Original MacDigest illustration. The waveform is a visual metaphor, not recorded or generated audio.
OpenAI’s prompt-to-voice lets developers describe, save and reuse an application’s voice without making a recording. The company released it to all API customers on October 16, 2026, through the API and voice Playground. The key constraint: prompt-generated voices work with GPT-Live only.
Check compatibility first
OpenAI’s custom-voices guide excludes the Text to Speech, Realtime and audio Chat Completions APIs. Recording-based custom voices remain a separate workflow with eligibility and speaker-consent requirements.
The GPT-Live session reference accepts a custom voice ID in session.audio.output.voice. Choose it when creating the session; switching voices requires a new session.
Budget for listening and reuse
Creating a prompt-generated voice is free. Speech generation incurs regular GPT-Live charges, and listening in the Playground runs a billable session.
Organizations can save up to 100 prompt-generated voices, including those an application creates for customers. Reuse saved IDs when designing customer-facing personalization.
Review retention before deployment
OpenAI’s data-controls table marks /v1/audio/voices as ineligible for Zero Data Retention, with application state retained until deletion. GPT-Live sessions are separately ZDR-eligible with limitations; ZDR disables stored recordings, downloads and forking. Check both stages before promising a retention policy.
OpenAI also requires adult voices and disclosure that speech is AI-generated, and prohibits impersonation.
Start with your existing audio endpoint, retention requirements and expected voice count. Then evaluate pronunciation and cost in your own application. This report is documentation-based; MacDigest has not tested the feature.
