Google has launched Gemini 3.5 Transcribe, a new speech-to-text model that automatically detects and transcribes over 85 languages, removes filler words, and fixes self-corrections as you speak. Google reports a 2.6% average word error rate on non-streaming audio and says time to final transcription is 70% faster than its previous Chirp 3 model.
Try It: Turn a Rambling Voice Note Into Clean Text
Open Google AI Studio, pick the Gemini 3.5 Transcribe live model, and record a messy voice memo. Because the model strips "ums" and "ahs", auto-formats text, and resolves corrections like "let's meet Tuesday, no, Wednesday" on the fly, the output reads closer to a finished script than a raw transcript. Developers can wire the same capability into an app through the Interactions API.
Why It Matters for Creators
Transcription is the unglamorous first step in most audio and video work: podcast show notes, subtitle files, interview pulls, and voice-driven drafting. A 70% cut in time-to-final and a sub-3% error rate mean fewer manual fixes before a transcript is usable. Because the model also streams partial results through the Live API, real-time captioning and live note-taking become viable, not just batch jobs. Multi-speaker identification with timestamps for up to three speakers makes it practical for interviews and panels without hand-labeling who said what.
Key Details
Languages: Detects and transcribes 85+ languages automatically.
Accuracy: 2.6% average WER non-streaming and 4.0% streaming (Artificial Analysis); 5.04% and 5.50% on the FLEURS benchmark.
Speed: 70% faster time to final transcription than Chirp 3.
Where it runs today: Gboard's Rambler feature on Android, the Gemini app on macOS with voice commands, and Google Antigravity, with Chrome support coming.
Extras: Custom vocabulary for specialized jargon and speaker timestamps for up to three voices.
What to Do Next
Test it against your current tool on a real recording before switching. If you already lean on OpenAI's GPT-Transcribe or an open model like Cohere Transcribe, run the same audio through Gemini 3.5 Transcribe and compare word error rate, speaker labels, and latency on your own material. Google's launch coverage has the full rollout details.
