Many AI researchers believe AI models that listen and speak to people are trailing text-based models in capabilities and performance. That gap is even bigger when you look at audio models for languages other than English, especially outside the Western hemisphere. That could be a big issue for companies like OpenAI that are aggressively trying to expand internationally with personal AI devices controlled through voice . (As to why it’s so important for OpenAI to find new sources of ChatGPT users, check out yesterday’s newsletter here .) ChatGPT today supports more than 80 written languages and 37 spoken languages. But it’s still missing a lot of important languages. Take India, where today ChatGPT is used by 100 million people at least once a week, or nearly 11% of the chatbot’s 910 million total weekly active users, CEO Sam Altman said earlier this month. India has 22 official languages and hundreds more unofficial languages and dialects, many of which ChatGPT doesn’t yet support. And many transactions handled in India—from businesses ordering more inventory to planning logistics over WhatsApp voice memos—happen over the phone. In other words, people are doing those tasks by speaking out loud. To automate that work, ChatGPT must improve its ability to listen and speak out loud—and add support for the additional languages.