AI
Google's Gemini TTS Now Offers Self-Serve Voice Replication, Sidestepping OpenAI's Sales Gatekeeping
Google has opened up custom voice creation directly to developers via API, eliminating the manual sales process that OpenAI requires for voice customization.
The New Stack
Google has rolled out voice replication and design capabilities in Gemini 3.8 Flash TTS, giving developers direct self-serve access to create and manage custom voices—a stark contrast to OpenAI's sales-restricted approach.
How Voice Replication Works#
The new Voices endpoint (POST /v1beta/voices) requires two clean audio samples from 10 to 30 seconds each, plus a separate consent recording where the speaker confirms ownership. Google verifies that both the reference clip and consent recording belong to the same person before generating a voice ID that persists in the developer's project for up to a year.
Developers can store up to 200 custom voices per project and manage them through standard API operations. For temporary use cases, setting store=False returns an encrypted key that expires after a week and stays with the application. All synthesized audio carries Google's SynthID watermark, and replicated voices include C2PA content credentials for transparency tracking.
One important caveat: Google doesn't offer voice replication through AI Studio in Illinois, Texas, the European Economic Area, the U.K., Switzerland, or India.
Generating Voices From Description#
Rather than requiring recordings, developers can craft voices entirely through natural-language prompts by describing role, accent, and character traits. Google supports voice design across more than 100 languages and dialects, with a library exceeding 2,000 production-ready options. The platform provides 30 prebuilt studio voices and hundreds more that can be filtered by language, accent, pitch, and use case.
To avoid voice drift across requests, Google recommends creating a voice once and reusing its ID rather than re-describing the same persona repeatedly. A remixing feature for adjusting timbre, pitch, pace, and accent is coming soon.
How Prompting Works Now#
Gemini 3.8 treats input text as verbatim transcript—a breaking change from the 3.1 preview. Sustained vocal directions like whispering or sarcasm now use speech_metadata annotations, while momentary sounds (<sigh>, <cough>, <short pause>) go inline in angle brackets. For multi-speaker scripts, listener reactions wrapped in pipes (e.g., |mhm|) create backchannels without extra dialogue turns.
