Build with Gemini 3.5 Transcribe: live captions, recorded audio and smart speech-to-text

What changed

Google has introduced Gemini 3.5 Transcribe as its latest speech-to-text model, positioning it as a Gemini-native alternative to conventional ASR pipelines for developers building voice interfaces, captions and post-call analytics [1]. The model is available in public preview for developers through the Gemini API in Google AI Studio and Google Antigravity, while enterprises can access it through the Gemini Enterprise Agent Platform [1]. Google says the model is already powering consumer voice experiences including Rambler on Android and the Gemini app on macOS, with Chrome support planned next [1].

The key product shift is that transcription is no longer framed as a narrow speech-recognition utility. Gemini 3.5 Transcribe is described as an LLM-based transcription model that converts raw audio into polished, formatted text and is designed to handle noisy audio, jargon, self-corrections and disfluencies [1]. For builders, that means the first design decision is not “which speech API should I call?” but “which transcript do I need: a literal record, a readable document, or an incremental stream?”

Choose the right API surface

There are two Gemini API paths to consider. For recorded audio, use gemini-3.5-transcribe through the Interactions API; for real-time streaming, use gemini-3.5-transcribe-live through the Live API [1]. The model page confirms those two versions: a unary gemini-3.5-transcribe endpoint and a Live API gemini-3.5-transcribe-live endpoint [2].

The recorded-audio path supports files up to one hour, but the limit drops to 30 minutes when speaker diarization or word-level timestamps are enabled [2]. Live transcription sessions support up to 10 minutes and do not support word-level timestamps or speaker diarization, so they are best for captions, voice input fields and real-time UX rather than forensic transcripts [2]. The model page also states that the Gemini Developer API model does not support the Batch API, Flex inference or Priority inference, so teams should distinguish “file transcription” from true batch infrastructure [2].

On Google Cloud’s Gemini Enterprise Agent Platform, Google documents preview-suffixed model IDs: gemini-3.5-transcribe-preview for synchronous transcription and gemini-3.5-transcribe-live-preview for streaming [6]. That enterprise page lists global availability and a 15-minute file-processing limit in its preview documentation, which is narrower than the Gemini Developer API’s one-hour file limit, so teams should validate limits against the exact platform they deploy on [6].

Build the recorded-audio pipeline

A practical recorded-audio pipeline starts with file ingestion. Google’s audio transcription guide shows the expected pattern: upload an audio file, pass the returned file URI and MIME type into an Interactions API request, and set the model to gemini-3.5-transcribe [3]. For short internal tools, that can be enough: upload the meeting recording, call the model, store output_text, and run downstream summarization or extraction separately.

Production systems should decide early between verbatim and smart modes. The default verbatim mode preserves filler words, repetitions, pauses and false starts, while smart mode removes disfluencies, resolves inline corrections and formats items such as lists, dates, currencies and numbers [3]. That is a meaningful product choice: a legal or compliance transcript may need verbatim output, while a CRM call note or clinic intake summary may benefit from smart formatting.

Speaker diarization and timestamps belong on the recorded path. Google’s guide says speaker diarization labels distinct speakers, word-level timestamps provide start and end offsets, and both are configured inside verbatim mode [3]. Smart transcription cannot be combined with timestamp granularities or diarization mode, so a robust app may need two workflows: one for a faithful, annotated record and another for a cleaned reading copy [3].

Custom vocabulary should be treated like a configuration asset. Gemini 3.5 Transcribe can accept up to 1,000 custom terms, acronyms or proper names, but Google says best results typically come from up to 100 targeted terms [2]. In practice, that means shipping vocabulary lists per customer, department or use case: drug names for healthcare, SKU names for support, ticker symbols for finance, and internal project codenames for enterprise meetings.

Build the live experience

Live transcription is a different system. Google’s Live API guide says developers can connect over WebSockets or the Google Gen AI SDK, stream continuous audio input and receive incremental text transcriptions as speech occurs [4]. The dedicated Live Transcription mode uses response_modalities=["TEXT"], distinguishes itself from a live conversational agent, and treats the input stream as raw 16-bit PCM audio [4].

