Gemini 3.5 Transcribe adds real-time speech-to-text in 85 languages

• Gemini 3.5 Transcribe splits speech-to-text into live and recorded-audio endpoints, with streaming, speaker identification, custom vocabulary, and function calling, as OfficialLoganK's launch post describes.
• The model turns raw speech into formatted text by removing filler words and handling self-corrections, while GoogleDeepMind's feature thread says it recognizes product names, IDs, and other custom vocabulary across 85-plus languages.
• Google’s FLEURS charts place the new model ahead of Chirp 3 in both modes, according to GoogleDeepMind's benchmark post.
• Public preview reaches Google AI Studio, Antigravity, and enterprise surfaces, while OfficialLoganK's overview says Google products already use the model.
• Transcription costs roughly $0.005 a minute, or about $0.30 an hour, according to OfficialLoganK's pricing reply.

In Google’s launch post, Antigravity pairs transcription with a user-permitted screen context and chat history, while the macOS Gemini app can route voice commands to other Gemini models for file summaries or image generation. The Apache-licensed Jot demo maps hold fn, speak, and paste polished text wherever the cursor sits.

• Developers get gemini-3.5-transcribe-live through the Live API for continuous, bidirectional audio streaming with sub-second latency, says OfficialLoganK's launch post.
• The Interactions API gets gemini-3.5-transcribe for recordings, meetings, and calls, with word-level timestamps and speaker attribution in Google’s launch post.
• Public preview is available in Google AI Studio and Antigravity for developers, plus Gemini Enterprise Agent Platform for enterprises; GoogleDeepMind's release thread also points to the Gemini macOS app and Android Gboard.
• The consumer rollout is Gemini on macOS in English and Rambler on Android in selected countries and languages, with Chrome announced as next; OfficialLoganK's price reply puts usage at about $0.005 per minute.

First-party

• Non-streaming FLEURS WER: Chirp 3, 5.66% → Gemini 3.5 Transcribe, 5.04%, -0.62 points, in GoogleDeepMind's non-streaming chart.
• Streaming FLEURS WER: Chirp 3, 7.32% → Gemini 3.5 Transcribe Live, 5.50%, -1.82 points, in GoogleDeepMind's streaming chart.
• Streaming FLEURS WER: OpenAI GPT Live Transcribe, 8.97% → Gemini 3.5 Transcribe Live, 5.50%, -3.47 points, in GoogleDeepMind's streaming chart.

Third-party evaluators

Google’s launch post attributes average WERs of 4.0% for streaming and 2.6% for non-streaming to Artificial Analysis, but does not give an earlier Artificial Analysis score for either mode.

Customer-reported

The same post names Vivo, Intellitek Health, and Lingopal as early users praising latency, accuracy, and language coverage, but publishes no customer benchmark values.

Smart transcription intentionally removes “ums” and “ahs,” repairs self-corrections, and auto-formats output, as GoogleDeepMind's feature thread shows. That makes its polished text a different artifact from a verbatim record.

Pre-recorded speaker attribution works natively for up to three speakers; Google calls three-plus speaker support experimental in its launch details.

Google’s model card lists hallucinations, occasional slowness, and timeouts among Transcribe and Transcribe Live’s known limitations. It gives the underlying model a January 2025 knowledge cutoff.

The product boundary is an API boundary:

• Live transcription: gemini-3.5-transcribe-live runs continuous, bidirectional streams through the Live API for captions and interactive voice interfaces, according to Google’s announcement.
• Recorded audio: gemini-3.5-transcribe uses the Interactions API for completed media, returning speaker attribution and word timestamps.
• Language cleanup: smart transcription handles self-corrections, filler removal, and formatting; custom vocabulary supplies the specialized names and spellings a project needs.
• Action routing: function calling can hand a spoken request to other Gemini models for image generation or file analysis. Google says that capability is currently available in the Gemini macOS app.
• Token envelope: both Transcribe variants accept audio and text in a 96K-token context window and return text up to 32K tokens, per the Gemini 3.5 Audio model card. The card says both are based on Gemini 3 Pro.

Rambler is the Android implementation: OfficialLoganK's reply identifies Gemini 3.5 Transcribe as the model powering the new feature. Google says Rambler formats spoken thoughts, strips fillers, and accepts voice edits, spelling corrections, and style changes.

Google also lists Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel, and Vision Agents alongside the Live API rollout. Its announcement describes those platforms as handling real-time media infrastructure around voice-driven interfaces.

The macOS Gemini app uses the model for dictated text and voice commands with screen context. Antigravity adds file names, agent thoughts, and active-document context when the user permits it.

gregisenberg proposed vibe-coding a personalized Wispr Flow-style app with the model; AmirMushich's follow-up said they built and tested one, reporting about 25 seconds to decipher “drunk dad’s mumbling.”

Jot, built by Ammaar Reshi and explicitly marked as unsupported by Google, is a small public version of that workflow: hold fn, dictate, and insert cleaned text at the active cursor. Its source repository is Apache 2.0 licensed and uses Python, Shell, and Swift.