
Research from Ampixa on speech and text technology for Nepali and the languages of Nepal, built to run on the device.
Ampixa. Deterministic pronunciation frontends and text-to-speech models for Limbu, Magar, Bantawa, Gurung, Dhimal and Nepali. These languages share Devanagari but not Nepali's sounds, and a generic Nepali frontend erases the contrasts that tell their words apart. The paper introduces a contrast-preservation test that needs no full pronunciation dictionary.
A benchmark of Nepali text-to-speech systems rated by native speakers, and a predictor of listener scores.
Neural text-to-speech small enough to run faster than real time on a microcontroller, in a browser or on an ordinary CPU.
Small models for recognising handwritten Devanagari.