Google launches Gemini 3.5 Transcribe with live transcription, Smart formatting, and support for 85+ languages

Two models: one for live speech, one for recorded files

Gemini 3.5 Transcribe Live has the model ID gemini-3.5-transcribe-live. It receives audio through a WebSocket connection over the Live API and is designed for sub-second responses, making it suitable for real-time captions, live streaming transcription, phone assistants, and voice agents.

The Live version caps each session at 10 minutes. It doesn't support speaker diarization or word-level timestamps, only phrase-level timing information. For longer meetings, apps need to chain sessions together, or hand the recording to the recorded version once it's finished.

Gemini 3.5 Transcribe has the model ID gemini-3.5-transcribe. It processes uploaded recordings through the Interactions API and is optimized for a lower final word error rate, while also supporting speaker diarization, word-level timestamps, and more complete transcription output.

Recordings can be up to 1 hour long. If speaker diarization or word-level timestamps are enabled, the limit drops to 30 minutes per file. The official docs state that up to 8 speakers are supported, though anything beyond 3 is still experimental.

Google AI Studio offers hands-on experiences for both live and recorded transcription. Developers can upload their own meeting, call, or interview recordings to decide which model fits their needs.

Key features: Smart cleanup, multilingual support, custom vocabulary, and annotations

Smart mode tidies up spoken content

Standard verbatim transcription keeps every "um," "uh," repeated phrase, and false start. Smart mode removes filler words, merges self-corrections, and reformats the text to reflect the final intended meaning.

In the ordering demo below, the speaker first asks for a large hazelnut latte, then switches to caramel, cancels an extra shot, changes the ice level, and finally downsizes from large to medium. The Smart output only keeps the final confirmed order.

Smart mode is well suited for dictating messages, emails, notes, and forms. If the use case requires preserving every utterance exactly as spoken, verbatim transcription is the way to go.

Over 85 languages with seamless switching

The models support more than 85 languages. Speakers can switch languages mid-sentence without manually specifying a language in advance.

The following demo cycles through English, Spanish, and Hindi, and also tests recognition of alphanumeric combinations.

Google says both live and recorded versions support multilingual transcription, but actual performance will still vary based on accent, recording quality, background noise, and language combinations.

Add custom terms: jargon, names, product names, and code

Developers can supply custom vocabulary to help the model recognize company names, product names, industry terms, people's names, and identifiers. The docs allow up to 1000 terms, though Google recommends keeping the most important ones under 100.

The live demo below incorporates background noise, custom terms like Neurowave and Server ID B742X, plus switching between English and Czech.

This demo shows these features can be combined, but it doesn't guarantee identical results across all noise levels and accents.

The recorded version supports speaker diarization and word-level timestamps

The recorded version can distinguish between different speakers and return the time position for each word. These features are useful for meeting minutes, customer service quality reviews, interview subtitles, and scenarios where replaying the original audio is necessary.

Smart mode cannot be used together with speaker diarization or word-level timestamps. Smart mode removes words and merges corrections, while word-level timestamps require output that stays aligned with the source audio. The docs also warn that enabling word-level timestamps may degrade transcription quality.

Official benchmark data: 4.0% streaming WER, 2.6% non-streaming WER

In Google's overall English transcription tests, the real-time streaming model achieved a word error rate (WER) of 4.0%, while the recorded non-streaming model hit 2.6%. A lower WER means fewer words were substituted, dropped, or added.

GOOGLE BENCHMARK · LOWER WER IS BETTER

On the multilingual FLEURS benchmark, the live version scored a WER of 5.50%, and the recorded version 5.04%. Both sets of numbers show that processing the full audio yields lower error rates, while the live version prioritizes speed.

GOOGLE BENCHMARK · LOWER WER IS BETTER

Compared with Chirp 3, Google also claims that Gemini 3.5 Transcribe reduces the time from audio end to final text by 70%. This is a relative improvement in finalization time, not a 70% reduction in error rate. These figures come from Google's release materials, and the article doesn't provide enough information for outsiders to fully replicate the experiments, so real-world results will need to be tested with your own recordings.

Already integrated into Gboard, Gemini macOS, Antigravity, and AI Studio

Google is rolling out these transcription models across its own input tools and developer products.

Gboard Rambler

Rambler is the dictation feature in Gboard. Users can speak naturally, pause, and correct themselves, and the system turns it all into a polished message ready to send.

This demo directly shows Smart mode in action on mobile input: after the user changes a date, restaurant, and phrasing, the output only keeps the latest version.

Gemini macOS

Gemini macOS can read the currently selected screen content after user authorization, then combine it with voice commands to complete tasks. In the demo, it converts a photo of a whiteboard into a table, uses Nano Banana to generate a birthday image, and searches the web to fill in an event address.

Here, speech-to-text is handled by Gemini 3.5 Transcribe, while screen understanding, web search, and image generation come from other models and tools within the Gemini app.

Antigravity

Antigravity combines voice commands with the code context in the editor. In the demo, the user asks it to modify a Flutter app's theme, then continues testing or kicks off a build.

The transcription model is responsible for turning speech into text; reading code, planning steps, and executing builds are Antigravity's agent capabilities. The Gemini 3.5 Transcribe model itself doesn't support function calling.

Google AI Studio already offers hands-on access to both models. The Chrome integration is still marked as "coming soon." Companies like Deepgram, AssemblyAI, and Krisp appear on the early partner list; these are integrations mentioned in the release materials, not independent evaluations.

Currently in Public Preview: know these limitations upfront

Both models are currently in Public Preview. You can start testing and integrating them, but before moving to production, you'll want to validate performance with your own accents, noise conditions, specialized terms, and long recordings.

Here are the main limitations:

• Live sessions are capped at 10 minutes each, with no speaker diarization or word-level timestamps.
• The recorded version handles up to 1 hour of audio; enabling speaker diarization or word-level timestamps reduces that to 30 minutes.
• Smart mode cannot be combined with speaker diarization or word-level timestamps.
• Word-level timestamps may lower transcription quality.
• The Chrome integration hasn't launched yet.

Based on typical token usage, Google estimates the Live version costs roughly $0.009 per minute and the recorded version about $0.005 per minute. That translates to around $0.54 and $0.30 for an hour of audio, respectively. Actual billing is calculated based on input and output tokens, and may also include storage, moderation, and additional model inference costs.

Choosing between them is straightforward:

• For real-time captions, live streaming, phone assistants, and voice agents, go with the Live version.
• For meetings, interviews, customer service recordings, and subtitle production, use the recorded version.
• Turn on Smart mode when you want text ready to send.
• For verbatim evidence, speaker labels, or timestamps, disable Smart and stick with verbatim.

Gemini 3.5 Transcribe is now available for testing through the Gemini API, Vertex AI, and Google AI Studio. Its main selling points come down to four things: real-time responses, high accuracy on recordings, automatic Smart cleanup, and support for multiple languages and custom vocabulary.