This is an experimental Bangla language learning app specially targeted for non-Bengali speakers. We mainly focused on implementing a practice session in this app to increase effectiveness of learning. 

Future Prospects

Currently, we live in the era of communication. Because of globalization, the entire planet has shrunk to the size of a single city. And there are so many people who speak Bengali that it ranks as the seventh most spoken language worldwide. The country of Bangladesh is a promising destination for vacationers. Our model can understand numerical, alphabetical, and linguistic input. You can use it as a tool to study Bangla. The absence of data is our project's biggest hurdle. We anticipate that the model's accuracy will increase if more varied data is introduced. A refined version of our prototype concept could serve as a platform for communication between Bengalis and those who are not from the region. We can create a version with the most used words and sentences, which could be useful as a means of communication for visitors. The potential of this methodology to teach Bangla to infants is enormous. Everything these days has a technological slant. They can be used as a tool for self-learning and to pique babies' interest with a sophisticated and fancy interface. This approach is widely embraced as a means of combining educational content with enjoyable activities. Illiterate people may be taught the basic essentials with this model. Teaching them at the elementary level will equip them to comprehend any writing.

Conclusion

We have successfully developed an app to recognize Bengali digits letters words and sentences. Our model's results are respectable. Google's search engine has consistently surpassed local engine in terms of recognition. The model could be trained to recognize objects using a KNN, SVM, decision tree, or ensembling technique. In every case, the neural network model has been found to be superior. When given fewer data points, neural networks struggle. Consequently, the limitation of data and the lack of diversity in data samples pose the largest difficulty to training local engines. We need a large and versatile dataset to fully exploit the potential of this software. Unfortunately, we have had some difficulty incorporating the Bangla typeface into MATLAB. The problem has been addressed via our Python interface. To be of practical use, this model can be enhanced. Our goal is to get local engine performance worthy of note by providing it with a suitable dataset and a few tweaks. 

Contributors: 

Supta Dev (User Interface)

Samiul Islam (Speech Recognition Algorithm)

Tonmoy Hossain (Tuning Prediction Model)

Sajib Biswas Shuvo (Team Lead)