Gemini 3.5 Transcribereview, pricing and limits

Built for meeting notes, dictation, and multilingual support transcripts rather than general chat, Gemini 3.5 Transcribe replaces Google's Chirp 3 with a 70% faster time-to-final-transcript. A single batch call tops out around 60 minutes of audio (dropping to 30 minutes with diarization) and reliably identifies only 2 speakers before attribution turns experimental.

Gemini 3.5 Transcribe is a speech-to-text model that Google DeepMind put into public preview on August 26, 2026, automatically detecting and transcribing more than 85 languages. It differs from Google's general Gemini 3.5 chat models by only converting audio into text, adding speaker diarization, timestamps and disfluency cleanup along the way.

Provider: Google DeepMind · Family: Gemini 3.5

More about Google DeepMind on HokAI

Context window: 96,000 tokens

Input modalities: audio · Output: text

About Gemini 3.5 Transcribe

Gemini 3.5 Transcribe is Google's dedicated speech-to-text model, built by Google DeepMind and shown first at Google I/O on May 19, 2026 before entering public preview on August 26, 2026 with its own write-up on the Google blog. Unlike the general-purpose Gemini 3.5 chat models, it is a narrow specialist: it takes an audio file or stream and returns transcribed text, nothing else. It ships in two forms, gemini-3.5-transcribe for pre-recorded audio through the Interactions API and gemini-3.5-transcribe-live for real-time streaming through the Live API, and it directly succeeds Chirp 3, Google Cloud's prior transcription model from 2025. Google has not disclosed the model's architecture or parameter count, consistent with how it treats every Gemini-family model. On accuracy, Google reports a 2.6% average word error rate in non-streaming mode and 4.0% in streaming mode, as measured by Artificial Analysis; on the FLEURS multilingual benchmark it scores 5.04% WER non-streaming and 5.50% streaming. Google describes this as a major advancement over Chirp 3, with a meaningfully faster time-to-final-transcript. Third-party evaluations put its accuracy in the same tier as the leading open-weight ASR models rather than clearly ahead of them, though no independent lab has yet run a direct adversarial head-to-head against them. A single batch request has a per-request duration ceiling that shortens further once diarization or word-level timestamps are requested. Google's published token-rate assumptions, roughly 25 audio tokens per second of input and about 175 text tokens per minute of output, imply a working context in the tens of thousands of tokens for a long recording, though Google has not published an exact context-window figure the way it does for its chat models. The model's only input modality is audio; its only output is text. It picks up whichever language is being spoken without a manual setting, including switching languages mid-recording, and supports two modes: verbatim, which preserves every filler word and false start, and smart, which strips repetitions and self-corrections and auto-formats the result. It separates individual speakers within a recording, generates word-level timestamps in MM:SS format, accepts custom-vocabulary phrases to bias recognition toward names and jargon, and can hand off to other Gemini models inside the macOS Gemini app. Google prices the batch endpoint per token at a lower rate than the real-time streaming endpoint, with both billed by audio tokens in and text tokens out. Both endpoints include a free tier in Google AI Studio, and Vertex AI costs a bit more for the same model there. See the pricing FAQ below for the exact per-token and per-minute figures. Developers reach the model through the Gemini API in Google AI Studio, through Google Cloud's Gemini Enterprise Agent Platform, and through Vertex AI, with official client libraries covering both major scripting languages developers use for the Gemini API. On the consumer and workspace side it already powers Rambler dictation in Gboard on Android, the Gemini app on macOS, and Search Live, with a Chrome integration announced as coming soon and enterprise customer-experience support in progress. Google states that its Gemini models, including its audio family, go through training-time evaluations and pre-release ethics and safety reviews carried out with its internal safety and responsibility teams, consistent with its AI Principles. Google has not published a dedicated model card or system card specifically for Gemini 3.5 Transcribe as of this writing, and no separate training-data cutoff date has been disclosed for the model. It fits teams building dictation, meeting-notes, or multilingual support-transcription features who want disfluency cleanup and formatting handled in the same call as the transcription, without a second LLM pass. It is a weaker fit for anyone who needs a verified, audited speaker count beyond a couple of participants, since Google itself labels three-plus speaker attribution experimental, and it is not an option for teams that need to self-host or fine-tune an ASR model, since it ships as a closed, API-only model with no open weights. As the successor to Chirp 3, Gemini 3.5 Transcribe was previewed at Google I/O on May 19, 2026 and formally launched into public preview on August 26, 2026 alongside its gemini-3.5-transcribe-live sibling. Google has not announced a retirement date for Chirp 3 or a further version beyond this preview as of the model's launch week.

Pricing

Standard tier: $2/$12 per 1M input/output tokens (~$0.005/min blended). Live tier: $3.50/$21 per 1M input/output (~$0.009/min blended). Vertex AI adds a 10-20% premium over AI Studio for the same model.

Key Features

• Speaker Diarization: Attributes speech to up to 8 distinct speakers per recording, though Google flags attribution for 3 or more speakers as experimental.
• Word-Level Timestamps: Generates MM:SS timestamps for individual words, useful for captioning and search-by-word, at some cost to measured accuracy versus running without them.
• Smart Transcription Mode: Strips filler words, repetitions and self-corrections and auto-formats the output, as an alternative to the verbatim mode that preserves every spoken word.
• Custom Vocabulary: Accepts up to 1,000 domain-specific phrases to bias recognition toward names, acronyms and jargon, though Google recommends around 100 terms for best results.
• Broad Language Auto-Detection: Automatically detects and transcribes dozens of spoken languages and locales, including mid-recording code-switching, without a manual language setting.

Pros

• Cuts time to final transcript by 70% versus Google's prior Chirp 3 model.
• Bundles transcription, disfluency cleanup, diarization and timestamps into one API call instead of a transcribe-then-clean pipeline.
• Auto-detects the spoken language and handles mid-recording code-switching without a manual language setting.

Cons

• Speaker attribution is explicitly experimental past 2 speakers, so diarization accuracy is unproven for panels or large calls.
• Batch audio is capped at 1 hour per request, dropping to 30 minutes once diarization or timestamps are enabled.
• Proprietary and API-only: there are no open weights, so it cannot be self-hosted or fine-tuned.

Benchmarks

• wer streaming pct: 4
• wer non streaming pct: 2.6
• wer fleurs streaming pct: 5.5
• wer fleurs non streaming pct: 5.04
• latency improvement vs chirp3 pct: 70

Visit Gemini 3.5 Transcribe Official Page