Google opens Gemini 3.8 voice replication through the API, while OpenAI still requires calling sales for the same

A team building a voice application or an audiobook pipeline has, for the past year, had two choices to create a custom voice: pay a specialised synthesis vendor or go through OpenAI's custom voice sales process. Google DeepMind's Gemini 3.8 text-to-speech launch on 2026-09-23 opens a third: a self-serve POST /v1beta/voices endpoint that turns a short recording into a voice ID and keeps it in the developer's project.

What shipped

Google released two models. Gemini 3.8 Flash TTS is the higher-quality model, aimed at what the launch calls "deep creative direction and character design": complex dialogue, vocal tags, regional dialects and long narration. Gemini 3.8 Flash-Lite TTS is the cheaper, faster option and the direct replacement for the earlier gemini-3.1-flash-tts-preview, tuned for bulk production and voice agents. The two share an API schema, so moving between them is a one-parameter change, and both support voice design and replication.

Voice design generates a persona from a written description of role, accent and character. Google says it works across more than 100 languages and dialects, and lists 130 supported languages for Flash TTS and 101 for Flash-Lite. The launch also claims a library of more than 2,000 production-ready voices.

The replication endpoint

Amanda Caswell at The New Stack, in a detailed writeup on 2026-09-24, covers the replication API. Two clean 10-to-30-second recordings from the same speaker are required, plus a separate consent recording in which the speaker reads a statement agreeing to a synthetic version being created. Once approved, Google returns a voice_... ID and keeps it in the project for a year alongside any designed voices. A project can hold up to 200 voices in total.

Voice replication can also run without storing the profile: setting store=False returns an encrypted voicekey_... instead, which the application supplies each time and which expires after seven days. The launch says Google marks generated audio with SynthID and adds C2PA content credentials, and Caswell notes voice replication is not offered through AI Studio in Illinois, Texas, the European Economic Area, the United Kingdom, Switzerland or India.

Where OpenAI sits

For comparison, Caswell writes that OpenAI also offers custom voices, but "access is tighter": customers have to go through sales, are limited to 20 voices per organisation and must provide a consent recording alongside a voice sample of up to 30 seconds. OpenAI does not offer prompt-based voice design from a written description at all; its 13 built-in voices can be steered for tone or speed and applications must disclose the audio is AI-generated.

Two things to watch when adopting it

Gemini 3.8 treats the input text as a strict verbatim transcript, which Google's docs describe as a breaking change from the 3.1 preview model: stage directions embedded in prompts no longer work, and sustained direction now belongs in a speech_metadata annotation while momentary sounds sit inline in angle brackets. Native two-speaker generation is limited to prebuilt voices; dialogue between designed or replicated voices has to be generated turn by turn and stitched from the 24 kHz PCM output.

Google recommends creating a voice once and reusing its ID, because the docs say repeatedly sending long persona descriptions is the most common cause of voice drift. API access through Gemini Enterprise is listed as coming soon.

Source

• Gemini 3.8 text-to-speech says hello, Google DeepMind, 2026-09-23
• OpenAI makes you call sales for a custom voice. Google just made it self-serve., Amanda Caswell, The New Stack, 2026-09-24