Google’s newest speech recognition model can follow conversations in more than 85 languages and, in pre-recorded mode, averages just 2.6 percent word errors, according to independent testing by Artificial Analysis. Streaming transcription runs slightly looser at 4.0 percent, and Google says the model finalizes transcripts 70 percent faster than Chirp 3, its predecessor.
Gemini 3.5 Transcribe is really two products. One endpoint, gemini-3.5-transcribe, digests uploaded files through the Interactions API. The other, gemini-3.5-transcribe-live, handles two-way streaming through the Live API. They carry different features, limits, and prices, so builders have to pick a lane.
The live endpoint emits interim transcriptions while a speaker is still talking, then a final version when the turn ends. Audio arrives as raw 16-bit PCM at 16kHz mono in 100ms chunks, and ephemeral tokens let mobile apps stream without an API key. But sessions stop at 10 minutes, and live mode lacks speaker diarization and word-level timestamps.
Those capabilities move to the Interactions endpoint instead, alongside custom vocabulary lists of up to 1,000 terms. There are no open weights, and no self-hosted deployment: the model is managed service only.
