actually no, Wednesday” and a traditional transcription engine hands you back that entire sentence, corrections included, forcing you to go back and edit it yourself.
Gemini 3.5 Transcribe handles that correction in real time. Google says the model listens for self-corrections mid-sentence and outputs only the corrected intent, so “let’s meet Tuesday, actually no, Wednesday” becomes “let’s meet Wednesday” in the final text. It also strips filler words like “um” and “ah” automatically and formats text as a user speaks rather than leaving formatting for later. Google has described this behavior as the model tidying up while you talk instead of transcribing you verbatim. That distinction, formatted intent versus verbatim audio, is the core difference between Gemini 3.5 Transcribe and the engine it replaced.
The Rollout: Gmail, Docs, Keep, and Chrome Voice Typing
Chirp 3 was mostly a backend model, used inside specific transcription workflows and developer tools. Gemini 3.5 Transcribe is shipping as a front-facing feature across Google’s productivity stack. Google’s plans, per its own announcements, call for the model to power Search Live’s voice interactions, dictation inside Docs and Keep, and voice-to-text in Gmail. Chrome gets arguably the widest-reaching change: voice typing built on Gemini 3.5 Transcribe in any text field on the web, not just inside a Google app.
That last piece matters more than it sounds. A browser-level dictation layer means every web form, every comment box, every third-party web app gets Gemini-quality transcription without that app’s developers writing a line of integration code. It’s the same distribution logic Google used to spread spell-check and autofill: bake the feature into the browser and let adoption happen passively. Whether Chrome’s voice typing actually reaches that scale by the end of 2026 is still an open question, since Google has not published a full rollout timeline covering every market and language.
Pixel, Android, and the Gemini App: Where It Ships First
On mobile, Gemini 3.5 Transcribe already powers Rambler, Google’s dictation feature on Android devices including Pixel phones, and it runs inside the Gemini app itself. That’s a deliberate sequencing choice. Pixel and the Gemini app are first-party surfaces where Google controls the full stack, so bugs get caught and fixed before the feature reaches Gmail’s billion-plus user base or the open web through Chrome. Developers get access to the same underlying model through Google’s Studio and Antigravity platforms, letting third parties build custom vocabularies for jargon-heavy fields like law, medicine, or engineering before the consumer rollout is even complete.
Inside the Benchmark Numbers
General Word Error Rate
Word error rate, or WER, is the standard yardstick for transcription accuracy: the percentage of words a model gets wrong compared to a human-verified transcript. Lower is better. According to measurements published by benchmarking firm Artificial Analysis, Gemini 3.5 Transcribe scores a 4.0% WER in streaming mode (transcribing audio as it’s spoken) and 2.6% WER in non-streaming mode, where the model processes a complete recording after the fact. Non-streaming scores are typically better across the industry because the model has the full audio context available rather than reacting word by word.
FLEURS Multilingual Results
On the FLEURS benchmark, a multilingual speech dataset covering dozens of languages that’s widely used to test cross-language performance, Google reports 5.50% WER streaming and 5.04% WER non-streaming for Gemini 3.5 Transcribe. Google also says the model cuts time-to-final-transcription by 70% compared to its prior approach, a metric aimed less at accuracy and more at how fast a user sees finished, formatted text appear after they stop speaking.
