Voice assistants can understand English almost instantly. Many can also recognize widely spoken languages such as Spanish, Hindi and Mandarin. But try speaking to one in Nepal Bhasa, and you probably won’t get very far.
A group of researchers from Kathmandu University is working to change that.
They have created Nwāchā Munā, a collection of recorded and transcribed Nepal Bhasa speech that can be used to train computers to recognize the language.
What Is Nepal Bhasa?
Nepal Bhasa, also commonly called Newari, is the native language of the Newar community and has been spoken in the Kathmandu Valley for centuries. Its name literally means “language of Nepal.”
Despite the similar name, Nepal Bhasa is not a dialect of Nepali. It is a separate language belonging to the Sino-Tibetan language family. According to the researchers, it is spoken by more than 860,000 people in Nepal. However, UNESCO classifies it as a “definitely endangered” language.
Like many Indigenous languages, Nepal Bhasa has far fewer digital resources than major languages. This makes it difficult to develop tools such as speech recognition, voice typing and automatic transcription.
Building a Nepal Bhasa Speech Dataset
Voice-recognition systems need recordings of people speaking, along with accurate written versions of what they said. These recordings help the system learn how spoken sounds relate to written words.
For the Nwāchā Munā project, researchers created a dataset containing 5.39 hours of Nepal Bhasa speech and 5,727 recorded sentences.
The recordings included 18 native speakers from Banepa, Dhulikhel, Panauti and Patan. The participants came from different age groups and included 10 men and eight women. Most of the original recordings were made using smartphone microphones.
The researchers also collected written material from Nepal Bhasa Wikipedia, school textbooks, newspapers, literary works and everyday conversations.
What Did the Researchers Find?
The researchers tested whether a speech-recognition system already trained in Nepali could be adapted to understand Nepal Bhasa.
Before it was trained with the new dataset, the system had a character error rate of 52.54 percent. After it was trained using the Nepal Bhasa recordings and additional data-processing techniques, the error rate dropped to 17.59 percent.
In simple terms, the computer became much better at recognizing individual characters in spoken Nepal Bhasa. However, it still had a word error rate of 52.61 percent, meaning it frequently made mistakes when forming complete words. The results are promising, but the technology is still at an early stage.
The researchers also found that the Nepali-trained system performed about as well as Whisper-Small, a much larger multilingual AI model. This suggests that knowledge from a nearby language such as Nepali can help build technology for Nepal Bhasa without requiring an extremely large and expensive system.
Why Does This Matter?
This does not mean that Siri, Alexa or other voice assistants can suddenly understand Nepal Bhasa. The project is still in its early stages, and the dataset is based on a relatively small number of speakers.
However, it creates a useful starting point.
In the future, this type of research could support Nepal Bhasa voice typing, automatic subtitles, audio transcription, educational applications and digital archives. It could also make it easier to preserve recorded stories, songs and conversations for future generations.
Technology alone cannot save a language. Nepal Bhasa will continue to live only if people speak it, teach it to their children and use it in everyday life.
But in a world where people increasingly communicate through phones, computers and AI tools, making a language available digitally is an important step.
The researchers have publicly released the dataset and their findings so that others can continue improving the technology. Their complete 10-page research paper, “Nwāchā Munā: A Devanagari Speech Corpus and Proximal Transfer Benchmark for Nepal Bhasha ASR,” is available to read online or download as a PDF.
Editorial note: This article was prepared with AI assistance and reviewed and fact-checked by NepaliSite using the original research sources.
0 Comments