What Google actually shipped
On August 26, 2026, Google announced Gemini 3.5 Transcribe — a speech-to-text model in two flavours: a batch model for recorded audio and a Live variant for streaming.
What it claims, from Google's own announcement:
- 2.6% word error rate non-streaming, 4.0% streaming, measured by Artificial Analysis
- 0.40 seconds from end of speech to final text in streaming mode
- 85+ languages with automatic detection
- Up to 3 speakers identified, with timestamps
- Filler-word removal and self-correction handling built in — it drops the "umm" and keeps the corrected version when you restate something
- Custom vocabulary, and function calling to hand work off to other Gemini models
- Roughly an hour of audio in a single pass
That last one deserves more attention than it got. The Whisper API caps uploads at 25MB — about 20 minutes — so long recordings have to be split and stitched back together, and every cut point is a place where a word can go missing. An hour in one request removes that whole class of bug.
The benchmark trap everyone fell into
Within a day of the announcement, half a dozen posts compared Gemini's 2.6% against Whisper's "2.1% on LibriSpeech" or similar, and declared a winner.
Those numbers are not comparable, and it is worth being precise about why.
LibriSpeech test-clean is read English audiobook speech, recorded cleanly. It is the benchmark where MetaWhisp's own on-device model scores 2.76% WER — a number I quote, and one that describes a very specific, very easy condition.
AA-WER v2, the Artificial Analysis benchmark Google cited, is a different animal: roughly 8 hours of audio blended from agent conversations (50%), European Parliament recordings (25%), and earnings calls (25%), across varied accents and acoustic conditions. Messy, real, interruption-heavy speech.
A model scoring 2.6% on the hard benchmark and a model scoring 2.76% on the easy one are not tied. They were never on the same test.
Accuracy, all on one scale
Artificial Analysis runs every model through the identical AA-WER v2 set, which makes it the only apples-to-apples comparison available right now. Lower is better.
| Model | WER | Price / 1,000 min | Speed |
|---|---|---|---|
| ElevenLabs Scribe v2 | 2.2% | $3.67 | — |
| Gemini 3.5 Transcribe | 2.6% | $5.00 | 78× |
| AssemblyAI Universal-3 Pro | 3.1% | $3.50 | — |
| AssemblyAI Universal | 3.8% | — | — |
| OpenAI GPT-4o Transcribe | 4.0% | $6.00 | — |
| Whisper Large v3 | 4.1% | — | — |
| Whisper Large v2 | 4.1% | — | — |
| Deepgram Nova-3 | 5.2% | $4.30 | 501× |
Two things jump out. Gemini bought its accuracy at a speed cost — 78× realtime against Nova-3's 501×. And Deepgram, which sits last on accuracy here, is six times faster and cheaper per minute. If you are transcribing a million minutes of call-centre audio overnight, that trade is not obviously wrong.
Prices, pulled today
Every figure below came off the vendors' own pricing pages on 2026-08-28. Speech-to-text pricing changes often, so check before you budget.
Gemini 3.5 Transcribe
Google prices audio in and text out separately:
- Batch: $0.003/min audio in + $0.002/min text out — ~$0.005/min blended
- Live: $0.005/min in + $0.004/min out — ~$0.009/min blended
- Free tier available
The blended rate assumes 25 audio tokens per second in and 175 text tokens per minute out. Talk faster than average and your bill moves.
Deepgram
Deepgram is the most granular of the three, and diarization is included rather than billed extra:
| Model | Pre-recorded | Streaming |
|---|---|---|
| Nova-3 Monolingual | $0.0043 | $0.0048 |
| Nova-3 Multilingual | $0.0052 | $0.0058 |
| Flux English | — | $0.0065 |
| Flux Multilingual | — | $0.0078 |
| Whisper Large (hosted by Deepgram) | $0.0048 | — |
$200 in free credit to start, no expiry. Smart formatting and speaker diarization are bundled in. Redaction adds $0.0020/min, keyterm prompting $0.0013/min, entity detection $0.0017/min. The Growth plan starts at $4K/year for up to 20% off.
OpenAI
- Whisper API: $0.006/min
- gpt-4o-transcribe: $0.006/min
- gpt-4o-mini-transcribe: $0.003/min — the cheapest hosted option here
- gpt-4o-transcribe-diarize: $0.006/min, diarization included
- Live transcription: $0.017/min — notably expensive next to Gemini Live at ~$0.009
Whisper on your own machine
$0.00 per minute. The weights are open, the model runs on Apple Silicon, and there is no meter. That is the whole pitch, and it is a category difference rather than a discount: at 1,000 hours of audio, Gemini runs about $300 and local Whisper runs the electricity.