| Metric | Streaming Mode | Non-Streaming Mode | Source |
|---|---|---|---|
| Word Error Rate (general) | 4.0% | 2.6% | Artificial Analysis |
| Word Error Rate (FLEURS multilingual) | 5.50% | 5.04% | |
| Languages supported | 85+ | ||
| Max speakers with timestamps | 3 | ||
| Time-to-final-transcription improvement | 70% faster | ||
How the Correction Logic Works, Conceptually
Google hasn’t published the production code behind Gemini 3.5 Transcribe’s self-correction handling, but the concept is straightforward enough to sketch out. The model has to hold a rolling window of recently spoken words, detect a correction signal (words like “actually,” “no wait,” or a repeated noun phrase), and then discard the superseded segment before emitting final text. A simplified, illustrative version of that logic looks like this:
function process_utterance(audio_stream):
window = rolling_buffer(seconds=8)
draft = speech_to_text(window)
if correction_marker_detected(draft):
superseded_span = find_superseded_segment(draft)
draft = remove_span(draft, superseded_span)
draft = strip_filler_words(draft)
draft = apply_live_formatting(draft)
return draftThis is not Google’s actual implementation, just a plain-language sketch of the problem the model has to solve. The hard part isn’t detecting the word “actually,” it’s deciding how much of the preceding sentence to throw away without also deleting content the speaker meant to keep.
Where Gemini 3.5 Transcribe Stands Against Whisper, Otter.ai, and Dragon
OpenAI’s Whisper large-v3 remains the most widely deployed open transcription model outside Google’s ecosystem, commonly cited as supporting around 99 languages. Independent benchmark roundups circulating in 2025 and 2026 put Whisper’s word error rate around 2.7% on clean, single-speaker audio, climbing toward 10% or higher on harder multilingual test sets, a wider spread than Google reports for Gemini 3.5 Transcribe on FLEURS. Whisper’s core model is also, by design, a batch transcriber rather than a native streaming one, though OpenAI’s API pricing page now lists a separate real-time model, gpt-realtime-whisper, priced at $0.017 per minute for live transcription, versus $0.006 per minute for standard Whisper transcription.
Otter.ai, built for meeting transcription rather than general dictation, doesn’t publish a formal WER figure on its pricing page, but it does publish tiers: a free Basic plan, Pro at $16.99 a month (or $8.33 a month billed annually), Business at $30 a month (or roughly $20 annually), and custom Enterprise pricing. Otter’s pitch has always been meeting-specific features like speaker labels and searchable transcripts rather than raw transcription speed, which is a different product category than what Google is now building directly into Chrome and Gmail.
Nuance’s Dragon line, now under Microsoft after Nuance’s acquisition, still dominates professional dictation in fields like law and medicine, where accuracy on specialized vocabulary matters more than raw speed. Dragon Professional v16 is commonly listed around $699 as a one-time purchase across reseller pages, while Dragon Medical One, aimed at clinicians, runs on a subscription model cited around $79 to $99 per user per month depending on contract length, according to a review published by health-tech firm Commure. Dragon doesn’t publish a public WER figure either, leaning instead on decades of specialized medical and legal vocabulary training as its competitive edge.
Pricing at a Glance
| Product | Pricing | Streaming | Positioning |
|---|---|---|---|
| Gemini 3.5 Transcribe | Free in Google apps; developer access via Studio/Antigravity | Yes (4.0% WER) | Consumer dictation, browser-wide via Chrome |
| OpenAI Whisper / gpt-realtime-whisper | $0.006/min (batch); $0.017/min (live) | Yes, via separate real-time model | Developer API, open-source base model |
| Otter.ai | Free to $30/mo (Business) | Yes, via RTMP for higher tiers | Meeting transcription and notes |
| Nuance Dragon Professional | ~$699 one-time (v16) | Live dictation | Professional desktop dictation |
| Nuance Dragon Medical One | ~$79-99/user/month | Streamed to cloud | Clinical documentation |
Why Bundling Beats Building a Standalone App
None of Gemini 3.5 Transcribe’s individual numbers, not the 2.6% WER, not the 85-plus languages, represent a dramatic leap over what the best transcription models already did in 2025. What’s different is distribution. Whisper needs a developer to integrate it. Otter.ai needs a user to open a separate app and remember to hit record. Dragon needs a purchase decision and, often, a specialized workflow built around it. Gemini 3.5 Transcribe needs none of that inside Google’s own products, because it’s already sitting behind the microphone icon a user has clicked a thousand times before in Gmail or Docs.
