LLamjinLab
AI NewsWords 1593Read time≈ 4 min

Google Launches Gemini 3.5 Transcribe With Smart Dictation and Two Speech-to-Text APIs

Gemini 3.5 Transcribe adds smart dictation, 85+ languages, live streaming, speaker labels, timestamps, and paid API access.

AIGeminiGoogleTool Use

Google introduced Gemini 3.5 Transcribe on August 26, 2026, as a dedicated speech-to-text model for live voice interfaces and recorded audio. The public-preview release supports more than 85 languages and adds an optional “smart” mode that removes filler words, resolves spoken corrections, and formats unstructured speech into readable text.

The model is available to developers through two separate Gemini API endpoints. gemini-3.5-transcribe-live handles real-time audio over the Live API, while gemini-3.5-transcribe processes uploaded recordings through the Interactions API. Google has also deployed the underlying transcription technology across selected consumer products, including Rambler in Gboard and the Gemini app for macOS.

The distinction between transcription and broader assistant behavior matters. Gemini 3.5 Transcribe itself produces text; Google’s model reference lists function calling, file search, image generation, code execution, and thinking as unsupported. When the macOS app uses a dictated command to summarize a file or generate an image, other Gemini models and product-level tools perform that work after the speech has been transcribed.

1. Two Models Serve Different Audio Workflows

The live model connects through a bidirectional WebSocket session and emits incremental text while audio is still arriving. Google describes the latency as sub-second, making the endpoint suitable for captions, dictation interfaces, and voice applications that need visible text before a speaker finishes a long passage.

Live sessions accept raw 16-bit PCM audio, with Google’s implementation guide specifying mono audio at 16 kHz and recommending chunks of roughly 100 milliseconds. A continuous live transcription session is limited to 10 minutes. Applications that need longer conversations must manage session turnover in their own infrastructure.

The non-streaming model accepts prerecorded audio files and can process up to one hour in a request. That limit falls to 30 minutes when word-level timestamps or speaker diarization are enabled. Unlike the live endpoint, it can attach start and end offsets to individual words and identify speakers within a recording.

Google’s launch announcement describes speaker attribution for up to three speakers. The API reference exposes support for as many as eight, but labels attribution involving three or more speakers as experimental. Developers handling larger meetings should therefore treat the eight-speaker figure as a technical ceiling, not a guarantee of stable attribution quality.

Both endpoints automatically identify supported languages and can change languages within a session or sentence without requiring developers to configure a single language in advance. If the language is known, an application can supply a BCP-47 code to bias recognition toward it.

Custom vocabulary offers a second form of steering. Developers can provide as many as 1,000 terms for names, acronyms, technical expressions, or other uncommon phrases. Google says results are typically best with lists of no more than 100 terms, suggesting that targeted vocabularies are preferable to broad dictionaries.

2. Smart Transcription Changes the Output, Not Just Its Presentation

Gemini 3.5 Transcribe has two output modes. The default VERBATIM mode preserves filler words, repetitions, false starts, and spoken self-corrections. SMART mode instead attempts to produce the text the speaker intended to communicate.

In smart mode, the model can remove expressions such as “um” and “uh,” eliminate stuttering, and apply punctuation and sentence casing. It also resolves inline corrections: a speaker who initially says Tuesday and then corrects the date to Wednesday receives a transcript containing only Wednesday.

The formatting system can convert a spoken sequence into paragraphs, bullet points, or numbered lists. It also applies inverse text normalization, turning spoken quantities into compact written forms—for example, converting “twenty six million dollars” into “$26M.”

This behavior makes smart mode useful for messages, notes, prompts, and first drafts. It is not equivalent to a literal record, however. Because the model deliberately deletes and rewrites parts of the utterance, organizations producing legal transcripts, research data, quotations, or compliance records should use verbatim output and review the result against the original audio.

Google also imposes a technical separation between cleaned output and detailed annotations. Smart transcription cannot be combined with word-level timestamps or speaker diarization in the recorded-audio API. Those features require verbatim mode. Live transcription supports smart formatting but offers neither word-level timestamps nor speaker diarization; it emits timestamps at the utterance level instead.

3. Independent Testing Places the Model Among the Most Accurate Hosted Systems

Google calls Gemini 3.5 Transcribe its most precise speech-to-text model to date and positions it as the successor to Chirp 3. On the multilingual FLEURS benchmark, Google reports a word error rate of 5.50% for streaming transcription and 5.04% for non-streaming processing across a selection of leading languages and locales.

