Gemini 3.5 Transcribe reaches public preview with separate streaming and pre-recorded APIs

Sources: primary, discussion
Summary: Google put Gemini 3.5 Transcribe into public preview with separate APIs for streaming and pre-recorded audio. Google cites Artificial Analysis word error rates of 4.0 percent streaming and 2.6 percent non-streaming. Google also states a 70 percent improvement in time to final transcription over Chirp 3.
Why it matters: Published error rates and latency deltas against a named predecessor make replacing an existing speech pipeline a measurable decision.