Google shipped a new speech-to-text model on August 26, 2026, and by the numbers it just quietly reset the bar for transcription accuracy. Gemini 3.5 Transcribe lands with an average word error rate of 2.6% for non-streaming audio, support for more than 85 languages, and a feature set that pulls diarization, word-level timestamps, and custom vocabulary biasing into a single API call. Google announced the release in a blog post titled “Intelligent transcription with Gemini 3.5 Transcribe,” and the model is already rolling into the Gemini API, Google AI Studio, and a string of consumer products.
This is not a marginal spec bump. Speech-to-text has quietly become one of the most contested corners of applied AI, feeding everything from call-center analytics to meeting notes to voice assistants. Google’s timing, less than two weeks after it was still previewing Gemini 3.7 and 3.8 Flash, signals a company pushing hard on every input and output modality at once. Below, we break down what shipped, what it means for competitors, and where the speech AI market goes from here.
What Google Actually Announced
Google’s own blog post frames Gemini 3.5 Transcribe as “our most precise speech-to-text model yet, designed for intelligent voice interactions.” The release covers two distinct models rather than one: gemini-3.5-transcribe, built for non-streaming, pre-recorded audio such as meeting recordings and call logs, and gemini-3.5-transcribe-live, a streaming variant that runs over WebSockets through the Gemini Live API.
That split matters. Non-streaming models can look at the whole audio clip before committing to an output, which is why the non-streaming variant posts the lower error rate. The live model has to commit to text in near real time, which is a harder problem and shows up in its higher, though still competitive, error rate. Google’s developer documentation describes Gemini 3.5 Transcribe as a speech-to-text model based on Gemini’s underlying audio understanding capabilities, rather than a bolted-on legacy ASR (automatic speech recognition) engine.
Google is not shy about the ambition here. On X, the official Google account posted: “We’re introducing Gemini 3.5 Transcribe, our most precise speech-to-text model yet” (source). The Google for Developers account echoed the same line almost verbatim: “Gemini 3.5 Transcribe is our most precise speech-to-text model yet, designed for intelligent voice interactions” (source). Repetition across official channels usually signals a coordinated launch rather than a quiet API update, and this one hit blog, docs, and social simultaneously.
The Word Error Rate Numbers, Explained
Word error rate (WER) is the standard yardstick for transcription quality. It counts substitutions, deletions, and insertions against a reference transcript, then divides by total words. Lower is better, and gains below a full percentage point are usually treated as meaningful at this end of the scale.
Google reports an average WER of 4.0% for the streaming model and 2.6% for the non-streaming model, with both figures measured by Artificial Analysis, an independent AI benchmarking firm. That third-party measurement matters. Vendors self-reporting benchmarks on their own launch day is common; having an outside evaluator attached to the number gives it more weight with engineering teams deciding whether to switch providers.
It is worth being precise about what is and is not confirmed here. Google and Artificial Analysis have confirmed the 2.6% non-streaming WER figure and the 85+ language support separately. The sources reviewed for this article do not explicitly state that the 2.6% figure was measured across all 85+ languages as a single blended average versus a narrower benchmark set, so treat that combination as directionally accurate rather than a single certified statistic.
Feature Breakdown: Diarization, Timestamps, and Vocabulary Biasing
The headline WER number gets the clicks, but the feature list is what determines whether engineering teams actually migrate. Gemini 3.5 Transcribe ships with speaker diarization, meaning the model can separate and label distinct speakers in an audio file, an essential feature for call centers, podcasts, and multi-party meetings. Word-level timestamps are also included, letting developers align transcript text to exact audio positions for captioning or search.
Custom vocabulary biasing, meanwhile, lets developers feed the model domain-specific terms, product names, or jargon that a general-purpose model would otherwise mangle. Google’s documentation specifies support for up to 1,000 custom terms per request. That is a genuinely useful ceiling for enterprise deployments dealing with medical terminology, legal names, or internal product codenames that a stock model has never seen.
Language coverage rounds out the pitch. Google says the model supports more than 85 languages with utterance-based language detection, meaning it can identify which language is being spoken at the sentence or phrase level rather than requiring the developer to hardcode a language setting upfront. For global products with unpredictable user bases, that removes a real integration headache.
| Feature | Gemini 3.5 Transcribe (Non-Streaming) | Gemini 3.5 Transcribe Live (Streaming) |
|---|---|---|
| Reported average WER | 2.6% | 4.0% |
| Language support | 85+ languages | 85+ languages |
| Speaker diarization | Supported | Not supported |
| Word-level timestamps | Supported | Not supported |
| Custom vocabulary biasing | Up to 1,000 terms | Up to 1,000 terms |
| Primary use case | Recorded audio, meetings, call logs | Live voice interactions, assistants |
| Access method | Gemini API | Gemini Live API (WebSockets) |
The asymmetry between the two models is deliberate. Google’s transcription documentation confirms diarization and timestamps are reserved for the non-streaming model, likely because both features require the kind of full-context analysis that a live, low-latency pipeline cannot afford. That is a normal engineering tradeoff, not a limitation unique to Google, and it shows up in every competing streaming ASR product as well.
