What Gemini 3.5 Transcribe is, quickly
Until now, transcribing audio with Google meant one of two things: Google Cloud's Speech-to-Text API (the Chirp line), or handing an audio file to a general Gemini chat model and asking nicely. Gemini 3.5 Transcribe, announced around August 26, 2026, is the first model Google built specifically for the job, and it's positioned as the successor to Chirp 3.
Google's model page describes it as "a speech-to-text model based on Gemini's audio understanding capabilities" with language detection, speaker diarization, word-level timestamps, Smart transcription, and custom vocabulary biasing. It's really two endpoints: gemini-3.5-transcribe for pre-recorded files and gemini-3.5-transcribe-live for real-time streaming. I dug into the full pricing and spec breakdown separately; this review is about whether it's actually good.
Here's the model page itself, which is where most developers will start:
One thing to flag before you build on it: Google's own docs can't agree on its status. The models overview page tags it "New Stable," while the launch blog calls it "in public preview." It's already live in consumer products (Rambler on Gboard, the Gemini app on macOS), so treat it as production-adjacent but not fully settled.
How I judged it
I didn't have months of production logs to draw on, so I'll be straight about the basis for this review: I read the model card and pricing docs in full, checked Google's benchmark claims against how they're framed, and leaned heavily on the developers who were A/B testing it against Whisper, Deepgram, and ElevenLabs Scribe in the days after launch. Transcription is one of those areas where the vendor's clean-benchmark number and your real-world result can be worlds apart, so the community signal matters as much as the spec sheet.
The short version of what that turned up: the numbers are real and good, the pricing is very competitive, and the loudest complaints all cluster around two specific things. Let's take them in order.
Accuracy: the lab numbers are great, real audio is the test
On paper, this model is very good. Google reports the word error rates below, all attributed to Artificial Analysis:
A 2.6% word error rate on non-streaming English is roughly one wrong word in forty, which is excellent. Google also claims time to final transcription improves by 70% versus Chirp 3, though that framing is Google's own, so I'd file it as a vendor figure rather than an independent benchmark.
The more useful signal is what people found running it on their own audio. And here it splits. On the strong-positive side, this was about as glowing as review feedback gets:
"I am using it all day as a main dictation model, and I can say it's the best model in terms of accuracy, latency, and formatting I have ever used."
But the launch thread is full of people running comparisons and landing somewhere cooler. One who does exactly this for a living moved back to a hybrid setup:
"Curious if this suffers from the same hallucination problems as Chirp or not... we've moved back to Whisper for timestamp accuracy and then Gemini Flash for transcription correction."
My read: the WER on a clean benchmark and the WER on your actual audio (noisy calls, multiple accents, cross-talk) are different animals. The lab numbers are a strong starting point, not a promise. If you're evaluating this for real work, run it against your own recordings before you commit a pipeline to it.
The Smart transcription catch
This is the part of the review I'd underline. Smart transcription is the feature Google leans on hardest, and it's also the single most common complaint. A raw transcript of human speech is messy, people say "um," restart sentences, mumble numbers. Smart transcription cleans all of that up automatically.
Per the launch post, it handles self-corrections ("let's meet Tuesday, no, Wednesday" becomes "Wednesday"), strips filler words, and formats alphanumeric strings intelligently. For dictating an email or tidying meeting notes, that's the kind of thing standalone dictation tools charge extra for, so getting it in the base model is a genuine plus.
The problem is that "cleaning up" is a judgment call, and the model doesn't always share yours. One developer testing it on a Pixel 11 Pro put it bluntly:
"The main issue is when you want to say something precise with specific wording it might 'simplify' it and break the meaning. Something like 'I hesitated to check it, I should have verified' => 'I should have verified'... I tried the sentence few times and it always removed the first part."
So the mental model to carry into a purchase decision is this: Smart transcription is a feature, not a neutral transcription. For anything where the exact words matter (legal, research, compliance, medical), that's a real trade-off to weigh, not a footnote. It's the one thing about this model I'd want to test hardest on my own audio before trusting it.
Diarization stalls past three speakers
The second recurring complaint is multi-speaker handling, and if you do meeting transcription, this is the one to pay attention to. Google's docs list two different caps: the model page says the file model supports "up to 8 speakers (Attribution for 3+ speakers is experimental)," while the launch blog says "up to three." The reliable ceiling today is three, and I'd plan around that number. The live streaming model has no diarization at all, since it prioritizes low latency.
That gap showed up loudly in testing:
"Still no real time diarization beyond 3 people (and even then experimental) when others do it very well, like Soniox and Deepgram. For something like meeting notes this is critical."
If your use case is a two-person support call or a single-speaker voice note, this won't touch you. If it's a six-person standup, it's a real limitation, and it's the kind of thing a spec sheet lets you overlook until you're three weeks into a build.
Pricing: this is where it wins
Google didn't put pricing in the model announcement, which annoyed developers enough that one just posted them to Hacker News himself. Here's the real picture from the pricing page:
| Model | Audio input | Text output | Blended rate |
|---|---|---|---|
gemini-3.5-transcribe (file) | $2.00 / 1M tokens (~$0.003/min) | $12.00 / 1M tokens (~$0.002/min) | ~$0.005 / min |
gemini-3.5-transcribe-live (streaming) | $3.50 / 1M tokens (~$0.005/min) | $21.00 / 1M tokens (~$0.004/min) | ~$0.009 / min |
For comparison, Google's own Cloud Speech-to-Text v2 standard recognition starts at $0.016/min, while its dynamic batch tier is $0.003/min. So the new file model sits neatly below standard real-time recognition and just above the cheapest batch option. When you do the math, compare like for like: real-time against real-time, batch against batch.
