Google finally hired a digital stenographer that ignores your "uhms" and "ahs"

The new Gemini 3.5 Transcribe model processes audio in under a second, handling multiple languages on the fly while tagging speakers like a pro. It automatically scrubs away verbal filler and mid-sentence corrections, turning spontaneous babble into clean, readable text before the user even finishes a sentence.

Google is positioning this as the backbone for future voice-controlled agents, moving far beyond the simple transcription tools of the past. Testing on the FLEURS benchmark shows a 5% error rate, with Artificial Analysis reporting a 70% speed boost over the previous Chirp 3 architecture. Real-world performance in noisy environments, however, remains the ultimate test of its digital patience.

The era of humans needing to speak like robotic command-line interfaces is finally dying. By offloading the burden of articulation to the machine, the keyboard is being handed a genuine challenge for the first time in decades.

Source: Google Blog