Why This Launch Comes Now
Google has been releasing Gemini variants at a pace that would have seemed reckless two years ago. Shattered.io covered Gemini 3.8 Flash entering preview testing just 14 days after Gemini 3.7 shipped. Gemini 3.5 Transcribe lands into that same cadence, a company iterating across text, multimodal, and now audio-specific models on overlapping timelines rather than a single yearly flagship release.
Speech-to-text is also becoming infrastructure for a much bigger bet: voice-first AI assistants. A transcription model with low WER and native diarization is the input layer for every voice agent, meeting summarizer, and dictation tool Google plans to ship next. Getting the transcription layer right first is a prerequisite, not an afterthought, and it explains why Google positioned the launch as feeding directly into the Gemini app, Gemini Enterprise, and products like Docs, Keep, Gmail, and Gboard rather than as a standalone research artifact.
The enterprise angle deserves attention too. Google is putting Gemini 3.5 Transcribe into public preview via the Gemini Enterprise Agent Platform, with a stated plan to extend it to Gemini Enterprise for Customer Experience. Call-center transcription and analytics is a market where incumbents like Deepgram, AssemblyAI, and Microsoft’s Azure AI Speech have built entire businesses. Google entering with a model that has an independently measured sub-3% WER on recorded audio is a direct shot at that segment.
Competitive Landscape: Where Gemini 3.5 Transcribe Fits
The speech-to-text market was never short on serious players. OpenAI’s Whisper remains the most widely deployed open-source ASR model, and OpenAI also offers hosted transcription through its API platform. ElevenLabs built a business around high-fidelity voice AI and now offers its own speech-to-text product. Deepgram and AssemblyAI both specialize in developer-first transcription APIs with strong enterprise footholds, and Microsoft bundles transcription into Azure AI Speech as part of its broader cloud AI services suite.
None of the coverage reviewed for this article includes a direct, apples-to-apples benchmark pitting Gemini 3.5 Transcribe against Whisper, Deepgram, AssemblyAI, or Azure Speech on the same test set. That comparison will matter enormously once independent evaluators run it, and it is the single most important open question hanging over this launch. Until then, the safest claim is a structural one: Google now has a model with a third-party-verified sub-3% WER on recorded audio, entering a market where several rivals have not published comparably specific, externally-measured numbers.
What Google does have that most rivals cannot easily replicate is distribution. Whisper, Deepgram, and AssemblyAI all need customers to integrate an API. Google can push Gemini 3.5 Transcribe directly into Search Live, Gemini Live, Docs, Keep, Gmail, and Gboard, reaching users who never touch a developer console. That distribution advantage, not the WER number alone, is the more durable competitive edge.
| Provider | Model / Product | Type | Notable Distribution |
|---|---|---|---|
| Gemini 3.5 Transcribe / Live | Non-streaming + streaming | Gemini API, AI Studio, Gemini Enterprise, Docs, Gmail, Gboard | |
| OpenAI | Whisper / API transcription | Open-source + hosted API | Developer API, open-source community deployments |
| ElevenLabs | Speech-to-Text | Hosted API | Voice AI developer ecosystem |
| Deepgram | Nova family | Hosted API | Enterprise call-center and contact-center integrations |
| AssemblyAI | Universal / Slam models | Hosted API | Developer-first API, media and podcast platforms |
| Microsoft | Azure AI Speech | Cloud service | Azure ecosystem, Microsoft 365 integrations |
Availability: Who Gets Access and When
Developers can access Gemini 3.5 Transcribe now through the Gemini API in Google AI Studio, according to Google’s own release materials. The model is also available through Antigravity, Google’s developer tooling brand for agentic workflows. For enterprise customers, the model is in public preview through the Gemini Enterprise Agent Platform, with Gemini Enterprise for Customer Experience listed as a coming-soon integration rather than a day-one feature.
Consumer rollout is happening in parallel. Google says the model is reaching the Gemini app on macOS and the Rambler app on Android, and reports from Android Authority note it is spreading across Search Live, Gemini Live, Docs, Keep, Gmail, the Gemini app, and Gboard. That is a wide surface area for a single model launch, and it suggests Google is treating transcription quality as foundational infrastructure rather than a niche feature for one product line.
