Gemini API: Voice Editing Arrives in Gemini 3.5 Transcribe

Gemini CLIView original changelog

Google's Gemini API now lets the gemini-3.5-transcribe and gemini-3.5-transcribe-live models edit, rewrite and format speech. The capability works on dictated speech or on existing text passed in as source_text. Developers building dictation and voice-driven writing tools can use it without a separate text-editing step.

Key Takeaways

  • Voice editing is now available in both Gemini 3.5 Transcribe models, so speech can be edited, rewritten and formatted rather than only transcribed.
  • The feature works on dictated speech as well as on existing text supplied through the source_text input.
  • Both the recorded-audio model (gemini-3.5-transcribe) and the live streaming model (gemini-3.5-transcribe-live) support it.
  • Developers can fold editing and transcription into one step instead of chaining a separate text-editing model.
  • The change suits dictation and voice-driven writing tools where users speak corrections or formatting instructions.
  • Google directs developers to the Audio transcription and Live transcription guides for implementation details.

Voice Editing Comes to Gemini 3.5 Transcribe

Google added a voice editing capability to the Gemini 3.5 Transcribe models in the Gemini API. Both gemini-3.5-transcribe and gemini-3.5-transcribe-live can now edit, rewrite and format speech, which extends them beyond plain speech-to-text into a tool for shaping text by voice.

What the Models Can Now Do

The new capability applies to two kinds of input. The first is dictated speech, where the model takes what a person says and edits, rewrites or formats it. The second is existing text passed in through the source_text field, which lets an application hand the model a draft and have the model change it according to spoken instructions.

Gemini 3.5 Transcribe is offered through two endpoints. The standard model, gemini-3.5-transcribe, handles recorded audio. The live model, gemini-3.5-transcribe-live, handles streaming audio. Voice editing is available on both, so apps can use it on a recorded clip or in a real-time session.

Why It Matters for Developers

Dictation tools have usually needed two stages: transcribe the speech, then send the text to a separate model to clean it up or apply changes. With voice editing built into the transcription models, an application can ask for editing behavior in the same step. This suits dictation features, note-taking apps and voice-driven writing assistants, where a user may speak a correction or a formatting instruction and expect the text to change.

Where to Learn More

Google's release notes point developers to the Audio transcription and Live transcription guides in the Gemini API documentation for usage details.