Gemini 3.8 TTS Adds Voice Remixing: Edit Prebuilt or Designed Voices With Natural-Language Instructions Into New Reusable Voice IDs

Background and the Industry Pain Point

Teams building voice products often get stuck at "almost right": a prebuilt voice is close, but the accent, pacing or perceived age is off, while designing a new voice from scratch with Voice Design makes it hard to stay consistent with an existing brand voice. Gemini 3.8 TTS already supported prebuilt voices and Voice Design, but had no way to adjust an existing voice.

Architecture Highlights and Mechanism

According to the official Gemini API changelog (November 4) and the Voice remixing guide, the feature reuses the Voice Design endpoint POST /v1beta/voices (type=prompted, store=true) and adds a prompted.base_voice field for the source voice, with prompted.input describing the changes in natural language. The model applies these relative instructions to the base voice's acoustic profile, stores a new voice_ resource in the project and returns a 24kHz WAV preview, without modifying the original. The new ID can be passed as base_voice again to refine it step by step. Thor (@thorwebdev), who works on developer experience at Google DeepMind, also introduced the feature on X.

Benchmarks and Measured Results

This is an API feature update. Google published no benchmarks, latency or audio quality evaluations, and did not describe separate pricing; check the Gemini API pricing page for billing.

Getting Started for Developers

Try it interactively on the Generate speech page in Google AI Studio, or call client.voices.create() in the google-genai SDK with a base_voice such as "Puck" and an edit instruction. Use the returned voice_ ID in speech_config with gemini-3.8-flash-tts or gemini-3.8-flash-lite-tts. It suits audiobook narration, brand voices and multi-character dubbing where you need several variants of one voice.