Google launches Gemini 3.5 Transcribe with two separate APIs for recordings and live streaming

Google has released Gemini 3.5 Transcribe, a speech-to-text model that arrives as two distinct endpoints rather than one — an architectural choice that dictates entirely different capabilities, limits, and pricing. gemini-3.5-transcribe handles recorded files through the Interactions API, while gemini-3.5-transcribe-live targets real-time bidirectional streaming via the Live API. According to Artificial Analysis testing, the average error rate sits at 4.0% for streaming and 2.6% for non-live processing, with a 70% improvement in time-to-final-transcript versus Chirp 3, Google's previous model. Automatic language identification covers more than 85 languages and handles mid-sentence switching.

Two interfaces, two different products

The Live API delivers continuous sub-second transcription. It emits interim_input_transcription for speculative segments while the speaker is still talking, then input_transcription once the turn ends. Audio enters as raw 16-bit PCM at 16 kHz mono, in 100 ms chunks, with support for automatic, hybrid, or manual voice-activity detection. Ephemeral tokens let mobile and web clients stream without holding an API key. The limits are hard: a live session caps at 10 continuous minutes, no diarization (speaker identification), and no word-level timestamps.

The Interactions API fills in what streaming cannot do. It offers diarization, word-level start and end timestamps, and custom vocabulary biasing — up to 1,000 terms, with best results under 100. Standard requests accept up to one hour of audio; that drops to half an hour when diarization or word timestamps are enabled. This is the core design difference: readable transcripts and auditable transcripts are now two separate API calls.

Verbatim versus smart, the choice that dictates features

Both interfaces expose two modes. verbatim is the default and returns everything: fillers, repetitions, false starts. smart strips disfluencies, resolves spoken self-corrections inline, and applies built-in formatting. Google's own example: "Um, so for the meeting, I think we need to, uh, invite Alice and... wait no, Bob and Carol." Verbatim keeps it all. Smart returns "For the meeting, I think we need to invite Bob and Carol." The trade-off is sharp: smart mode cannot be combined with word-level timestamps or diarization. Anyone who needs a clean transcript for reading and an accurate one for analysis will have to run two separate calls.

Performance: the numbers and their context

On FLEURS, a multilingual benchmark, the model reports 5.50% WER for streaming and 5.04% for non-live on a set of top languages and locales. Those figures are higher than the claimed averages (2.6% / 4.0%), suggesting the averages lean on easier languages or conditions. The 70% improvement in time-to-final-transcript versus Chirp 3 matters for real-time applications, but Google has not published absolute latency metrics — only the relative gain. There are no open weights and no self-hosted path; this is a managed-service decision, not an infrastructure one.

Availability: three tracks, all in public preview

Individual developers and early-stage startups start on the free tier of Google AI Studio. Mid-market teams move to a paid tier for higher rate limits, which also guarantees their content will not be used to improve Google's products. Regulated enterprises route through the Gemini Enterprise Agent Platform, which adds guaranteed throughput, compliance controls, and volume discounts. Both the developer and enterprise tracks are in public preview, so production commitments should be calibrated accordingly.