---
type: "post"
id: "prompting-gpt-live-for-languages-and-regional-accents-25c589"
slug: "prompting-gpt-live-for-languages-and-regional-accents-25c589"
title: "Prompting GPT-Live for languages and regional accents"
description: "OpenAI’s GPT-Live guide shows how to shape multilingual voice agents through voice selection, regional voice design, session instructions and application-specific listening tests."
retrieval_nugget: "OpenAI’s GPT-Live guide shows how to shape multilingual voice agents through voice selection, regional voice design, session instructions and application-specific listening tests. For teams building speech interfaces, its practical point is that accent, language switching, number handling and conversational policy need separate design choices and repeatable checks—not one generic instruction to sound natural."
published_at: "2026-11-20"
updated_at: "2026-11-20"
record_date: "2026-11-20"
date_kind: "published_at"
topics: ["prompts","multimodal"]
entities: ["OpenAI","GPT-Live","Nira","Bossa","Tempo"]
related_patterns: ["model-answers-become-task-interfaces"]
novelty_label: "new_fact"
what_changed: null
prior_refs: []
source_url: "https://developers.openai.com/blog/prompting-multilingual-voice-agents"
source_urls: ["https://developers.openai.com/blog/prompting-multilingual-voice-agents"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:prompting-gpt-live-for-languages-and-regional-accents-25c589"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/prompting-gpt-live-for-languages-and-regional-accents-25c589/","markdown":"https://newruntime.com/posts/prompting-gpt-live-for-languages-and-regional-accents-25c589.md","json":"https://newruntime.com/posts/prompting-gpt-live-for-languages-and-regional-accents-25c589.json"}
markdown_url: "https://newruntime.com/posts/prompting-gpt-live-for-languages-and-regional-accents-25c589.md"
---

# Prompting GPT-Live for languages and regional accents

## Retrieval answer

OpenAI’s GPT-Live guide shows how to shape multilingual voice agents through voice selection, regional voice design, session instructions and application-specific listening tests. For teams building speech interfaces, its practical point is that accent, language switching, number handling and conversational policy need separate design choices and repeatable checks—not one generic instruction to sound natural.

## Summary
OpenAI’s GPT-Live guide shows how to shape multilingual voice agents through voice selection, regional voice design, session instructions and application-specific listening tests. For teams building speech interfaces, its practical point is that accent, language switching, number handling and conversational policy need separate design choices and repeatable checks—not one generic instruction to sound natural.

## What happened
OpenAI’s guide begins with existing GPT-Live voices. It recommends testing built-in options before creating a new voice: Bossa and Tempo have Brazilian Portuguese influence, while Nira is a Hindi voice. The guide’s catalog also maps voices to regional influences and presentation, including Australian, British, Irish, North American, Filipino and Southern U.S. English, as well as several other languages. API voice names should be lowercase, and the current catalog is in OpenAI’s voice-options guide.

For a more specific sound, OpenAI recommends prompt-to-voice, with a description that names the region alongside pitch, pace and vocal character. Its Spanish example specifies a standard central/northern Peninsular Castilian accent, including native vowels, consonants, stress and sentence melody, and pairs that description with a Spanish `script_hint` containing ordinary connected speech. The guide suggests generating candidates, choosing a saved voice and testing it on a fresh sentence not used in the creation script. For bilingual voices, the script hint should include both languages. It also gives an Emirati Arabic example and says voice design should focus on sound; decisions such as which dialect to use for a caller or when to switch to English belong in session instructions.

The guide treats evaluation as a separate step. It proposes a small listening set of fresh sentences, realistic conversations and application-specific vocabulary or values. Checks include pronunciation in full context, regional number conventions such as lakh and crore in India, spelling and correcting an identifier, and mixed-language turns with product names and prices. For dates, amounts and identifiers, OpenAI says to validate the value sent to the backend as well as what the caller hears; ambiguous dates such as `04/05` may merit clarification and reading the month aloud.