That’s the same playbook Google ran with Google Docs’ built-in grammar suggestions and Gboard’s predictive text: ship the feature inside software people already have open, and let the accuracy improvements arrive as invisible background updates rather than something users have to seek out or pay for separately. It’s a harder strategy for competitors to counter with a better model alone, since the fight isn’t purely about accuracy anymore. It’s about how many keyboards, browsers, and inboxes a company controls.
A Short History of the Dictation Problem
Dragon NaturallySpeaking launched in 1997 as the first commercially viable continuous-speech dictation product, and for nearly two decades it stayed the default answer to “how do I talk instead of type.” Apple and Google both shipped cloud-based dictation on smartphones in the early 2010s, trading Dragon’s local processing for server-side models that improved faster but required a data connection. OpenAI’s Whisper, released as an open-weight model in 2022, reset expectations again by making near-state-of-the-art transcription free for anyone to run, pushing WER figures for major languages down into the low single digits for the first time outside proprietary systems.
Every one of those jumps, though, still left users editing out “um,” fixing self-corrections by hand, and reformatting a wall of run-on text into paragraphs. Gemini 3.5 Transcribe is the first mainstream product from a major platform to treat that cleanup step as the model’s job rather than the user’s, and to ship it as a default rather than a premium add-on.
Multilingual Reach and Speaker Attribution
Google states Gemini 3.5 Transcribe supports more than 85 languages, and can attribute speech to as many as three separate speakers with timestamps in a single recording. That speaker-attribution cap is notably lower than what dedicated meeting-transcription tools like Otter.ai support, since Otter is built around multi-person calls with far more than three participants. It suggests Google is optimizing Gemini 3.5 Transcribe first for single-speaker dictation and small conversations rather than large meeting rooms, which lines up with its initial rollout into Gmail composition and Docs dictation rather than into Google Meet’s transcript feature.
Custom Vocabularies and the Developer Angle
Beyond the consumer rollout, Google is opening Gemini 3.5 Transcribe to developers through its Studio and Antigravity platforms, with support for custom vocabularies covering jargon and technical terms. That’s a direct pitch to the same specialized markets Dragon Medical and Dragon Legal have owned for years: a hospital or law firm could, in principle, feed the model a list of drug names or case terminology and get accuracy improvements on exactly the words a general-purpose model would otherwise flub. Whether Google can build the compliance, security, and integration relationships those industries expect from an incumbent like Nuance is a separate, harder problem than shipping an accurate model.
Market Impact: Who Feels This First
The most exposed players are transcription-focused startups and mid-market tools that compete on general dictation accuracy rather than a specialized niche. If Chrome-level voice typing works as advertised across enough websites, some fraction of casual dictation use cases (drafting an email, filling out a web form, taking a quick note) simply stop needing a separate app or extension at all. Otter.ai’s meeting-transcript niche is more insulated, since it depends on features like calendar integration and searchable meeting archives that a browser-level dictation layer doesn’t replicate. Dragon’s medical and legal markets are the most insulated of all, protected by compliance requirements, existing IT contracts, and specialized vocabulary training that a general consumer feature won’t match overnight.
For OpenAI, the pressure is different: Whisper’s real value has always been as an API developers build on top of, not a consumer feature competing for Gmail’s compose box. Google folding a comparable model directly into its own products doesn’t threaten Whisper’s developer business so much as it narrows the case for building a custom dictation feature from scratch when Chrome might just do it natively soon.
What Could Go Wrong
Automatic correction and filler-word removal sound like pure upside until a model guesses wrong about what a speaker meant to keep. If Gemini 3.5 Transcribe misreads a self-correction and deletes content a user actually wanted, that error is invisible at the moment of dictation, since the deleted words never appear in the final text for the user to catch. That’s a different failure mode than a simple misspelled word, which is easy to spot and fix. Google has not published error rates specific to its correction-detection feature separate from general WER, so how often that specific failure happens in practice remains untested by outside reviewers as of this writing.
