stoplime / openaudioai

Open Audiobooks

OpenAudioAI

The goal of this project is to eventually create an audiobook system that can autodetect who the speakers is and change its voice settings accordingly. To do that, our first step is to create a speaker segmentation system which can detect who the speaker is purely text-based. There is not very many papers about speaker detection that is purely text-based, so we created one here.

Author 1: Steffen Lim Email: [email protected] Author 2: Sams Khan Email: [email protected]

https://docs.google.com/document/d/1fe0iCKSy-xPkjXN8e_PjmjxUOdGY7QqCyvKVCkFh2d8/edit?usp=sharing

To do: Dataset Preparation:

• [] 1) Split LSTM networks to context and main networks

• [] 2) Implement concept from the main paper dynamic attention from sentence embedding to speaker embedding

  Current model Idea:

openaudioai's Issues

Train the model for longer

Try training the model for more than 120 epochs. This will probably need some sort of tracking, either through tensorboard or through manual since it saves the output of every epoch in a separate document

Find NLP big researchers

Find and get an understanding of who and what big research topics are on NLP projects. Try to find a list of research topics where the big researchers research about.

Run a forward pass through the given model

The existing code did not have any way to generate an output from the network.

• Make a save of the model after training
• Figure out how the input of the model is setup
• Run a forward pass

Summarize and organize sources and info about Hierarchical and attention based models

what is so special about attention based and hierarchical models

Find specific papers that was succesful with attention based models and herarchical models.
Simplify understanding of these ideas and what makes them special.

Download Code and run.

https://arxiv.org/pdf/1703.07713.pdf

Find Feeds on NLP info

Find newsletters, podcasts, twitter feeds and any feeds that provide new information on NLP.

Find NLP research hubs

List all the good research hubs found in the comments. Some research hubs may include Subreddits, discord/slack channels, or forums.

Running forward pass and configuring data for sample code

Literature Review Paper Trail

Leave a list of all the useful papers that pertain to the project here in the comments. These would include projects that try to do this exact thing, projects that do similar things, and projects that may help this project in any way.

Disecting the original code to understand how to implement our model

Run the whole data through the model

Use the extra data folder and compile it to be able to train on it. then run an evaluation on it.

Find the popular NLP Technique

List out all the popular Techniques being used in NLP. Such examples would be like attention CNN replacing recurrent networks.

Brainstorm for new model architecture for speaker detection in Text.