Gemini 3.5 Transcribe: Features, API, and Use Cases

Google introduced Gemini 3.5 Transcribe on August 26, 2026 for recorded and live speech-to-text. gemini-3.5-transcribe processes audio files, while gemini-3.5-transcribe-live streams incremental text from live audio. Google’s Gemini API release notes list both models as generally available.

What did Google announce with Gemini 3.5 Transcribe?

In its launch announcement, Google calls Gemini 3.5 Transcribe its most precise speech-to-text model yet. It targets filler words, self-corrections, accents, background noise, language switching, and specialized terms.

The file model can return speaker labels and word-level timestamps. The Live model prioritizes low-latency output and does not currently provide either feature.

Gemini 3.5 Transcribe features that matter

Smart transcription and verbatim transcription

Smart mode removes filler words, resolves self-corrections, adds punctuation, and formats dictated content. Verbatim mode preserves repetitions and false starts, making it better for research, legal review, and other workflows that need a closer record of the audio.

Automatic detection across 85+ languages

The models support more than 85 languages, regional accents, and language switching. They can detect languages automatically or accept BCP-47 hints such as en-US and es-ES.

Real-world audio and custom vocabulary

Google highlights noisy-audio performance and recognition of details such as postal codes and order IDs. Citing Artificial Analysis, it reports average word error rates of 2.6% for file transcription and 4.0% for streaming. Results still depend on audio quality, language, accent, overlapping speech, and vocabulary.

Developers can bias recognition toward names, acronyms, jargon, and unusual spellings with custom_vocabulary. The API accepts up to 1,000 terms; Google recommends a focused list of up to 100 for the best results.

Speaker diarization and word-level timestamps

The file model can label up to eight speakers, although attribution for three or more is experimental. It also supports word-level timestamps. Both features require verbatim mode, and Google notes that timestamps may reduce accuracy.

Where can Gemini 3.5 Transcribe be used?

• Meetings and interviews: create searchable notes, separate speakers, and link text to audio with timestamps.
• Voice notes and dictation: turn speech with pauses and corrections into formatted text.
• Customer support: recognize product names and identifiers, then index or analyze call transcripts.
• Live captions: deliver interim and final text for events, classrooms, broadcasts, or voice interfaces.
• Multilingual church services: transcribe the source speech before translating captions for each listener.

Sensitive, legal, medical, and accessibility use cases still need appropriate consent, review, retention controls, and testing with real audio.

How do you call the Gemini 3.5 Transcribe API?

For a recording, install Google’s Gen AI SDK, set GEMINI_API_KEY on the server, upload the audio, and pass its URI to the Interactions API.

npm install @google/genai

import { GoogleGenAI } from "@google/genai";

const client = new GoogleGenAI({});

const audioFile = await client.files.upload({
  file: "./meeting.mp3",
  config: { mime_type: "audio/mp3" },
});

const interaction = await client.interactions.create({
  model: "gemini-3.5-transcribe",
  input: [
    {
      type: "audio",
      uri: audioFile.uri,
      mime_type: audioFile.mimeType,
    },
  ],
  generation_config: {
    transcription_config: {
      language_codes: [],
      custom_vocabulary: ["SpeedSound", "Gemini"],
      mode: { type: "smart" },
    },
  },
});

console.log(interaction.output_text);

This enables automatic language detection and smart cleanup. For diarization or timestamps, use verbatim mode with diarization_mode: "speaker" or timestamp_granularities: ["word"].

See Google’s audio transcription guide for more options. Microphone audio should use the Live transcription guide and gemini-3.5-transcribe-live. Keep API keys server-side, validate uploads, and delete temporary files after processing.

How SpeedSound uses Gemini 3.5 Transcribe

The SpeedSound Audio to Text tool uses Gemini 3.5 Transcribe as its default model. Users can upload audio, enable automatic detection or add a language hint, choose clean or verbatim output, and copy or download the transcript. SpeedSound deletes the temporary Gemini upload after processing.

SpeedSound Live covers the related church use case by transcribing speech, translating it, and delivering captions to guests’ phones. Its current production speech-recognition path is separate from the uploaded-audio Gemini integration; it should not be presented as the same implementation.

What are the current limitations?

• File transcription supports up to one hour, or 30 minutes with diarization or timestamps.
• Live sessions support up to 10 minutes and need continuity logic for longer events.
• Smart mode cannot be combined with diarization or word timestamps.
• Live mode does not currently support diarization or word timestamps.
• Standalone Gemini 3.5 Transcribe returns text; other actions require separate models or orchestration.

Gemini 3.5 Transcribe FAQ

Is Gemini 3.5 Transcribe available through an API?

Yes. Use the Interactions API for files and the Live API for streaming audio.

Does Gemini 3.5 Transcribe work in real time?

Yes. Real-time transcription uses gemini-3.5-transcribe-live over a bidirectional WebSocket.

What is the difference between smart and verbatim mode?

Smart mode cleans and formats speech. Verbatim mode stays closer to the audio and is required for diarization and timestamps.

How many languages does Gemini 3.5 Transcribe support?

More than 85, with automatic detection, regional accents, and language switching. Quality varies by audio and language.

Can it identify multiple speakers?

The file model can label up to eight speakers; three or more remains experimental. Live mode does not support diarization.

Can Gemini 3.5 Transcribe handle specialized terminology?

Yes. Custom vocabulary can bias recognition toward names, jargon, acronyms, and uncommon spellings.

Can I try Gemini 3.5 Transcribe without building an API integration?

Yes. Upload a recording with the SpeedSound Audio to Text tool. For church captions, see SpeedSound Live.