Gemini 3.5 Transcribe is Google's newest speech-to-text model, and it replaces Chirp 3 as the transcription engine behind Search Live, Gemini Live, Docs, Gmail, Keep, and the Rambler dictation feature in Gboard. Google shipped it on August 26, 2026, and it's already selectable in Google AI Studio under two model IDs — gemini-3.5-transcribe for uploaded audio files and gemini-3.5-transcribe-live for real-time streaming.
Short answer: Gemini 3.5 Transcribe is Google's speech-to-text model, launched August 26, 2026, with a 4.0% word-error rate on streaming audio and 2.6% on pre-recorded files, support for 85+ languages, and speaker diarization for up to 8 people. Use it via the Gemini API as
gemini-3.5-transcribe, priced from $0.003 per minute of audio, with a free tier in Google AI Studio.
In my testing, I cross-checked every number in this guide against Google’s own launch post and its live API documentation, because launch-week figures for a brand-new model tend to get miscopied by aggregator sites within a day or two. It posts a 4.0% word-error rate on streaming audio and 2.6% on pre-recorded files across Google's internal real-world test set, with higher error rates — 5.50% and 5.04% — on the public FLEURS benchmark, which is a harder, more linguistically diverse test than most vendors report against. It automatically detects language across more than 85 locales, handles mid-sentence code-switching, tags up to eight speakers in a recording, and returns word-level timestamps on request. Here's exactly how to start using it, what it costs per minute, and the mistakes that waste your first few requests.
What you'll need
A Google account is enough to try it for free in Google AI Studio, where the free tier covers both transcription models with no card on file. For anything past casual testing, you'll need a Google AI Studio or Vertex AI project with billing enabled, since paid-tier usage is billed per token (or the equivalent per-minute rate) from the first call past the free quota. If you're building against the API directly, you need the google-genai SDK for Python or JavaScript, or you can call the REST endpoint with curl and an API key. If you just want dictation, not the developer API, you don't need any of this — it already powers Rambler in Gboard on Android and the Gemini app on macOS, and it's rolling out to Chrome next, per Google's launch post.
Step-by-step: setting it up
1. Decide whether you need the batch model or the live model
gemini-3.5-transcribe takes an uploaded audio file and returns a finished transcript — use it for recordings, podcasts, and meeting audio. gemini-3.5-transcribe-live streams over WebSockets for real-time captioning or live dictation. In my testing, most one-off transcription jobs only need the batch model; the live model earns its higher price when latency actually matters.
2. Generate an API key in Google AI Studio
Sign in at Google AI Studio and create an API key. Both transcribe models are visible in the model picker under "Try it in Google AI Studio," and the free tier is enough to run a handful of test files before you need to attach billing.
3. Upload the audio and call the model
For files longer than a few seconds, Google's docs recommend uploading through the Files API first, then passing the file's URI to the model rather than inlining raw audio:
“`python from google import genai
client = genai.Client() audio_file = client.files.upload(file="path/to/sample.mp3")
interaction = client.interactions.create( model="gemini-3.5-transcribe", input=[{"type": "audio", "uri": audio_file.uri, "mime_type": audio_file.mime_type}], ) print(interaction.output_text) “`
4. Add language hints if you know the language in advance
By default, the model auto-detects the spoken language and switches languages mid-file when speakers code-switch. If you already know the language, passing a BCP-47 code like es-ES in language_codes improves accuracy over letting it guess.
5. Turn on diarization or timestamps only if you need them
Both are opt-in through the mode object, and Google's own docs warn that word-level timestamps can measurably degrade overall transcription accuracy — so don't enable it by default just because it's available.
6. Choose verbatim or smart transcription
Verbatim is the default and keeps every "um," false start, and self-correction exactly as spoken. Smart transcription cleans up disfluencies, resolves spoken corrections, and auto-formats lists and numbers — but it can't be combined with diarization or timestamps in the same request.
Example prompts you can copy
These are the request bodies I'd actually reach for, adapted from Google's own docs:
Plain transcription, auto-detected language: “python interaction = client.interactions.create( model="gemini-3.5-transcribe", input=[{"type": "audio", "uri": audio_file.uri, "mime_type": audio_file.mime_type}], ) “
Meeting recording with speaker labels and timestamps: “python generation_config = { "transcription_config": { "mode": {"type": "verbatim", "diarization_mode": "speaker", "timestamp_granularities": ["word"]} } } “
Jargon-heavy audio with custom vocabulary: “python generation_config = { "transcription_config": {"custom_vocabulary": ["Gemini", "Kubernetes", "BigQuery"]} } “
A rambling voice memo you want cleaned up, not verbatim: “python generation_config = {"transcription_config": {"mode": {"type": "smart"}}} “ Google's own example shows what smart mode does to a real utterance: "Um, so for the meeting, I think we should, uh, invite Alice and, wait no, Bob and Carol" becomes "For the meeting, I think we should invite Bob and Carol."
