Google splits transcription in two, and quietly absorbs your cleanup layer

The layer coming out of the stack is usually small and annoying to own: a filler-word list, a rule for collapsing repeated tokens, often a second model call that rewrites the raw string into something a downstream parser will accept. Google now describes that work as part of transcription itself, including self-corrections like "let's meet Tuesday, no, Wednesday" and automatic formatting ^[4]. The maintenance goes away, and so does the unedited string. Runtimewire reports that Google's audio transcription documentation does not establish separate verbatim and smart modes for the model ^[12]. If you retain transcripts for dispute resolution, clinical notes or QA scoring, settle that before the migration rather than after.

Then the accuracy split, which is the part the two endpoints make unavoidable. Google cites Artificial Analysis at 4.0% word error rate streaming and 2.6% non-streaming ^[5]. That is 1.4 points, or about 54% more errors on the interactive path ^[1]: roughly one wrong word in 25 live, against one in 38 for recorded audio ^[3]. The batch endpoint is also the one carrying speaker attribution and word-level timestamps ^[3]. Post-call analytics gets the better model and the richer output; the live agent gets neither, which argues for transcribing twice when the transcript is the record of the conversation and not just its input.

Multilingual deployments should size against a different figure again. On FLEURS the model is reported at 5.04% non-streaming and 5.50% streaming ^[7], and 5.04 is roughly 1.9 times the headline 2.6 ^[2]. Google says the model auto-detects and transcribes more than 85 languages ^[8]; runtimewire says the published documentation does not establish that 85-language count ^[12]. Two publishers, one product, and the gap sits in the exact place a procurement review looks.

That gap is wider than language counts. Runtimewire, writing about the August 26 announcement, says it disclosed no pricing, no benchmarks and no public model identifier, and that the Gemini catalog lists Gemini 3.5 Flash and Gemini 3.5 Live Translate but not Transcribe, with no price on the pricing page ^[10][11]. Google's own post names both identifiers, gemini-3.5-transcribe-live on the Live API and gemini-3.5-transcribe on the Interactions API ^[2][3], and publishes the error rates ^[5]. You can write the call. You cannot yet write the invoice, and Google already runs Cloud Speech-to-Text while general-purpose Gemini models accept audio ^[14], so this is a choice among its own paths as much as against a specialist vendor.

The feature that actually decides pipeline viability is the narrowest documented one. Wrong order IDs and phone numbers make a transcript useless to downstream automation even when the prose around them is clean ^[16], and Google's claim to handle alphanumeric entities in noise ^[15] rests on custom vocabulary that runtimewire says was demonstrated in a simulated interface carrying a warning that compatibility and availability vary ^[13].