That distinction matters. A live agent listens, reasons and may speak back; Live Transcription is a low-latency speech-to-text pipeline that emits streaming text [4]. If you are building captions, dictation, moderation previews or live meeting overlays, you likely want Live Transcription. If you are building a voice assistant that calls tools and responds in audio, you likely want the broader Live Agent pattern, not the dedicated transcription pipeline.

The Live API exposes interim and final results. Google documents server_content.interim_input_transcription for partial hypotheses while a user is speaking and server_content.input_transcription for the finalized transcript after a speech turn ends [4]. A good UI should visibly treat those states differently: interim text can appear in a lighter style, while final text can be committed to the transcript, indexed and sent to downstream models.

Live mode also supports automatic language detection, custom vocabulary and smart transcription [4]. It does not support speaker diarization, and word-level timestamps are not available over Live API sessions [4]. That creates a common hybrid architecture: use Live API for the in-the-moment experience, then run the stored audio through recorded transcription later if speaker labels, precise timestamps or audit records are required.

Accuracy, language and formatting

Google cites Artificial Analysis measurements of 4.0% average word error rate for streaming and 2.6% for non-streaming use cases, plus a 70% improvement in time to final transcription compared with Chirp 3 [1]. On the FLEURS benchmark across selected top languages and locales, Google reports 5.50% WER in streaming mode and 5.04% in non-streaming mode [1]. Those numbers are useful for screening, but production teams should test their own audio because microphone quality, accents, turn-taking and domain vocabulary often dominate real-world performance.

Language support is broad. The Gemini API model page says Gemini 3.5 Transcribe automatically detects more than 85 languages, handles code-switching and supports BCP-47 language codes such as en-US, fr-FR, de-DE, es-US, pt-BR, ja-JP and many others [2]. Google’s transcription guide says omitting language_codes or passing an empty array enables automatic detection, while known language hints can improve accuracy [3].

Smart formatting is one of the model’s most product-facing capabilities. Google says the model can clean up fillers, handle self-corrections and apply structured formatting, while the developer guide describes inverse text normalization such as turning spoken currency into compact written forms [1][3]. That is valuable for forms, notes and command input, but it also means builders should expose a “raw transcript” option in contexts where exact wording matters.

Pricing and adoption plan

Google’s pricing page lists gemini-3.5-transcribe at $2.00 per 1 million input tokens, estimated at $0.003 per audio minute, and $12.00 per 1 million output tokens, estimated at $0.002 per text minute, for an effective blended rate of about $0.005 per minute [5]. The live model is priced higher: gemini-3.5-transcribe-live is listed at $3.50 per 1 million input tokens, estimated at $0.005 per audio minute, and $21.00 per 1 million output tokens, estimated at $0.004 per text minute, for an effective blended rate of about $0.009 per minute [5].

The sensible build plan is therefore staged. First, benchmark the recorded endpoint on real examples and compare verbatim, smart, vocabulary and diarization settings. Second, prototype Live API streaming only where latency changes the product experience. Third, design retention, consent and review workflows before transcripts flow into search indexes, CRMs or agent tools. Gemini 3.5 Transcribe gives developers a cleaner speech layer, but the winning products will be the ones that choose the right transcript for the job.

Sources from the last 72 hours

1. [1]Intelligent transcription with Gemini 3.5 TranscribeAug 26, 2026, 12:00 AM UTC
2. [2]Gemini 3.5 Transcribe | Gemini API | Google AI for DevelopersAug 26, 2026, 12:00 AM UTC
3. [3]Audio transcription | Gemini API | Google AI for DevelopersAug 26, 2026, 12:00 AM UTC
4. [4]Live transcription with Gemini Live API | Gemini API | Google AI for DevelopersAug 26, 2026, 12:00 AM UTC
5. [5]Gemini Developer API pricing | Gemini API | Google AI for DevelopersAug 26, 2026, 12:00 AM UTC
6. [6]Gemini 3.5 Transcribe | Gemini Enterprise Agent Platform | Google Cloud DocumentationAug 26, 2026, 12:00 AM UTC

AI-generated article based on recent web research, then preserved as a dated editorial snapshot.