Pricing has not been confirmed in any source reviewed for this article. Google has not published per-minute or per-request pricing for Gemini 3.5 Transcribe as of this writing, which is typical for a same-day launch where usage-based pricing tiers often follow in subsequent developer documentation updates.
Historical Context: Google’s Speech AI Trajectory
Google’s investment in speech recognition predates the current generative AI boom by well over a decade, running through Google Voice Search, the original Cloud Speech-to-Text API, and the speech components embedded in Android’s on-device assistant stack. What changed with the Gemini era is the shift from purpose-built ASR architectures to a single multimodal foundation model handling audio as just another input type alongside text and images.
That shift has tradeoffs. A dedicated ASR model can, in theory, be smaller and cheaper to run than a general multimodal model repurposed for transcription. Google’s answer with Gemini 3.5 Transcribe is to split the difference: keep the Gemini audio-understanding backbone that gives it contextual smarts (handling self-corrections, stripping filler words, formatting output based on apparent intent) while packaging it as a dedicated, purpose-named product rather than asking developers to prompt a general chat model into transcription mode.
The “smart transcription” framing is a genuine departure from older dictation tools. Google’s product materials describe the model handling self-corrections automatically, stripping filler words, and formatting output based on the speaker’s apparent intent rather than transcribing every syllable literally. That is a materially different product philosophy from a raw ASR pipeline that transcribes every “um” and false start verbatim, and it points toward transcription increasingly being judged on readability, not just literal accuracy.
Market Impact: What Sub-3% WER Means for Enterprise Buyers
For enterprise buyers evaluating transcription vendors, WER differences in the low single digits translate directly into downstream costs. A call center processing tens of thousands of hours of audio a month sees compounding value from every fraction of a percentage point improvement, because errors propagate into search, analytics, and compliance workflows built on top of the transcript. A model that misses one word in forty versus one word in twenty-five is a meaningfully different foundation for automated QA or sentiment analysis.
The custom vocabulary ceiling of 1,000 terms is also a practical enterprise signal. Domains like healthcare, legal, and financial services depend heavily on proper nouns and technical jargon that general models routinely mistranscribe. A higher term ceiling reduces the manual post-processing and correction work that has historically made transcription tools frustrating for specialized industries.
There is also a procurement dimension. Enterprises already standardized on Google Cloud or Workspace get a lower integration cost by adopting Gemini 3.5 Transcribe compared to bolting on a third-party vendor, simply because billing, identity, and data governance are already unified. That bundling dynamic has played out before in cloud infrastructure and is likely to repeat in AI services, where the biggest threat to standalone speech AI vendors is not a worse product but a more convenient one.
What Developers Are Saying
Reaction from Google’s own developer relations channels has been consistent in emphasizing the same core feature set: smart transcription, function calling, lower WER, custom vocabulary, multi-speaker identification, and language breadth. Google’s official developer account posted that Gemini 3.5 Transcribe “is our most precise speech-to-text model yet, designed for intelligent voice interactions” (source), a line that has now been repeated across the company’s blog, documentation, and social channels.
Google’s official blog post itself states: “Today, we’re introducing Gemini 3.5 Transcribe, our most precise speech-to-text model yet, designed for intelligent voice interactions” (source). The company’s developer documentation frames the release in more technical terms: “Gemini 3.5 Transcribe is a speech-to-text model based on Gemini’s audio understanding capabilities” (source).
On the transcription pipeline specifically, Google’s documentation notes: “The Gemini API converts speech in audio files into text using the Gemini 3.5 Transcribe model” (source). Taken together, these statements describe a company treating transcription as a core, named capability of the Gemini platform rather than a side feature, which tracks with how aggressively it is being distributed across Google’s consumer and enterprise product lines.
Technical Access: A Quick Look at the API
For developers evaluating the model, Google’s documentation describes a straightforward request pattern through the Gemini API for the non-streaming model, and a WebSocket connection through the Gemini Live API for the streaming variant. A simplified request for the non-streaming model looks like this:
from google import genai
client = genai.Client()
response = client.models.generate_content(
model="gemini-3.5-transcribe",
contents=[
{"audio_url": "path/to/audio-file.wav"},
],
config={
"diarization": True,
"word_timestamps": True,
"vocabulary_terms": ["Kubernetes", "shattered.io", "Gemini"]
}
)
print(response.text)This example is illustrative of the documented capabilities (diarization, timestamps, and vocabulary biasing as configurable options) rather than a verbatim code sample from Google’s docs. Developers should consult the official Gemini API transcription documentation for exact syntax before building against it in production.
