Get API access
Model release

Gemini 3.5 Transcribe Reaches General Availability

Google released Gemini 3.5 Transcribe on August 26, 2026. This non-streaming speech-to-text model supports 85+ languages, diarization, and custom vocabulary…

Released Published 3 min readModel evidence ↗
Gemini 3.5 Transcribe release illustration — Google

Official Release Details

The source event date is August 26, 2026. On this date, Google confirmed that Gemini 3.5 Transcribe is generally available. This release marks the deployment of a dedicated speech-to-text model based on Gemini's audio understanding. The model is identified as gemini-3.5-transcribe within the official API documentation.

  • Provider: Google
  • Model Name: Gemini 3.5 Transcribe
  • Input Modality: Audio
  • Output Modality: Text
  • Context Window: 98,304 tokens
  • Max Output Tokens: 32,768

Documented Capabilities

According to the official changelog, the model offers high-accuracy and low-latency performance in a non-streaming format. It includes utterance-based language detection supporting more than 85 languages. Additionally, the system provides speaker diarization and word-level timestamps. Users can apply custom vocabulary biasing with a limit of up to 1,000 terms. Open weights are not confirmed for this release.

  • Non-streaming speech-to-text processing
  • Utterance-based language detection (85+ languages)
  • Speaker diarization support
  • Word-level timestamp generation
  • Custom vocabulary biasing (max 1,000 terms)

Evaluation Procedure

Engineers should validate the model against specific use cases using the following steps. First, test the model with audio samples containing multiple speakers to measure diarization accuracy. Second, compare the generated word-level timestamps against manual annotations to assess precision. Third, validate language detection by submitting audio clips in different languages from the supported list. Finally, measure the impact of custom vocabulary biasing by comparing transcription results with and without the 1,000-term limit applied. These proposed tests allow for objective assessment without assuming unverified benefits.

Primary sources

QUICK ANSWERS

Frequently asked questions

What is the maximum number of terms allowed for custom vocabulary biasing?

The source states that custom vocabulary biasing supports up to 1,000 terms.

Does the model support streaming input?

No, the source describes the model as providing non-streaming speech-to-text.