Google launches Gemini 3.5 Transcribe API with 85+ languages

• gemini-3.5-transcribe and gemini-3.5-transcribe-live split recorded and live audio work: Google's launch post introduced the pair, while Google's rollout post placed them in the API, AI Studio and Antigravity.
• Smart transcription edits speech as it recognizes it, handling self-corrections, removing fillers and formatting text, according to Google's feature list.
• Recorded jobs add word-level timestamps and attribution for up to three speakers; one early user's API rundown says the live endpoint is bidirectional and targets sub-second interactions.
• Google’s FLEURS charts put batch WER at 5.04% and streaming WER at 5.50%, as shown in the benchmark comparison.
• Google quoted roughly $0.005 per minute in a pricing reply, while Artificial Analysis's results estimates $5 per 1,000 minutes for batch and $9 for Live.

Chrome talk-to-type remains a future feature in Google's launch post. Vercel's AI Gateway changelog accepts a browser ReadableStream of 16 kHz, 16-bit PCM and exposes SMART and VERBATIM modes. The model card assigns the transcription models a January 2025 knowledge cutoff.

• gemini-3.5-transcribe-live streams continuously through the Gemini Live API; gemini-3.5-transcribe processes completed recordings through the Interactions API, per Google's launch post.
• Both models are in public preview through the Gemini API, Google AI Studio and Antigravity, as the API announcement states.
• Enterprise access is in public preview through Gemini Enterprise Agent Platform, and Google says it is coming to Gemini Enterprise for Customer Experience; Google's reply also lists AI Studio and the wider product suite.
• Consumer availability covers the Gemini app on macOS in English and Rambler on Android in select countries and languages, according to Google's Rambler post and Google's macOS post.
• Google describes base pricing as about $0.005 per minute in a pricing reply; Artificial Analysis's launch-day comparison normalizes that to $5 per 1,000 batch minutes and estimates $9 per 1,000 Live minutes.

The live model supplies continuous bidirectional transcription with sub-second latency. The recorded model adds timestamps and speaker attribution, a distinction set out in Google's launch post.

Google lists automatic detection across 85-plus languages, custom vocabulary, noise handling and recognition of phone numbers, postal codes and order IDs in the DeepMind feature list. Live streams can also follow language switches, according to Vercel's implementation notes.

The model card says both transcription variants are based on Gemini 3 Pro, accept audio and text in a 96K-token context window, and return up to 32K text tokens. Function calling is documented for the macOS Gemini app, where Google says background Gemini models can handle image generation and local-file analysis.

First-party

• FLEURS, non-streaming WER: Chirp 3 5.66% → Gemini 3.5 Transcribe 5.04%, -0.62 points, per the FLEURS batch chart.
• FLEURS, non-streaming WER: OpenAI GPT Live Transcribe 7.47% → Gemini 3.5 Transcribe 5.04%, -2.43 points, per the FLEURS batch chart.
• FLEURS, streaming WER: Chirp 3 7.32% → Gemini 3.5 Transcribe Live 5.50%, -1.82 points, per the FLEURS streaming chart.
• FLEURS, streaming WER: OpenAI GPT Live Transcribe 8.97% → Gemini 3.5 Transcribe Live 5.50%, -3.47 points, per the FLEURS streaming chart.

Third-party evaluators

• Batch throughput: ElevenLabs Scribe v2 55x realtime → Gemini 3.5 Transcribe about 84x, +29x, according to Artificial Analysis's comparison.
• First partial, WER: GPT Live Transcribe 6.3% → Gemini 3.5 Transcribe Live 5.8%, -0.5 points; latency: 0.26 seconds → 0.25 seconds, -0.01 seconds, per Artificial Analysis.

Customer-reported

Google names Vivo, Intellitek Health and Lingopal as early users in its launch post, but it did not publish a customer benchmark figure.

Artificial Analysis’s methodology asks for a verbatim transcript, while Google's smart-transcription path removes fillers and resolves self-corrections. Those are different output targets.

The recorded endpoint supports timestamped attribution for three speakers; Google calls support beyond three speakers experimental. The model card also lists foundation-model hallucinations, occasional slowness and timeout issues for both transcription variants.

One technical walkthrough reports a current 10-minute cap on individual Live sessions, according to one API rundown.

Vercel added both model slugs to AI Gateway on launch day.

Google has already put the model behind Rambler, which turns Android dictation into edited text and supports voice-driven corrections and style changes, as Google's reply confirms. The Gemini macOS app uses it for spoken requests that can reach image generation, search, text summaries and file analysis, according to Google's macOS post.

Google's launch post also names Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel and Vision Agents among platforms using the Gemini Live API for voice interfaces.

Pipecat v1.8 added launch-day support for Gemini 3.5 Transcribe alongside 233 changelog entries.

Its voice-agent plumbing changes include:

• processors and services now initialize and connect concurrently before pipeline startup;
• provider-native turn detection proposes boundaries while Pipecat's turn strategy decides the final boundary;
• Pipecat Evals can emit results.jsonl, score individual turns and run directories of scenarios;
• long-running async tools can opt into LLM-driven cancellation, and timed-out calls are cancelled automatically.

Pipecat said the framework now has more than 180 integrations.