“My school had taken me for a field trip to a paralysis care centre, and I saw the difficulties the patients had in communicating with their caretakers and their surroundings. That’s where the idea came to me. What if there was a machine learning system which could understand what they’re saying?” said 11th-grade student Pranet Khetan. The young innovator has developed a small device that can translate slurred speech.
The device, according to Khetan, is India’s first open-source Automatic Speech Recognition (ASR) framework specially designed for Hindi dysarthric speech. Talking about the device, Khetan said that what struck him most profoundly was dysarthria – a motor speech disorder affecting patients with paralysis, Parkinson’s disease, COPD, and other neurological conditions.
“The advantage with dysarthria is that the speech processing in their brain is completely fine. They know what they want to say, and they understand everything everyone else is saying. It’s just that their muscles aren’t able to properly produce speech. But the meaning and that intent are still there,” Khetan told indianexpress.com.
Hindi speech recognition
Khetan has named the device Paraspeak, and it aims to address a critical gap in the field. In India, while there is no specific data on dysarthria, it is a fairly common speech disorder. The teen said that while there is some research for English dysarthric speech recognition, there are no solutions for Hindi-speaking patients. “I chose Hindi because it’s spoken by over 40 per cent of India. But surprisingly, there is no dataset for Hindi dysarthric speech. No one has ever gone out and collected Hindi dysarthric speech for the purposes of training a model,” he explained.
This led to one of the challenging aspects of his project – data collection. As part of his research, the teen visited NGOs and care centres and created the first-ever dataset of Hindi dysarthric speech comprising 42 minutes of recordings from 28 patients. Later, with the help of sophisticated data augmentation techniques, he expanded it to 20 hours of synthetic training data. Even a simple sentence with three or four words would take time and effort for those affected by dysarthria.
“It was really emotional, because what happens with dysarthria is that people aren’t able to understand what you say. So eventually they withdraw from their surroundings and stop trying to speak in the first place,” he said.
“Many of these patients were putting in a lot of effort, and there were many cases where I was telling them to stop, the doctor was telling them to stop, but they still kept going. They were really happy to contribute.” Following real-time testing, Khetan filed a patent for the device.
Story continues below this ad
The tech behind Paraspeak
Khetan shared that the small device is built on the transformer architecture, the technology that is powering large language models like ChatGPT and Claude. However, for the device, he adapted the architecture for speech recognition. When it comes to function, the system takes slow, unclear speech from dysarthric patients and converts it into clear, understandable speech that can be played aloud.
The device comes in a compact form factor, as small as a webcam, that can be worn around the neck. “All this device needs to operate is an internet connection. It has a long battery life. It can last more than 10 hours at least,” Khetan explained. The patient needs to simply press a button to activate recording, speak into the built-in microphone, and the audio is processed by AI models running on cloud servers before returning clear speech output.
What makes Paraspeak unique is its scalability. Unlike existing research solutions that work only with individual patients, Paraspeak is designed to work across multiple speakers with the same underlying system. “Current research is very limited in this area. Such models only work with one patient. So let’s say I take one patient’s data, and I train my model on them. It will only work with them. It’s not scalable across multiple speakers,” he noted.
Designing on a budget
The speech recognition system stands out for its remarkably low cost of construction. Khetan said that the device costs Rs 2,000 to manufacture, with a monthly subscription fee of around Rs 200. The subscription here is internet data packs from telecom providers. According to the teen, the low cost of the device makes it accessible.
Story continues below this ad
Talking about the cost efficiency, Khetan admitted that it was achieved through numerous iterations over a year. The first prototype was bulky with many wires attached; however, he progressively miniaturised it using custom printed circuit board designs. “I have designed my own custom printed circuit boards. It’s almost like how a smartphone would work on the inside. Very compact, custom electronics, which I’ve tried to design to create this as small as possible,” said Khetan, who is a student of Shiv Nadar School, Gurgaon.
As of now Paraspeak isn’t just a proof of concept; according to Khetan, it is a working solution that has been tested on real patients. “I tested it on seven patients. I also recorded videos of three of them with a variety of diseases like congenital disorders, paralysis, Parkinson’s, and COPD, and they were content with the fact that the device worked for them,” said Khetan, whose innovation also got recognised at the Regeneron International Science and Engineering Fair (ISEF) 2025, held in Ohio, USA. His work was also recognised at India’s IRIS National Fair.
As of now dysarthria is present in over 75 per cent of Parkinson’s patients and mostly all late-stage Amyotrophic Lateral Sclerosis (ALS) patients. Technologies like Paraspeak can bring hope to many. Khetan shared that as of today there is no market-ready solution for this disorder, making it a unique piece of tech.
When asked what excites him most about technology, Khetan said, “I’m really passionate about assistive technology. There’s so much in the technology spectrum which can be applied to assistive technology to tangibly improve the lives of so many people across the world. The best solutions are those with a tangible human impact, where you can instantly see a change in someone’s life.”