Speech Enlightenment

Inspiration

On different topics, many people have excellent knowledge,idea or experience. But due to poor speaking skills, they get failed to grab people attention and loosing their interest in topic. This gave us the motivation to make a product which can help people to deliver an excellent speech and sharing their knowledge more effectively.

What it does

Segments the audio files and provides insightful tags on various time stamps.

How I built it

Extracted 34 audio domain features using PyAudioAnalyzer.
Generated dynamic clusters of audio features using gaussian model mixture.
These 14 clusters are further converted as label for supervised learning task.
Training the model on these labels.
Tested on both the given dataset and random youtube video (TED Talk).

Disclaimer: Unsupervised task was not required as the data already contained information about the clusters and their respective values.

Challenges I ran into

Scaling the dataset as features need to be mapped to a specific range and labels need to be converted to discrete values.
Deciding the segment step size and window size and overlapping parameters.
Trade off between averaging after prediction using all data points or prediction after averaging short span sample.

Accomplishments that I'm proud of

The model is universal as it is based on phoneme extraction and analysis. Here we need to train only the upper layer for portability across languages.