Bahubhashi - The Speech Translator

GitHub Link

  • Recently many MNC’s are facing an issue where a customer of a particular region calls in and the customer experiences long wait hours as all the representatives are busy
  • While their call centers in other regions are not experiencing busy wait hours
  • To mitigate this issue there could be a system designed which takes an audio input in one language and translates it into another language
  • Designed a complex network of 22 layers in Keras to convert speech commands in English to text with 94.5% accuracy
  • Built an auto-encoder/decoder model involving four bi-directional RNNs and LSTM layers to convert text from English to German with a BLEU score of 0.6
  • This solution could be used where if the call centers in one region are facing high call volumes the calls could be directed to the regions facing lower call volumes and BahuBhashi could be used so that person on both ends can converse in their own language and find a solution to their issue.

Image Style Transfer

GitHub link

  • We give a photo from our camera roll, then we select a design to combine both the images and we get a resultant new image which has the content of our input image and style of the design image. In the world of deep learning this is called Style transfer
  • Three steps to get the target image:
      • Extract feature map from content image and a random generated image through VGG-19 and try to minimize the difference.
      • Evaluate the correlation between two patterns and to transform it. Here we use dot product of to represent the correlation.
      • Define the total loss and use Adam algorithm to optimize the result.
  • Generated an amalgamated image with one styling template and content image utilizing pre-trained VGG-19 image dataset and dependencies such as TensorFlow, NumPy, SciPy, Pillow in Python
  • Tuned numerous hyper parameters and model architecture and calculated styling and content losses to scrutinized model performance

Diabetes Prediction

  • Achieved a maximum accuracy of 92.5% by cross-validating and running predictions for diverse supervised machine learning models such as GLMs, GBMs, deep learning
  • Cleaned and provided modeling and analyses of structured and un-structured data with the help of pandas, seaborn and matplotlib for Kaggle dataset
  • Found the best result with Random Forest and SVM.

Hotel Management Database

  • Examined the requirements and Implemented an efficient and systematic system to manage all the Hotel Data reducing redundancy in data.
  • Developed a database for a Hotel Management system, designed the EER and conceptual schema from the given requirements.