Loading page…
The release turns speech recognition into a Gemini product layer for developers and Google surfaces, but a product-specific price remains unavailable for teams weighing production use.
Listen to this story
Google’s public preview of Gemini 3.5 Transcribe gives builders a choice between low-latency streaming for live voice products and richer processing for completed recordings. The latter can attach word-level timestamps and identify up to three speakers, while the system can clean filler words and handle alphanumeric terms across more than 85 languages. Google reports 4.0% streaming and 2.6% non-streaming word error rates, but those are company-provided launch figures.
The Live API uses gemini-3.5-transcribe-live for continuous bidirectional audio and Google-stated sub-second latency.
The Interactions API uses gemini-3.5-transcribe for recordings, with timestamps and speaker attribution for up to three speakers.
Speaker support beyond three people is experimental, limiting immediate use for larger meetings or group calls.
Google has released Gemini 3.5 Transcribe in public preview, giving developers separate APIs for real-time voice interaction and recorded-audio processing. The split matters because a live assistant needs speed, while a call archive needs speaker labels and timestamps that can be searched or routed downstream.
For interactive applications, the Live API exposes gemini-3.5-transcribe-live as a continuous, bidirectional stream with Google-stated sub-second latency. That is the route for voice agents and live captioning, where the system must keep accepting audio as it returns results.
The Interactions API uses a different model name, gemini-3.5-transcribe, for meetings, call logs, and other completed recordings. Its added outputs are word-level timestamps and attribution for as many as three speakers; Google labels support beyond three speakers experimental.
Google positions the Live API for continuous bidirectional transcription in interactive voice applications, with stated sub-second latency.
Google positions the Interactions API for completed audio and includes speaker attribution and word-level timestamps.
The model is designed to do more than render utterances word for word. Google says it can remove filler words, resolve a spoken self-correction, format the result, and adapt to a supplied custom vocabulary. That makes the output a cleaned-up working draft, not necessarily a verbatim record.
Google also emphasizes terms that frequently break automated workflows: postal codes, order IDs, and other alphanumeric entities, including in noisy recordings. It says the model automatically detects and transcribes more than 85 languages, and can handle regional accents and dialects.
Google says Artificial Analysis measured average word error rates of 4.0% for streaming and 2.6% for non-streaming use. Google also says time to final transcription improved 70% over Chirp 3, its prior transcription model. Those results distinguish the two modes on accuracy, but they remain performance figures presented in Google’s launch material.
The launch does establish model identifiers and API paths, addressing the practical question of how each mode is called. But Google’s pricing page does not list a price for Gemini 3.5 Transcribe. For a preview aimed at voice agents, captioning tools, and post-call analytics, that leaves the cost comparison unresolved.
Editorial analysis
Google’s pitch is not simply that it can recognize speech. It is that transcription can become an intent-aware interface: clean up dictation, preserve specialized terms, and feed voice-driven actions across its products. The more consequential test is commercial clarity. Google has supplied named API models and preview routes, but teams building high-volume call, captioning, or agent systems will need a product-specific price and operating terms before they can compare this service against other transcription options. Those details are the next meaningful signal to watch.
Loading discussion...
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.