For session behavior, OpenAI says to select the voice when creating the GPT-Live session; a saved generated voice uses `session.audio.output.voice.id`. Put role, language policy, delivery and conversation behavior in `instructions`, reinforcing the voice description. Its examples specify language switching rules, polite Hindi register, and a Spanish shopping assistant that asks one question at a time and does not confirm prices or order changes before system verification. For multilingual agents, the team should decide what triggers a language switch independently of the desired accent. OpenAI cautions that prompts can steer an accent but cannot guarantee it in every reply, and recommends testing both directions of a switch, including after reconnection.

OpenAI also sets boundaries around voice creation. Recording-based custom voices are limited to eligible customers, require express speaker consent and the rights needed for the recording and resulting voice, and require the consent recording and reference sample to come from the same speaker. Voice creation is ineligible for Zero Data Retention. The guide says listeners should be told when they are hearing AI-generated speech and warns against impersonating real people.

## What the archive adds
New Runtime covered OpenAI’s custom-voices guide on [October 22, 2026](https://newruntime.com/posts/openai-s-custom-voices-documentation-describes-prompt-to-dd276d/), describing voice identity as a configurable API resource. This is a new practical layer on that story: the GPT-Live guide explains how to specify regional characteristics and test them in application conversations.

## Trend signal
This is a concrete example of [model answers becoming task interfaces](https://newruntime.com/patterns/model-answers-become-task-interfaces/): a voice agent’s output is shaped for a particular interaction, with rules for language, register, clarification and backend-confirmed actions. The guide also emphasizes that a voice’s surface qualities do not define the full task contract; session instructions and application checks do.

## Why now
As teams move from voice demos to customer-facing assistants, regional speech and accurate handling of names, prices and identifiers become product requirements rather than polish. OpenAI’s guide addresses that shift with examples spanning shopping support and multilingual calls, while acknowledging that prompts alone do not ensure consistent accent behavior.

## Why it matters
Voice agents can sound fluent while still producing a poor or risky interaction: a caller may hear the wrong currency convention, have a name mispronounced, or receive an unverified price as fact. Separating voice design from conversation policy gives teams clearer places to diagnose those failures. A regional voice description can target sound; session instructions can define language switching and clarification; backend validation can protect values that must remain exact.

The guidance is especially useful for teams serving multiple locales. Rather than treating “speak Spanish” or “sound professional” as sufficient, they can define the intended variety and register, then test realistic phrases and edge cases. Consent and disclosure requirements also make voice choice an operational decision, not just an audio preference. The remaining work is empirical: test the specific voice, prompt, session behavior and application data path that customers will encounter.

## Limits and uncertainty
OpenAI presents examples and recommended practices, not measured results showing that a particular prompt reliably produces a specific accent or improves task success. The guide explicitly says accent steering is not guaranteed, including after language switches and session reconnections. It does not quantify error rates for pronunciation, date interpretation, identifier correction or mixed-language speech. Teams still need to validate those behaviors in their own application and confirm supported models and accents before using recording-based voices. Voice creation’s exclusion from Zero Data Retention is an important stated constraint; the guide points readers to separate data controls documentation for session settings.

## Skill to build now
Build a small listening set from real application tasks before tuning the voice: include names, product terms, local currency and number formats, ambiguous dates, and a caller correcting one character in an identifier. For each case, check both the spoken reply and the value your backend receives. Keep accent and language-switch rules separate in the voice description and session instructions, then rerun the same set after changing either.

## Reader takeaway
A convincing voice is not a complete voice agent. Choose or design the regional sound, specify conversation and switching behavior separately, and test the exact names, numbers and corrections your product handles—checking backend values as well as what users hear.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Prompts](https://newruntime.com/topics/prompts.md): Explore the prompts topic hub.
- topic hub: [Multimodal](https://newruntime.com/topics/multimodal.md): Explore the multimodal topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)
