External Publication
Visit Post

Trouble getting Realtime voices to sound naturally Mexican Spanish

OpenAI Developer Community May 20, 2026
Source
Hi everyone, I’m working on a voice agent using Realtime 2.0 in Spanish for users in Mexico. The agent understands and speaks Spanish well, and the wording is already localized for Mexico. Right now the prompt is not just asking for “Mexican Spanish”; it asks for a natural professional Mexican voice from central Mexico, with clear consonants, a crisp S in all positions, slightly reduced unstressed vowels, clear stressed vowels, moderate musical intonation, light emphasis on tonic syllables, a uniform/moderate rhythm, and connected speech instead of dry word-by-word pauses. I also tried being more specific by region, for example asking for a Bajío accent, a Yucatan accent, and even more specific references like D.F., Queretaro and Aguascalientes. The results changed a bit, but they still didn’t sound naturally Mexican or consistent enough for real calls. The issue is that the result still sounds somewhat generic or foreign. It is understandable, but the accent and intonation don’t really land as Mexican in a natural way. It feels like the prompt can influence wording and delivery a bit, but not the actual accent enough. A few things I’m trying to figure out: * Which voice has worked best for Mexican Spanish * Whether prompt instructions actually help with accent, or only with style * Any phrasing that improves pronunciation/intonation * Whether Custom Voices are the only realistic path for this kind of use case Not looking for perfect cloning or anything like that, just trying to make the agent sound more natural for Mexican callers. Thanks.

Discussion in the ATmosphere

Loading comments...