Overview / Usage

In this project we have created a machine learning model which classifies the audio files of TATA Motors call centre conversations into different categories by their intent ie. if the calls were for Breakdown , Complaint , Feedback etc .

Methodology / Approach

●Our dataset contains 84 audio files

●Model trained on 75 and validated on 9 samples

●We trained our model using Keras library’s Sequential Model

●Model consisted of combination of Bi-directional LSTM and dense layers

●The final activation function was Softmax

Technologies Used

Approach

●Source and nature of data

●We have created our own data . We made the text dataset based on the conversation between the TATA motors call centre agent and Customer ●** Pre-processing approach** ● We first converted our audio file into text using speech_recognition library. Then we translated our text to english as a whole by using googletrans Translator. Then we have cleaned out text data using nltk and re library. ●** Technology used** ●Neural Networks ●LSTM ● Speech Recognition ● Language Translation

​

​

Steps to run the code

●Loading and shuffling dataset

●Cleaning and tokenizing text data

●Padding in input sentences

●One hot encoding and train_test split

●Model creation

●Model training

●Loading test audio files and converting into text

●Translating Hindi Or Other Languages to english text

●Prediction of the class

​

Solution Architecture

​

●To solve the given problem we took the approach that an audio file must be converted to a text document which can be further used for training using Natural language Processing methods

●While predicting , as the problem statement stated that used language might be Hindi/English / Other . so we decided to translate the whole text data to English ( Which was the language we used for training )

●Model with best validation accuracy and with lowest validation loss was chosen.

​

​

Repository

https://github.com/bajpaiaviral/audio-intent-clasification