Google AI Releases Gemini 3.5 Transcribe: A Speech-to-Text Model Reporting 2.6% Average WER Across 85+ Languages

Gemini 3.5 Transcribe Splits Speech-to-Text Into Two Distinct Products

Google has released Gemini 3.5 Transcribe, a speech-to-text model designed for real-time voice interfaces and recorded audio, reporting a 2.6% average word error rate for non-streaming and covering more than 85 languages. The launch is notable not for a single model, but for a deliberate split into two separate API endpoints that serve different use cases. This is a managed-service decision, not an infrastructure one, with no open weights and no self-hosted path available.

The Two-Endpoint Strategy

The first endpoint, gemini-3.5-transcribe, handles pre-recorded files through the Interactions API. The second, gemini-3.5-transcribe-live, manages bidirectional streaming through the Live API. These are not variations of the same product. Their features, limits, and prices differ, so planning around that split is key for any deployment. The Live API delivers sub-second, continuous transcription. Audio streams in as raw 16-bit PCM at 16kHz mono, in 100ms chunks. It emits interim transcription for speculative partials while someone is still talking, then finalizes when the turn ends. Ephemeral tokens allow mobile and web clients to stream without holding an API key. But the constraints are real and should shape expectations. Live sessions cap at 10 minutes of continuous streaming. Speaker diarization is not supported. Word-level timestamps are not supported. If your application needs those features, you are looking at the wrong endpoint.

What the Interactions API Covers

The Interactions API steps in where streaming cannot. It offers speaker diarization, word-level start and end offsets, and custom vocabulary biasing. The vocabulary list accepts up to 1,000 terms, though best results stay below 100. Standard requests accept up to one hour of audio, dropping to 30 minutes once diarization or word timestamps are enabled. That tradeoff matters. A readable summary and an auditable transcript are now two different API calls. Smart mode cannot be combined with word timestamps or diarization, meaning a readable summary and an auditable transcript are now two different API calls.

Verbatim Versus Smart Mode

Both endpoints expose two transcription modes, and choosing between them is the real design decision. Verbatim is the default. It returns everything, including fillers, repetitions, and false starts. Smart mode removes disfluencies, resolves spoken self-corrections inline, and applies structured formatting. Google's documented example illustrates the difference clearly. The input: "Um, so for the meeting, I think we should, uh, invite Alice and, wait no, Bob and Carol." Verbatim keeps all of it. Smart returns "For the meeting, I think we should invite Bob and Carol." Smart mode cannot be combined with word timestamps or diarization. That is the tradeoff to plan around. If you need an auditable record with speaker attribution, you lose the clean formatting. If you need clean formatting, you lose the granular metadata.

Two endpoints, not one: streaming trades diarization and word timestamps for sub-second latency.

Performance Numbers That Matter

As measured by Artificial Analysis, Google reports an average word error rate of 4.0% for streaming and 2.6% for non-streaming. That headline figure of 2.6% is the non-streaming result. On the multilingual FLEURS benchmark, across a set of top languages and locales, the model reports 5.50% streaming and 5.04% non-streaming. Time to final transcription improves by 70% over Chirp 3, Google's previous transcription model. Language coverage spans more than 85 locales with automatic detection and code-switching handled without configuration. Mid-sentence code-switching requires no special setup.

Who Should Use This

Company level is flexible. Solo developers and startups can begin on the Gemini API free tier via Google AI Studio. Mid-market teams move to the paid tier for higher rate limits, which also guarantees content is not used to improve Google's products. Regulated enterprises route through the Gemini Enterprise Agent Platform, which adds provisioned throughput, compliance controls, and volume discounts. Both developer and enterprise tracks are in public preview, so treat production commitments accordingly. This is not a stable, final release with long-term guarantees.

Ecosystem and Ecosystem Limits

The Live API is already wired into LiveKit, Pipecat, Agora, Fishjam, Vercel, and Vision Agents. On the consumer side, the model powers Rambler on Android, the Gemini app on macOS, and Google Antigravity. Chrome is listed as coming soon. Blended cost runs about $0.005 per minute for batch and $0.009 per minute for live. No open weights exist. This is API-only, and that decision has consequences. You cannot run this on your own hardware. You cannot fine-tune it locally. You are renting the capability, not owning it.

Hard Limits to Know Before You Build

The constraints are specific and should be part of any architecture discussion. Live sessions cap at 10 minutes. Files cap at one hour. With diarization enabled, that drops to 30 minutes. Smart mode cannot combine with timestamps or diarization. Pick one per call. Industries likely to benefit include contact centers and CX platforms, clinical documentation, media captioning and localization, legal and insurance intake, meeting tooling, and voice-driven developer tools. Applications range from real-time voice agents and live captioning to post-call analytics pipelines, meeting transcription with speaker attribution, dictation, and voice-controlled interfaces. The 10-minute live session cap means long-form streaming applications need a different strategy. The 30-minute file limit with diarization means extended meetings require chunking. This is not a drop-in replacement for every transcription workflow. It is a tool with sharp edges, and knowing where those edges are will determine whether your deployment succeeds.

Frequently Asked Questions

What are the main differences between the two API endpoints of Gemini 3.5 Transcribe?

The first endpoint, gemini-3.5-transcribe, handles pre-recorded files through the Interactions API, while the second, gemini-3.5-transcribe-live, manages bidirectional streaming through the Live API. Their features, limits, and prices differ, with the live endpoint offering sub-second continuous streaming but no diarization or word timestamps, and the interactions endpoint supporting those features.

How does the Smart mode differ from Verbatim mode in transcription?

Verbatim mode is the default and returns everything including fillers, repetitions, and false starts. Smart mode removes disfluencies, resolves spoken self-corrections inline, and applies structured formatting, as shown by the example where the input is transformed to a cleaner output, but Smart mode cannot be combined with word timestamps or diarization.

What are the specific limitations on session durations for live and file transcriptions?

Live sessions cap at 10 minutes of continuous streaming, while standard file requests accept up to one hour of audio. However, when diarization or word timestamps are enabled, the file limit drops to 30 minutes.

Who are the intended users for Gemini 3.5 Transcribe according to the article?

Solo developers and startups can use the Gemini API free tier via Google AI Studio, mid-market teams can use the paid tier for higher rate limits, and regulated enterprises can use the Gemini Enterprise Agent Platform for provisioned throughput and compliance controls. Industries likely to benefit include contact centers, clinical documentation, media captioning, legal and insurance intake, meeting tooling, and voice-driven developer tools.

What performance metrics does the article report for Gemini 3.5 Transcribe?

Google reports an average word error rate of 4.0% for streaming and 2.6% for non-streaming. On the multilingual FLEURS benchmark, the model reports 5.50% streaming and 5.04% non-streaming, and time to final transcription improves by 70% over Chirp 3. Language coverage spans more than 85 locales with automatic detection and code-switching.