Word error rate measures substitutions, deletions, and insertions relative to a reference transcript, with lower scores indicating fewer errors. The figures are benchmark-specific and should not be read as universal error rates for every accent, microphone, or acoustic environment.

Google separately says Artificial Analysis measured average error rates of 4.0% for streaming and 2.6% for non-streaming use. It also reports a 70% reduction in the time required to produce a final transcript compared with Chirp 3.

Artificial Analysis’s public non-streaming leaderboard confirms the 2.6% result. At publication time, it placed Gemini 3.5 Transcribe fifth for accuracy among the systems listed, behind models with results between 1.7% and 2.4%. Its test combines roughly eight hours of audio from conversational, parliamentary, and corporate-earnings datasets rather than relying on a single clean-speech corpus.

The same independent table reports a processing speed factor of approximately 82.5, meaning the service processed about 82.5 seconds of audio per second under that benchmark’s conditions. Artificial Analysis estimated a cost of $5 per 1,000 minutes, matching Google’s approximate blended non-streaming API price.

These measurements support Google’s claim that the model is competitive, but not that it is uniformly the most accurate service available. Production evaluations still need representative recordings, particularly where code-switching, specialized vocabulary, overlapping speakers, or noisy channels determine whether a transcript is usable.

4. Pricing Starts at About Half a Cent per Recorded Minute

Google’s paid Gemini Developer API pricing lists non-streaming audio input at $2 per million tokens, estimated as $0.003 per minute. Text output costs $12 per million tokens, estimated as another $0.002 per minute. Google therefore gives a blended estimate of approximately $0.005 for each minute of recorded audio.

Live transcription costs more. Audio input is priced at $3.50 per million tokens, or approximately $0.005 per minute, while emitted text costs $21 per million tokens, estimated at $0.004 per minute. Google calculates an effective combined price of about $0.009 per live minute.

Both variants have a free API tier. Google’s pricing table also says free-tier content may be used to improve its products, while paid-tier content is not used for that purpose under the listed terms. Teams processing confidential recordings should evaluate the applicable service tier, regional availability, and data-governance terms before deployment.

Neither transcription model supports batch, flex, or priority inference. Google Search grounding is also unavailable. This is a specialized audio-to-text service rather than a general Gemini model with transcription added as one promptable capability.

5. Google Is Embedding Transcription Into Dictation and Agent Interfaces

For consumers, Gemini 3.5 Transcribe is available in English through the Gemini app on macOS and through Rambler on Android in selected countries and languages. Rambler can clean up dictated text, correct misspellings, make edits, and change writing style through voice instructions.

In Google Antigravity, transcription can use permitted screen context and chat history to improve its handling of filenames, active documents, and agent output. Google AI Studio also exposes the model in Build mode so developers can dictate while creating applications.

The Gemini macOS app combines transcription with screen context and other Gemini models. Spoken requests can trigger workflows involving local-file summaries, text reuse across applications, searches, or image generation. Those actions are an integration feature: the dedicated transcription model converts speech into text, while the surrounding application routes the resulting command.

Google says talk-to-type support for arbitrary web fields is coming to Chrome, but it is not part of the currently available release. Enterprise access is in public preview through Gemini Enterprise Agent Platform, with support for Gemini Enterprise for Customer Experience planned separately.

The launch therefore changes more than recognition accuracy. Google now offers one transcription layer across consumer dictation, development environments, enterprise audio processing, and real-time voice interfaces. Developers can choose cleaned or literal output, but they must also choose between the live endpoint’s low latency and the recorded-audio endpoint’s speaker labels and word-level timing.

Frequently Asked Questions

Is Gemini 3.5 Transcribe generally available?

No. Google lists the developer and enterprise services as public previews.

How many languages does the model support?

It automatically detects and transcribes more than 85 languages, including language changes within an utterance or session.

Can the live model identify speakers?

No. Speaker diarization and word-level timestamps are available only for prerecorded audio in verbatim mode.

Does smart transcription preserve the speaker’s exact words?

No. It can remove filler, resolve corrections, restructure text, and change formatting. Use verbatim mode when exact wording matters.

How much does the API cost?

Google estimates approximately $0.005 per minute for non-streaming transcription and $0.009 per minute for live transcription on the paid tier.

Sources

Share

Share this article