Gemini 3.5 Transcribe is Google's speech-to-text model, built to write spoken audio down as clean, punctuated, formatted text. A free tier lives in Google AI Studio, with API billing by audio volume beyond it. The model auto-detects over 85 languages, already powers the Rambler dictation feature on Android and the Gemini app on macOS, and sits on the same API as the rest of the Gemini family. Voice-app developers and heavy dictation fans are first in line.
Gemini 3.5 Transcribe ships with two output styles, a verbatim mode that keeps every hesitation and a smart mode that cleans the speech before writing it down. You say 'let's meet Tuesday, no, Wednesday' and only Wednesday survives in the final text, while the ums quietly vanish (nobody will miss them).
The rest hides in the technical details, and they matter daily. Where classic speech recognition plays the audio back word for word, this model behaves closer to a human proofreader.
The model posts an average word error rate of 2.6% on pre-recorded audio and 4% in streaming, as measured by Artificial Analysis at launch. Against Chirp 3, its Google predecessor, time to final transcription drops by 70%.
Rivals are wide awake, though. At launch the model sat fifth on the Artificial Analysis non-streaming board, behind ElevenLabs Scribe v2 among others, and OpenAI prices its own file transcription in the same range. Google's edge lies elsewhere, one single call returns speakers, timestamps and formatting, where other stacks bolt several pieces together for the same job.
Access runs through the Gemini API in Google AI Studio, through Google Cloud's agent platform, or through Google's own products. On Android, Rambler dictation already relies on it, and a Chrome rollout has been announced. The live variant targets real-time captioning and the kind of AI voice tools where every millisecond counts.
On the billing side, a minute of audio comes to a blended $0.005 for files and around $0.009 for live streaming, free tier included on both endpoints. Those figures describe the preview period, so check Google's pricing page before building a budget around them.
| Endpoint | Typical use | Known limit |
|---|---|---|
| gemini-3.5-transcribe | Audio files, meetings, recorded calls | 1 hour of audio per file |
| gemini-3.5-transcribe-live | Real-time streaming, captions, voice agents | 10-minute sessions |
| Rambler on Android | Cleaned-up dictation, voice editing | Consumer feature, no charge |
| Gemini app on macOS | Voice commands, search, file analysis | Consumer feature, no charge |
Partly, yes. A free tier is available in Google AI Studio, and consumer features built on the model, like Rambler on Android, cost nothing. Past the free allowance, the API bills by audio volume, and file transcription ranks among the cheaper managed speech services currently on the market.
It comes down to infrastructure. Whisper is open source and self-hostable at near-zero marginal cost, but it returns raw text with no cleanup and no speaker labels. Gemini 3.5 Transcribe delivers formatted, timestamped, speaker-attributed text in one call, with no server of your own to run.
Three speakers get reliable attribution on pre-recorded audio, timestamps included. Beyond three, the feature exists but Google still labels it experimental, so a six-person panel discussion will call for a careful proofread of the output before you publish anything.
Yes on both counts. Google highlights regional accents and dialects as a strength, and language detection works per utterance across the 85+ supported languages, so a conversation flipping between English and Spanish mid-sentence stays intact. Custom vocabulary also helps it catch names and jargon it would otherwise mangle.
Verdict: Hours of re-listening saved on every recorded call, that is the pitch in one line, and it lands squarely with voice-app builders, caption pipelines and dictation die-hards on Android.