To make that concrete, plug in your own volume:
A worked example: a small team transcribing 200 hours of support calls a month pays roughly $60 on the file model versus about $192 on Cloud Speech-to-Text standard. Real savings, but not life-changing at that scale, and honestly the bigger cost question is usually what you do with the transcripts afterward, not the transcription itself. More on that below.
How it stacks up against Whisper, Deepgram, and Scribe
I want to be careful here, because a fair head-to-head needs each vendor's own numbers, and the honest position is that the real A/B tests are being run by users, not by me. So instead of a spec table I can't fully source, here's what the people actually comparing these models found.
On cost, the reaction split almost down the middle. A developer who runs whispermemos.com did the math and came away impressed:
"I did the math based on our data from customers... and it seems to be $0.12 per hour of input audio... Scribe is $0.22. If the accuracy is close to Scribe, I think it's a good deal."
Others weren't convinced the accuracy gets close enough:
"Still more expensive and worse performing than ElevenLabs Scribe, unfortunately. Not sure who's the target audience for this."
And Whisper still has a loyal following, partly because it's free to self-host and partly because teams have already built pipelines around it. The most common pattern I saw was hybrid: Whisper for accurate word timestamps, then a Gemini model for the actual text. The short version is that Gemini 3.5 Transcribe is a strong, well-priced managed option, not an automatic winner over the specialist ASR tools. Which one wins depends on your audio and whether you need reliable diarization.
The verdict: who it's for
Pulling it together, here's where I land after the research:
| Gemini 3.5 Transcribe | |
|---|---|
| Strengths | Low price (~$0.005/min file), 2.6% English WER, built-in diarization + timestamps, 85+ languages, managed API |
| Weaknesses | Smart transcription can rewrite precise wording, diarization unreliable past 3 speakers, no batch/Flex tier, docs contradict themselves on status and speaker cap |
| Best for | Dictation, single-speaker voice notes, cost-sensitive transcription at scale |
| Look elsewhere if | You need verbatim output, 4+ speaker meetings, or the lowest possible batch cost |
The decision really comes down to your audio:
If I were shipping a dictation feature or transcribing a queue of single-speaker voice notes, I'd reach for this happily, the price and the English accuracy are hard to beat for a managed model. If I were building meeting notes for six-person calls, or anything where the exact words carry legal weight, I'd test Deepgram or ElevenLabs Scribe alongside it before deciding, and I wouldn't assume the benchmark numbers carry over to my audio.
For reference, if you've used Google Cloud's transcription console before, the file-upload workflow will feel familiar, you point it at audio and pick your options:
Where transcription fits for a support team
Here's the thing worth saying plainly, because it's the part a review can miss: Gemini 3.5 Transcribe is infrastructure. It turns audio into text, and it does that hard job well. But a transcript of a customer's phone call or voice note isn't an answer, it's just the start. Something still has to read that text, understand what the customer needs, pull the right information, and actually resolve the ticket.
That resolution layer is what I work on. I've spent the last few years around AI that sits on live support queues, and the pattern never changes: getting the words right is table stakes; acting on them correctly is the hard part. A transcription model gives you the words. It doesn't know your refund policy, your order system, or the last three tickets that customer opened.
This is the difference between a model and a teammate. Gemini 3.5 Transcribe is a model, brilliant at one narrow job. eesel is a teammate platform: you hire ready-to-work AI teammates for specific jobs, and the current roster includes an AI helpdesk teammate and an AI blog writer. The helpdesk teammate plugs into the helpdesk you already run, Zendesk, Freshdesk, Gorgias, and joins the existing queue, trained on your help center and past tickets.
Try eesel for voice and chat tickets
If you landed on a transcription review because your team is drowning in voice or chat tickets, remember the transcript is only half the job. The eesel AI helpdesk teammate picks up where a model like Gemini 3.5 Transcribe leaves off: it reads the customer's message, drafts or sends the reply, and closes the loop, across 1,000+ integrations.
The part I'd point to first is simulation mode. Before any AI touches a live customer, you run it against thousands of your own historical tickets and see exactly how it would have responded, so you're not flipping a switch and hoping, the same "test it on your own data first" advice I'd give for any transcription model, built into the product. Pricing is usage-based at $0.40 per ticket with no per-seat fees, and it's free to try until you've used $50. It's the difference between a smart transcription of a problem and an actual resolution of it.
Frequently Asked Questions
Is Gemini 3.5 Transcribe worth it?
How accurate is Gemini 3.5 Transcribe?
How much does Gemini 3.5 Transcribe cost?
Is Gemini 3.5 Transcribe better than Whisper?
Does Gemini 3.5 Transcribe support speaker diarization?
Article by
Kurnia Kharisma Agung Samiadjie
Kurnia is a software engineer and writer at eesel AI with two years of SEO experience, writing about AI tools, helpdesk software, and customer support. He pairs a developer's understanding of how these products are built with search-driven research into what actually ranks and resonates with the people searching for them.
