Week 2: Single line UI, Prompt files and Deep Learning

This week certainly started with a lot of clarity and direction. James provided several utilities to evaluate manipulate the prompt sentences and use them for pronunciation evaluation using logistic regression and DNNs. He also provided minor suggestions regarding UI which can be seen here -> link.

This week shall be dedicated to make these buttons functional and realize the 5th milestone of the proposal.

Past week we achieved the following:

1. Minor UI change mentioned above.

2. Code to generate prompt files in the following format:

line 1: Full text of the word or phrase for display, i.e., in upper
and lower case with full punctuation
all subsequent lines: <word in lower case> <single-character
part-of-speech-code or _ if missing> <phoneme(s) in lower case>...

3. Understood James' code for prompt file manipulation and evaluation with his DNN code achieving 87% accuracy over test set as compared to 75% using logistic regression. Yay :)

dims: 16 layers: 2 epochs: 500 accuracies: 73.41% 89.60% 76.88% 79.77% 83.24% 93.64% 90.17% 87.86% 98.27% mean: 85.87%

dims: 32 layers: 2 epochs: 500 accuracies: 77.46% 90.17% 77.46% 81.50% 83.82% 93.64% 90.17% 87.28% 98.27% mean: 86.64%

dims: 16 layers: 3 epochs: 500 accuracies: 77.46% 90.17% 78.61% 80.35% 82.08% 93.64% 91.33% 87.86% 98.27% mean: 86.64%

dims: 32 layers: 3 epochs: 500 accuracies: 75.14% 90.17% 72.25% 80.92% 83.24% 93.64% 90.17% 86.71% 98.27% mean: 85.61%

dims: 16 layers: 2 epochs: 1000 accuracies: 78.03% 89.60% 76.30% 84.97% 83.24% 93.64% 91.91% 87.86% 98.27% mean: 87.09%

dims: 32 layers: 2 epochs: 1000 accuracies: 78.61% 90.17% 75.72% 83.82% 84.39% 93.64% 91.91% 87.86% 98.27% mean: 87.15%

dims: 16 layers: 3 epochs: 1000 accuracies: 80.35% 90.17% 75.72% 82.66% 84.97% 93.64% 91.33% 87.86% 98.27% mean: 87.22%

dims: 32 layers: 3 epochs: 1000 accuracies: 79.77% 90.17% 79.19% 84.39% 82.08% 93.64% 91.91% 87.86% 98.27% mean: 87.48%

best: 87.48% dims: 32 layers: 3 epochs: 1000