What each one is actually good at
Gemini 3.5 Transcribe — accuracy on messy audio
Best hosted accuracy short of ElevenLabs, an hour of audio per request, filler-word cleanup and self-correction handling that produce text you can paste without editing. The 0.40s streaming latency is genuinely quick. Three speakers with timestamps covers most interviews and 1:1s, though it will not carry a six-person roundtable.
Deepgram Nova-3 — throughput and unit cost
501× realtime means an hour of audio in about seven seconds. Sub-300ms streaming latency. Diarization included, no per-speaker cap advertised. It is last on this accuracy table, but the gap between 5.2% and 2.6% matters far less when the output is feeding search or analytics rather than a document a human will read.
Whisper — the one you can run yourself
4.1% on hard audio, 99 languages, and the only option on this page you can point at a folder and run with the Wi-Fi off. Hosted, it is unremarkable at $0.006/min. Local, it is the only model here that never sends a client's voice to a third party. Whisper large-v3-turbo gives up about 0.3 percentage points of accuracy for roughly 4× the speed, which is the trade most desktop apps take — including this one.
Choosing by what you're actually doing
Transcripts a person will read
Interviews, podcasts, client calls. Accuracy and formatting win here — Gemini 3.5 Transcribe at ~$0.005/min, or ElevenLabs Scribe v2 if the 0.4 points matter more than the ecosystem.
Millions of minutes into a pipeline
Support calls feeding analytics or search. Deepgram Nova-3: 501× speed, $0.0043/min, diarization bundled. Nobody reads these transcripts word for word, and 5.2% doesn't break keyword search.
Audio that must not leave the machine
Therapy sessions, legal calls, anything under NDA, anything where uploading is the thing you are trying to avoid. Local Whisper is the only answer on this page — not because it wins the accuracy table, but because it is the only one that doesn't require trusting a vendor with the recording. See private voice-to-text on Mac.
Working offline
Planes, dead-zone conference rooms, spotty hotel Wi-Fi. Every hosted model on this page returns an error. Local transcription just works.
Where this leaves on-device Whisper
Being straight about it: Gemini 3.5 Transcribe is more accurate than Whisper on hard audio, by a real margin, on the only shared benchmark available. If your only criterion is "fewest wrong words on a noisy recording," Google's model wins and it costs half a cent a minute.
What that comparison leaves out is that accuracy is one axis of four. The other three — cost at volume, whether the audio leaves your machine, and whether it works with no connection — are exactly where a local model is unbeatable, and no amount of WER improvement changes them. A 2.6% cloud model still uploads the recording. A 4.1% local model still runs on a plane, for free, forever.
MetaWhisp runs Whisper large-v3-turbo on the Apple Neural Engine, on-device, with zero telemetry — that is the trade it makes, deliberately. If cloud accuracy is what your work needs, the honest advice is to use the cloud model, and this article is not going to pretend otherwise.
Frequently asked questions
Is Gemini 3.5 Transcribe more accurate than Whisper?
On the AA-WER v2 benchmark, yes — 2.6% against Whisper Large v3's 4.1%. Be careful with comparisons that put Gemini's 2.6% next to a Whisper score from LibriSpeech test-clean: those are different benchmarks, and LibriSpeech is substantially easier audio.
What does Gemini 3.5 Transcribe cost?
About $0.005 per minute blended for recorded audio ($0.003/min audio in plus $0.002/min text out), and about $0.009 per minute for the Live streaming variant. There is a free tier. Rates from Google's pricing page on 2026-08-28.
Is Deepgram cheaper than Gemini?
Slightly — Nova-3 Monolingual is $0.0043/min pre-recorded against Gemini's ~$0.005/min, and Deepgram includes diarization and gives $200 of free credit. Deepgram is also about 6× faster. It scores worse on accuracy: 5.2% versus 2.6%.
How many speakers can Gemini 3.5 Transcribe identify?
Up to three, with timestamps. Deepgram's diarization does not advertise a speaker cap, and OpenAI ships a dedicated gpt-4o-transcribe-diarize at $0.006/min.
Can I run Gemini 3.5 Transcribe locally?
No. It is API-only — the Live API and Interactions API, plus Google AI Studio and the Gemini Enterprise Agent Platform. Whisper is the model in this comparison whose weights you can download and run yourself.
Which is best for long recordings?
Gemini handles roughly an hour in a single pass. The Whisper API caps at 25MB (~20 minutes), so longer files need splitting and rejoining. Running Whisper locally has no upload limit at all — file length is bounded by your disk.
Sources
- Google — Intelligent transcription with Gemini 3.5 Transcribe (2026-08-26)
- Artificial Analysis — Speech-to-Text WER Index (AA-WER v2, retrieved 2026-08-28)
- Deepgram pricing (retrieved 2026-08-28)
- Gemini API pricing (retrieved 2026-08-28)
- OpenAI API pricing (retrieved 2026-08-28)