Risks and Open Questions
A few things remain genuinely unresolved. Pricing is not public yet, which matters enormously for high-volume use cases like call-center transcription where cost per minute can make or break a vendor decision. Independent, head-to-head benchmarks against Whisper, Deepgram, AssemblyAI, and Azure Speech using identical audio sets have not been published, so buyers should treat Google’s WER figures as directionally strong rather than as a settled competitive ranking.
There is also the multi-speaker diarization caveat buried in the documentation: attribution accuracy for three or more simultaneous speakers is marked experimental, capped in practice around eight speakers. That is a reasonable engineering limit, but it means large group calls or panel-style recordings may still need manual correction, undercutting some of the “fully automated” framing around the launch.
Data governance is the other open thread. Enterprises adopting Gemini 3.5 Transcribe for regulated audio, medical dictation, legal depositions, financial call recordings, will want clarity on data retention, training-data usage, and compliance certifications before committing production workloads. None of the sources reviewed for this piece addressed those specifics, and it is the kind of detail that typically surfaces in follow-up documentation rather than a launch blog post.
Five Predictions for the Next Six Months
- Independent benchmarks arrive within weeks. Given Artificial Analysis already measured the initial WER figures, expect a fuller head-to-head comparison against Whisper, Deepgram, and AssemblyAI to follow within one to two months of launch.
- Pricing lands as a usage-tiered model. Google typically publishes per-minute or per-token pricing for new Gemini capabilities within weeks of a preview launch, following the pattern set by prior Gemini API releases.
- Call-center and contact-center vendors respond with bundling deals. Expect Deepgram, AssemblyAI, and similar API-first vendors to emphasize accuracy-per-dollar and specialized industry tuning as their competitive counter, rather than trying to out-WER Google directly.
- Diarization for 3+ speakers matures out of experimental status. Multi-speaker attribution is the most obvious feature gap flagged in Google’s own documentation, and it is a likely candidate for the next incremental model update.
- Voice-first product features accelerate across Google Workspace. With transcription infrastructure now unified under one model family, expect faster rollout of meeting summarization, live captioning, and voice search improvements across Docs, Meet, and Gmail through the rest of 2026.
Frequently Asked Questions
What is Gemini 3.5 Transcribe?
Gemini 3.5 Transcribe is a speech-to-text model Google announced on August 26, 2026, built on Gemini’s audio understanding capabilities. It converts pre-recorded audio into text with reported low word error rates, speaker diarization, and word-level timestamps.
What is the difference between Gemini 3.5 Transcribe and Gemini 3.5 Transcribe Live?
Gemini 3.5 Transcribe handles non-streaming, pre-recorded audio and supports diarization and word-level timestamps. Gemini 3.5 Transcribe Live is a streaming model for real-time voice interactions delivered over WebSockets, but it does not support diarization or word-level timestamps.
How accurate is Gemini 3.5 Transcribe?
Google reports an average word error rate of 2.6% for the non-streaming model and 4.0% for the streaming Live model, with figures measured by the independent benchmarking firm Artificial Analysis.
How many languages does Gemini 3.5 Transcribe support?
Google states the model supports more than 85 languages, with automatic, utterance-based language detection that can identify language switches within a single audio file.
Does Gemini 3.5 Transcribe support speaker diarization?
Yes, for the non-streaming model, supporting up to eight speakers. Attribution accuracy for three or more simultaneous speakers is marked experimental in Google’s documentation. The streaming Live model does not support diarization.
How does Gemini 3.5 Transcribe compare to OpenAI Whisper?
No independent, head-to-head benchmark comparing Gemini 3.5 Transcribe directly against OpenAI Whisper on identical audio has been published as of this writing. Google’s reported 2.6% non-streaming WER figure is independently measured by Artificial Analysis, but a direct competitive comparison remains an open question.
Where can developers access Gemini 3.5 Transcribe?
Developers can access it through the Gemini API in Google AI Studio and through Antigravity. Enterprise customers can access it in public preview through the Gemini Enterprise Agent Platform.
Is Gemini 3.5 Transcribe available for free?
Google has not published pricing for Gemini 3.5 Transcribe as of this writing. Usage-based pricing typically follows in developer documentation updates after a preview launch.
- Google Tests Gemini 3.8 Flash 14 Days After 3.7 [2026]
- Nvidia Reportedly Buys Hugging Face for $12.9B [2026]
- Jalapeño Chip: OpenAI Targets Nvidia’s 75% Margin [2026]
- HBM4 Memory Hits 80% Yield, Powers Nvidia Rubin [2026]
- FinOps Shifts to AI: 98% Track Spend, 73% Blow Budget [2026]
- More AI & Machine Learning coverage