There’s also a privacy dimension that scales with reach. A model available in every Chrome text field processes audio from a much wider range of contexts than a dedicated dictation app a user opens deliberately, from casual browsing to sensitive account forms. Google’s privacy documentation for Gemini features covers how voice data is processed and retained, but a browser-wide rollout multiplies the number of contexts where that policy actually applies in practice.
Predictions: Where Dictation Goes From Here
A few reasonable predictions follow from the shape of this rollout. First, expect Microsoft to respond by pushing its own Copilot-branded dictation deeper into Windows and Office by early 2027, using the same bundling logic Google is running now, since Microsoft already owns Dragon’s enterprise relationships and has every incentive to fold that accuracy into consumer Windows dictation too. Second, expect OpenAI to keep competing on the developer side rather than chase consumer bundling directly, since Whisper’s business model was never about owning a browser or an email client.
Third, watch for Otter.ai and similar meeting-transcription vendors to lean harder into features a browser-level model can’t replicate, like calendar-synced meeting archives and CRM integrations, rather than compete on raw transcription accuracy where the gap with free, built-in options keeps narrowing. Fourth, expect regulatory attention on always-listening or browser-embedded voice features to grow in the EU specifically, given how aggressively Brussels has already moved on AI transparency rules this year. Fifth, expect Google to expand Gemini 3.5 Transcribe’s speaker-attribution limit beyond three speakers within the next product cycle, once the model proves itself in the smaller-scale dictation contexts it’s launching in now.
The Bigger Picture for Voice Interfaces
Dictation has always been a bottleneck feature: useful, but rarely good enough that people preferred it to typing for anything beyond short notes or accessibility needs. Gemini 3.5 Transcribe’s bet is that removing the cleanup work, not just improving raw accuracy, is what finally tips that balance for a meaningful share of everyday writing. Whether that holds up depends less on the WER numbers Google has already published and more on whether the correction logic actually stays out of a user’s way across the messy, interrupted, multitasking way most people actually talk while typing something else at the same time.
Frequently Asked Questions
What is Gemini 3.5 Transcribe?
It’s Google’s current speech-to-text model, which replaced the earlier Chirp 3 engine. It converts spoken audio into formatted text, automatically removes filler words, and corrects self-corrections made mid-sentence.
Which Google apps use Gemini 3.5 Transcribe?
Google has said it powers dictation in the Gemini app and Rambler on Android and Pixel devices today, with planned expansion into Gmail, Docs, Keep, Search Live, and Chrome’s voice typing feature.
How accurate is Gemini 3.5 Transcribe compared to Whisper?
Artificial Analysis measured Gemini 3.5 Transcribe at 2.6% WER non-streaming and 4.0% WER streaming. Independent roundups put OpenAI’s Whisper large-v3 around 2.7% WER on clean audio, rising toward 10% or higher on harder multilingual sets, though methodologies differ enough between tests that a direct one-to-one comparison should be read cautiously.
How many languages does Gemini 3.5 Transcribe support?
Google states more than 85 languages, according to its own product announcements.
Does Gemini 3.5 Transcribe work offline?
Google has not published details confirming offline support for Gemini 3.5 Transcribe. Its design, including cloud-based custom vocabulary and formatting features, suggests it is primarily a connected, server-processed model.
Is Gemini 3.5 Transcribe free to use?
Inside Google’s own apps like Gmail, Docs, and the Gemini app, it’s included at no separate charge. Developers accessing the underlying model through Studio and Antigravity work under Google’s standard API terms rather than a free consumer tier.
How many speakers can Gemini 3.5 Transcribe identify?
Google says the model can attribute speech to up to three separate speakers with timestamps, a smaller cap than dedicated meeting tools like Otter.ai support.
What happened to Chirp 3?
Chirp 3 was Google’s prior transcription engine. Gemini 3.5 Transcribe replaces it, according to Google’s own announcements, with improved accuracy and lower latency.
