The Pitch Is Not Accuracy, It Is Cleanup
Raw recognition accuracy has been a crowded competition for years. What differentiates this generation is post-processing: it actively drops filler words, resolves self-corrections toward the speaker's final intent, recognizes specialized terms and unusual spellings, and formats the result into readable paragraphs. The deliverable shifts from "a string of timestamped words" to "a record you can actually read." On the multilingual side it detects the language and then transcribes, covering more than 85 languages including regional accents and dialects, with no need to specify in advance. Speaker attribution is reliably supported for up to three speakers, with more than three marked experimental — a boundary worth noting if you are building meeting notes. Custom vocabulary is configurable for product names, personal names and industry jargon that default recognition handles poorly.
Two APIs for Two Situations
The model is reachable through two interfaces: the Live API for real-time streaming with sub-second latency, and the Interactions API for recorded audio with speaker identification and word-level timestamps. The model identifiers are gemini-3.5-transcribe for batch audio and gemini-3.5-transcribe-live for real-time streaming, with audio billed in tokens like any other Gemini input. On the product side it is already wired into Rambler in Gboard on Android and the Gemini app on macOS, with Chrome support coming. Reports mention a 96,000-token context window, roughly an hour of meeting audio.
How to Read the Numbers
The published figures are 4.0% average word error rate streaming and 2.6% non-streaming, attributed to Artificial Analysis. On the FLEURS benchmark across top languages and locales the numbers are 5.50% streaming and 5.04% non-streaming. The two sets differ because the evaluation corpora differ — cite which one you mean rather than blending them. All of it is vendor-supplied or vendor-cited, with no independent reproduction so far. For teams choosing a transcription tool, the thing to test is not those percentage points but three failure modes that matter more in production: whether code-switching mid-sentence causes language misdetection, whether proper nouns are unusable unless added to custom vocabulary, and how far speaker separation degrades past three participants. Trying it during public preview is cheap — just do not treat the benchmark scores as the selection decision.
via: Google's blog, The Decoder, 9to5Google