Teaching a Computer to Read the News in Nepali

For my final year project at IOE Pulchowk Campus, my teammates Mausam Kumar Sah, Rohit Joshi, Yugal Pariyar and I decided to tackle a problem that felt almost embarrassingly overlooked. Nepali is spoken by tens of millions of people, and yet there was almost no natural sounding speech synthesis built for it. Under the supervision of Dr. Arun Kumar Timalsina, and in collaboration with Prixa Technologies, we set out to change that.

Where We Started

The idea itself was simple to explain: a browser extension that scrapes Nepali text from a news website and reads it aloud as you browse. What sat behind that simple idea was a full speech pipeline. We trained a Tacotron2 model to turn text into a spectrogram, then paired it with a DeepGAN vocoder to turn that spectrogram into sound a person could actually listen to.

Building a Voice From Almost Nothing

Data turned out to be the hardest part of the whole project. Good Nepali speech datasets are rare, so we combined the OpenSLR Nepali dataset with our own recordings, close to two thousand sentences that we read aloud and captured ourselves. We pretrained the model on that combined set and then fine tuned it toward a single target voice, hoping the final result would sound like a person rather than a machine reading off a script.

Did It Actually Work

We ran a blind listening test with nineteen volunteers who had no idea how the audio was generated. They rated its naturalness at 3.91 out of 5, which is a genuinely encouraging number for a language with this little existing speech data to build on. It is not going to fool anyone into thinking a human is speaking, but it gets close enough to be genuinely useful, and closing that gap was the whole point.