Google’s Gemini 3.5 Transcribe is offered as two separate APIs: gemini-3.5-transcribe for batch files and gemini-3.5-transcribe-live for real‑time streaming. The live endpoint gives sub‑second interim results but does not provide speaker diarization or word‑level timestamps and is limited to 10‑minute continuous sessions. The batch endpoint supports diarization, word offsets, custom vocabularies up to 1 000 terms and can process up to one hour of audio (30 minutes when diarization or timestamps are enabled), but it is not real‑time. Both services expose verbatim and smart modes; smart mode strips fillers and self‑corrections but cannot be combined with timestamps or diarization, forcing a choice between a clean summary and an auditable transcript.
Reported word error rates are 4.0 % for streaming and 2.6 % for batch, with a 70 % faster final‑transcription latency versus the prior Chirp 3 model. Pricing is token‑based; a blended estimate is about $0.005 per minute for batch and $0.009 per minute for live usage. There are no open weights and no self‑hosted option—everything runs as a managed service.
To decide which API fits your need, first map the core requirement: if you need live captioning, voice agents, or interactive bots, start with the live API and accept its limits; if you need speaker attribution, timestamps, or custom vocabulary for post‑call analysis, media captioning, or compliance logging, use the batch API. Test both on the free tier in Google AI Studio, monitor actual token consumption, and consider chaining a live stream for immediate feedback followed by a batch pass for the final, diarized transcript when latency is not critical. Keep sessions under 10 minutes for live or split longer recordings into chunks for the batch path. Track usage to stay within expected cost bounds and enable ephemeral tokens for client‑side apps to avoid exposing API keys.
#AI #SpeechToText #Gemini #Transcription #VoiceAI #CloudAPI