Google AI Releases Gemini 3.5 Transcribe: A Speech-to-Text Model Reporting 2.6% Average WER Across 85+ Languages
MarkTechPost Asif Razzaq ● Covered by 6 sources
Google launched Gemini 3.5 Transcribe for speech-to-text. It’s split into live and batch APIs, with 85+ languages and no open weights.
Based on reporting by MarkTechPost, Asif Razzaq — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Google has put out Gemini 3.5 Transcribe, a speech-to-text model aimed at both live voice interfaces and recorded audio. The interesting part is that it isn’t one product wearing two hats. It comes as two separate API endpoints, and they behave differently enough that anyone planning a rollout has to treat them as distinct tools.
The live path, gemini-3.5-transcribe-live, is built for bidirectional streaming through the Live API. Google says it can deliver sub-second, continuous transcription, with interim partials while someone is still speaking and finalized text when the turn ends. Audio has to be sent as raw 16-bit PCM at 16kHz mono in 100ms chunks. That endpoint supports automatic, hybrid, and manual voice-activity detection, but it also has hard limits: sessions top out at 10 minutes of continuous streaming, and it does not support speaker diarization or word-level timestamps.
The batch path, gemini-3.5-transcribe, runs through the Interactions API and handles pre-recorded audio. This side gets the features the live endpoint leaves out: diarization, word-level start and end offsets, and custom vocabulary biasing. Google says the vocabulary list can hold up to 1,000 terms, though it recommends staying under 100 for the best results. Standard requests can take up to an hour of audio, but that drops to 30 minutes when diarization or word timestamps are turned on.
Both endpoints offer two modes, and that choice matters as much as the endpoint choice. Verbatim keeps fillers, repetitions, and false starts. Smart cleans those up and can even rewrite a spoken self-correction into a cleaner sentence. But smart cannot be combined with word timestamps or diarization, so teams have to choose between a polished transcript and a more forensic one.
On performance, Google points to an average word error rate of 4.0% for streaming and 2.6% for non-streaming, as measured by Artificial Analysis. On the multilingual FLEURS benchmark, it reports 5.50% streaming and 5.04% non-streaming. Google also says time to final transcription is 70% faster than Chirp 3, its previous transcription model, and automatic language detection covers more than 85 languages, including mid-sentence code-switching.
The model is API-only, with no open weights and no self-hosted route. Developers can start on the free tier in Google AI Studio, paid tiers add higher rate limits and a promise that customer content won’t be used to improve Google’s products, and the enterprise route adds provisioned throughput, compliance controls, and volume discounts. Both developer and enterprise tracks are still in public preview. The Live API is already connected into LiveKit, Pipecat, Agora, Fishjam, Vercel, and Vision Agents, while Google lists uses ranging from contact centers and clinical documentation to meeting transcription and voice-controlled interfaces.
My take — AI-written commentary, not fact-checked reporting
This is the usual Google playbook: strong model, tightly controlled delivery, and just enough product segmentation to make integration teams earn their keep. The real tell is the tradeoff between smart text and audit-friendly metadata; once again, convenience and compliance are being forced into separate lanes. Open-weight folks will call it closed, enterprise buyers will call it practical, and both will be right.
Read more about this at: MarkTechPost