Your voice service reads a reference number and a time. A pleasant voice can still make adjacent digits hard to hear. On October 16, 2026, OpenAI released voice generation from descriptions to API customers. Saved voices can be reused in GPT-Live. Prompt-based voice creation is free according to the release. Speech generation follows session billing. The separate process using a real sample requires a consent recording and a voice sample from the speaker.
Test with your own data
Create two descriptions without identifying a real person, one clear with brief pauses and one calm with a regular pace. Use a standard voice as a baseline. Have each read 20 fictional phrases with dates, letters and similar values. Randomize order and ask listeners to transcribe key details on headphones and on speakers with constant background noise. Compare exact words and digits, repetitions and listening time.
Making a decision
Keep the description, ID and date for the chosen voice. Disclose that it is synthetic and rerun the test when the voice or model changes. Compare cost per completed call, including the time and potential cost of repeated speech.