Additions to webapp, video streaming Reevaluated problem statement and goal through ethics assignment More tests with example voice tone analyzer using CNNs To Do: Journal recording storage on cloud Fix display bug for videos More Voice Tone Training
This week I finished the Google speech to text pipeline and a test script to demo my best network. I’m working on increasing my accuracy past 65%. I’m expecting to hit 70% accuracy next week. Additionally, I will work on creating a better dataset to train my final network on. I’ll also think of a way to visualize the classification of a journal entry (pie chart, bar chart, etc.).
Changed the program I’m using to get pitch and other features; no longer using Praat because it takes too long and is inefficient; using the python Librosa library instead (documentation here https://librosa.github.io/librosa/) Testing out example voice tone analyzer using CNNs; one project claims to have reached 70% accuracy Next week: Prepare demo Polish up web app More progress on voice tone emotion analysis
This week I trained finished training a few networks. The highest accuracy I achieved was 53% which is less than ideal. I trained on 40,000 tweets scraped from twitter and did some minimal preprocessing. Going forward, I want to implement the Google speech to text pipeline and train the network on more data with increased preprocessing. I’m expecting to get at least 65% accuracy. I plan on demo-ing the Google speech to text and text Read more…
This week we just continued working on our individual components of our project. Patrick continued training with images from AffectNet. Vinay continued training and working on the LSTM for sentiment analysis. Yoojin found a paper that detailed a CNN for tone recognition and collected the libraries and documentation to start building the network. We are also preparing for our demo in lab the following week. Vinay and Patrick are going to show how their algorithms Read more…
Completed: Trained with 10,000 images from AffectNet Accuracy still around 70% Lowest around 30% Highest at 77% Confusion matrix shows highest accuracy for happy, then anger, then sadness, surprise, fear, then contempt and disgust are not detected Downloaded 7 out of 11 parts of AffectNet so far To-do: Start batch training so that my computer memory isn’t taken over Take equal amounts of each emotion so that gradient descent won’t be skewed towards happiness Make Read more…
With the transition to online classes, I will no longer be building and designing the physical product for our project. Additionally, I will no longer need to manage porting code to the Raspberry PI or worry about its limited capabilities. The core of my workload going forward will consist of creating, training and tweaking a neural network that can recognize textual emotion. This week I aim to train the network and get a working product Read more…
This week: Talked about our progress so far, cleaned up end goals and worked on SOW Revised our team Gantt chart together Next week: All of us should have made significant process on our individual portions Start discussing bringing the algorithms together
This week was focused on gathering data, picking algorithms, and preprocessing/getting features of data Next week will be finishing gathering important features from data and starting to train. SOW and Gantt Chart: The original web application was going to communicate with the RPi and receive textual journal entries of the user. Since we will no longer be making the physical product, the tasks related to setting up communication between the Raspberry Pi and web Read more…
Completed: Personal SOW Downloaded AffectNet manually-annotated images Save trained SVM locally Testing different test/train splits Just testing and training further To-do: Train with more images Cross-validation on AffectNet Confusion matrix