Google Folds Speech-to-Text Into the Gemini API as Transcribe Enters Public Preview

Google launched Gemini 3.5 Transcribe on August 26, 2026, moving its transcription capability from the standalone Cloud Speech product line into the core Gemini API. The practical shift is not in headline accuracy scores — the model currently ranks fifth on independent benchmarks, behind ElevenLabs and Alibaba Cloud — but in what gets consolidated into a single API call: speaker diarization, word-level timestamps, and function calling, all in the base model, billed through the same token system as the rest of Gemini.

The launch replaces Chirp 3, Google's previous transcription offering, with a model developers call by a Gemini model ID. For teams assembling meeting-notes or voice-agent pipelines from multiple services — a transcriber, a diarizer, and a formatting pass — the new architecture is the more consequential development, even if the filler-word-removal demos are more photogenic.

Two Endpoints With Distinct Capabilities and Limits

Gemini 3.5 Transcribe ships as two separate API paths. The batch path, gemini-3.5-transcribe, routes through the Interactions API and handles pre-recorded audio: meetings, call-center logs, interview recordings. It supports speaker attribution for up to three speakers, generates word-level timestamps, and accepts up to one hour of audio per session — though that cap drops to 30 minutes when diarization or timestamps are enabled.

The streaming path, gemini-3.5-transcribe-live, runs over the Live API as a bidirectional WebSocket connection for interactive use: live captioning, voice interfaces, customer service calls. Google reports sub-second latency. This endpoint caps sessions at ten minutes and does not support speaker attribution or word-level timestamps. The two paths are separate SKUs with separate pricing: approximately $0.005 per minute for batch and $0.009 per minute for streaming, derived from Google's official 25-token-per-second audio billing rate — roughly $0.30 and $0.54 per hour of audio respectively. Both carry a free tier, which the pricing page explicitly notes may use data to improve Google's products; the paid tier does not.

What "Intelligent Transcription" Changes in Practice

Google's smart transcription mode applies model-level editing at inference: self-corrections are resolved ("let's meet Tuesday — no, Wednesday" becomes Wednesday's meeting), hesitation words are removed, and punctuation and formatting are applied. The model also accepts custom vocabulary lists to bias recognition toward domain-specific terminology — product names, medical terms, alphanumeric codes. Function calling allows the model to route tasks to other Gemini models, turning transcription into a component of a larger agent pipeline rather than a terminal step; this capability is currently available in the Gemini macOS app and coming to the batch API.

The editorial implication is material. Smart mode makes decisions speakers did not explicitly make, which is useful for meeting minutes and dictation but inappropriate where verbatim fidelity matters. Legal depositions, medical intake recordings, and journalistic interviews require the verbatim mode — which Google documents and preserves as the default. There is also a hard constraint: smart mode cannot be combined with speaker attribution or word-level timestamps. Teams must choose between readability and full traceability.

Where Gemini 3.5 Transcribe Actually Ranks

As measured by Artificial Analysis, whose non-streaming leaderboard covers 57 speech-to-text models, Gemini 3.5 Transcribe places fifth with a 2.6% word error rate. Alibaba Cloud's Fun-Realtime-ASR-preview leads at 1.7% WER, followed by ElevenLabs Scribe v2 at 2.2%, MAI-Transcribe-1.5 at 2.4%, and Smallest AI Pulse Pro also at 2.4%. OpenAI's widely deployed Whisper Large v3 sits around 4.2% on the same benchmark; Mistral's open-weight Voxtral Small reaches 2.8% with full self-hosted deployment control.

Google also claims a 70% improvement in time-to-final-transcription over Chirp 3, citing Artificial Analysis data — a figure the analytics firm noted represents Google-selected measurements on a model that had been available for roughly one day when reported, not an independently adversarial comparison. No external reproduction of a head-to-head test against open-weight alternatives existed as of this writing. Production teams evaluating Gemini 3.5 Transcribe should test with audio drawn from their own distribution — accent profiles, background noise levels, domain vocabulary density, and the presence of more than two speakers all affect real-world accuracy in ways that a single published WER figure does not capture.

The Case for API Consolidation

On raw non-streaming accuracy, ElevenLabs Scribe v2 and Alibaba Cloud's preview model lead. On streaming latency for voice agents, Deepgram Flux and ElevenLabs are stronger. On cost and control, Voxtral Small and NVIDIA Parakeet deliver self-hosted alternatives without API dependency.

What Gemini 3.5 Transcribe offers distinctively is surface consolidation within the Google stack. Before this launch, a diarized, timestamped, domain-adapted transcription pipeline on Google's infrastructure required Cloud Speech-to-Text, a separate diarization service, and a formatting layer — three billing lines, three integration points. A single call to gemini-3.5-transcribe now returns speaker-attributed, timestamped, formatted output with custom vocabulary applied. For teams running downstream Gemini models for summarization, extraction, or action-item detection, eliminating those service boundaries reduces both latency and complexity.

Whether that consolidation justifies a WER disadvantage against ElevenLabs depends on the workload. For multilingual meeting notes where Gemini will also process the output, the argument is strong. For legal or medical transcription where accuracy is the binding constraint, it is weaker — and the public preview status of the model means pricing, limits, and API behavior remain subject to change before general availability.

Constraints That Matter for Production Planning

Several hard limits bear repeating for any team evaluating Gemini 3.5 Transcribe for production: live sessions cap at ten minutes; batch sessions at one hour, or 30 minutes with speaker ID or timestamps enabled; the three-plus-speaker diarization is labeled experimental; and word-level timestamps are documented as potentially reducing overall transcription accuracy. The model is in public preview, meaning all of these parameters can change before GA.

On the consumer side, Gemini 3.5 Transcribe already powers Rambler in Gboard on Android and the Speak to Window feature in the Gemini macOS app, with Chrome talk-to-type, Docs Live, Keep, and Gmail integrations planned. The trajectory points toward transcription becoming the universal input layer for the Gemini stack — every spoken word entering Google's surfaces first passing through this model, then routing to whichever Gemini agent handles the follow-on task. The accuracy gap to the leaderboard's top two models, and whether it closes before GA, will determine whether Gemini 3.5 Transcribe becomes the default choice for Google-ecosystem developers or merely the most convenient one.