Gemini 3.5 Transcribe

Why teams choose it

Tradeoffs to know

• Developer and enterprise access is public preview; consumer surfaces and languages vary by product and region.
• Pre-recorded speaker attribution is documented for up to three speakers; more than three is experimental.
• WER and latency figures are vendor-reported via Artificial Analysis and should be remeasured on your audio domain.

When not to use this

• Not ideal for simple tasks where cheaper models in the same lineup are good enough.
• Avoid for regulated or high-stakes outputs without evaluations that mimic your tooling, data, and review process.
• Pair catalog notes with comparisons and your own benchmarks before declaring a routing winner.

Technical specs

Inputs

audio

Outputs

text

Capabilities

speech-to-text, streaming transcription, batch transcription, speaker attribution, custom vocabulary, multilingual ASR, disfluency cleanup

License

Proprietary API (public preview)

Model string

gemini-3-5-transcribe

Benchmarks

{
  "source": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-5-transcribe/",
  "vendor_reported": true,
  "aa_wer_streaming_pct": 4,
  "aa_wer_non_streaming_pct": 2.6,
  "fleurs_wer_streaming_pct": 5.5,
  "fleurs_wer_non_streaming_pct": 5.04
}