Generative AI Customer Experience Platform for Low-Resourced Languages

Enterprises and public services often struggle to serve customers in underrepresented languages due to limitations in current voice AI systems. This project addresses these gaps by improving speech recognition and text-to-speech (TTS) capabilities for low-resourced languages. Building on Proto’s existing natural language models, the new capability will leverage LLMs, transformer-based speech recognition, and variational auto-encoder-based TTS to significantly reduce the word error rate (WER) from 25% to as low as 5 to 10% in target languages. The project will leverage voice actor recordings and knowledge extraction from documents and online sources, making it possible to train voice AI agents for underserved languages. This project aims to enable equitable, high-quality service for everyone, especially citizens from emerging markets and immigrant communities. The resulting deployments in countries such as the Philippines, Rwanda, and Namibia are supported by Proto's partners, including the Gates Foundation, central banks, and social impact innovators.