DRONA

Student to Teacher ratio in India is one of the lowest in the world. COVID has accentuated this problem. The huge deficit in the number of teachers in India can only be solved with the help of artificial intelligence and technology. ( Tier-2 and 3 cities of India and rural areas have 1:50 teacher to student ratio and it is impossible to give personal attention to children ). Since the students are not personally attended in rural areas, the feedback loop is very long. AI can give immediate feedback without teacher intervention.

We have created a prototype of the solution.

Demo Link

Here the app listens to the student reading a book, and assess the speed of reading, pronunciation etc. It gives syllable level training - the sound of the syllable and the length of the syllable so that the user can improve his pronunciation. This app is a stand alone one which can be used by anyone who wants to learn English. 

Student to Teacher ratio in India is one of the lowest in the world. COVID has accentuated this problem. The huge deficit in the number of teachers in India can only be solved with the help of artificial intelligence and technology. ( Tier-2 and 3 cities of India and rural areas have 1:50 teacher to student ratio and it is impossible to give personal attention to children ). Since the students are not personally attended in rural areas, the feedback loop is very long. AI can give immediate feedback without teacher intervention.

We strongly believe that A.I. in education can make a dent. It can save millions of man-hours. It can train teachers highly effectively. The A.I can be a constant companion that works on a budget phone.

Our solution can convert any text based content ( stories, reading materials etc ) to an interactive book that can listen without any manual intervention. In the example of seKret stories book app - as soon as a book is opened we start a connection with the server and send audio inputs continuously. And parallely we are receiving processed outputs ( the recognized words ). 
In comparison, most of the other speech recognition apps send the entire audio file of the voice input to the server, get it processed and receive the output. The challenge for a real-time AI is the fact that it has to keep on adding new audio inputs ( new words )to the context and refine the output continuously. 
 
We are able to support a use case like book reading only because we have the real time processing advantage. ( A use case like book reading needs real time feedback. Recording the audio and sending it to the backend will not quite work there. In the example of quiz, the AI provides streamed text results - immediately returning text as the user is speaking and corrects any wrong outputs automatically. 

We have 200 million potential students in India. By providing an alternate learning opportunity, we are taking a huge load off the current educaitonal infrastructure.

Another potential use case is up-skilling. By providing English pronunciation training to the young population, they will be much prepared for the job market.

With over 200 millions school going children, the only way to make an impact is to embrace technology. Current educational infrastructure is failing miserably and need to take advantage of digital technology. 

By building an A.I. Based mentor / assessment solution that anybody can access, we are able to provide an alternate learning platform ( which is free ) that works on a mobile phone. 

We are currently partnering with multiple non profit organisations and help them customize and  deploy the solutions to their beneficiaries.

. 1Our system has the ability to understand when the user is talking, when the sentence is over and when the utterance is over. 

The AI is able to understand when the use starts talking and when a sentence or a paragraph is over by being context aware. It also understands single utterances. 

Identifying a single utterance is useful when the AI is expected to perform an action once the user has finished talking. ( For eg: in a question answer scenario, the AI can quickly assess if the answer is correct or not, or even direct to a second part of the question by recognizing when the user has stopped talking / finished answering ) 

2. The AI can isolate human voices from a variety of background noises in real time. Real-time is important since the AI does not have a lot of time to compare patterns of background noise to isolate. 

3. Correctness score algorithm. Correctness score is calculated based on syllable length and the pronunciation and matching is done by measuring the levenshtein distance ( distance between characters ). Levenshtein distance is the most effective way of processing similar words. We have written our own algorithm for this. 

4. It is configurable 

The correctness score can be personalized for the user by defining the threshold. The personalization can be done at user level or at content level. 5. Works with numbers 

Users can say "one hundred" or ”one zero zero” , and both will give the same results real time. ( Imagine a voice enabled lesson for multiplication table for instance ) 

6. It can tell whether the user has completed reading a text. 

The AI can track ( at a word level ) whether the user has read a particular text. This is useful in creating a learning path for the user and suggesting new materials to read. 

4000 + schools

 ~ 500,000 students

Number of words pronounced is a metric that we use.

~ 15 contributors and founders

The team consists mostly of software engineers and machine learning experts. Most of the works is currently being done to perfect the system, train the models etc

Too early for that.