Common mistakes to avoid
The mistake I'd flag first: trying to combine smart transcription with diarization or timestamps. They're mutually exclusive — smart mode only works inside verbatim's sibling settings, so if you need speaker labels and clean formatting, you have to pick one and post-process for the other. Second, enabling diarization or timestamps on a long file and getting cut off — standard requests support audio up to one hour, but that drops to 30 minutes the moment diarization or word-level timestamps are turned on. Third, stuffing the full 1,000-term custom vocabulary limit; Google's own docs say best results come from around 100 terms, not the maximum. Fourth, mixing up the two model IDs — gemini-3.5-transcribe and gemini-3.5-transcribe-live are separate APIs with separate pricing, so budgeting or benchmarking against the wrong one gives you the wrong number. Fifth, quoting the 4.0%/2.6% word-error rates as if they're the only numbers that matter — Google's own FLEURS benchmark results (5.50%/5.04%) are less flattering and worth citing alongside them.
Batch vs. live pricing
gemini-3.5-transcribe (batch) |
gemini-3.5-transcribe-live (streaming) |
|
|---|---|---|
| Best for | Uploaded files: recordings, podcasts, meetings | Real-time captioning, live dictation |
| Free tier | Free of charge | Free of charge |
| Input (paid) | $2.00/1M tokens, ~$0.003/min of audio | $3.50/1M tokens, ~$0.005/min of audio |
| Output (paid) | $12.00/1M tokens, ~$0.002/min of text | $21.00/1M tokens, ~$0.004/min of text |
| Blended rate | ~$0.005/min | ~$0.009/min |
| Max audio length | 1 hour (30 min with diarization/timestamps) | Streaming, no fixed file cap |
| Diarization | Up to 8 speakers (3+ experimental) | Not documented for live mode |
Pricing confirmed against Google’s Gemini API pricing page on August 27, 2026. The batch model is roughly half the live model's blended rate, which tracks with the live model doing continuous low-latency streaming work rather than a single finished pass over a file.
Tools that make this easier
If you haven't set up a Gemini account at all yet, my how to use Gemini guide covers the account and app basics this article assumes, and how to use Gemini AI goes deeper on the assistant itself. For the consumer side of what it now powers, how to use Gemini Live and how to use Gemini in Google Docs cover two of the surfaces it already runs on. If you're comparing this against OpenAI's voice stack instead of building on the API, how to use ChatGPT Voice is the closest equivalent. I ran a similar pricing-and-benchmark verification on Google's other recent launch in Gemini 3.7 Flash: pricing, benchmarks, and setup guide, useful if you're deciding which current Gemini model fits a given job. My starter kit for AI is the place to start if you haven't set up any model accounts yet, and how we test AI tools explains how I verify numbers like the ones above before publishing them.
Where this leaves you
It's a real upgrade over Chirp 3, not just a rename — a 70% faster time-to-final-transcript, per Google, plus diarization and custom vocabulary that Chirp 3 didn't offer. The free tier in AI Studio is generous enough to test both models before committing to billing, and the batch model's ~$0.005/min blended rate is cheap enough that cost won't be the reason you skip it. The setting most worth remembering is the mode exclusivity: pick verbatim with diarization and timestamps for meeting notes, or smart mode for a clean read — you can't have both in one request.
Last updated: August 27, 2026 · By Vishal Swami, Founder & Lead AI Reviewer, AISagely
Frequently Asked Questions
Is Gemini 3.5 Transcribe free to use?
Yes, up to the free tier limits in Google AI Studio, for both the batch and live models. Past that, it's pay-per-use at roughly $0.005 per minute blended for the batch model and $0.009 per minute for the live model, billed to whichever Google Cloud or AI Studio project holds your API key.
How long does it take to transcribe audio with it?
For a batch request, most of the wait is upload time plus processing, and Google says time-to-final-transcript is 70% faster than the older Chirp 3 model. For live use, transcription streams back in real time as audio comes in, so there's effectively no separate wait at all.
What is the easiest way to try it?
Open Google AI Studio, select gemini-3.5-transcribe from the model picker, and upload a short real recording rather than a clean test clip — that's where accents, background noise, and jargon actually show up.
Is Gemini 3.5 Transcribe better than Chirp 3?
On Google's own numbers, yes — it adds speaker diarization, word-level timestamps, custom vocabulary, and a smart-formatting mode that Chirp 3 didn't have, plus a 70% faster time-to-final-transcript. Chirp 3 is being phased out as it rolls out across Google's own products.
Does Gemini 3.5 Transcribe work on non-English audio?
Yes. It auto-detects language across more than 85 locales and can handle a speaker switching languages mid-sentence without you setting a language code in advance. Accuracy still improves if you pass a known BCP-47 language code instead of relying on auto-detection.