speaker_popularity

Automatically assessing public speakers popularity: A use-case for ted-talks

Structure

• audio_features: folder containing code for extracting high-level aggregations of the behavioral posteriors
• metadata: folder containing all generated datasets
  • ted_main.csv: The original TED talks dataset
  • merged_metadata.csv: The above dataset merged with the transcripts
  • merged_metadata_popularity_std.csv: The dataset enriched with the target metrics (popularity + ratings)
  • merged_metadata_popularity_features_std.csv: The dataset enriched with target metrics and features used in classifiers
  • embeddings_transcript_clean.csv: The embeddings of the transcripts
• mfccs: The mfccs per file, calculated with librosa
• modeling_api_results_embeddings: The responses of our API for each file, containing the behavioral embeddings and posteriors
• results: The directory where experiments results are saved
• scripts: Various scripts that are used by the process (explained below)

Install dependencies and environment

Make sure you have git lfs installed. Clone the repo and create a virtual environment (python 3.8) and install the required dependencies

git clone <URL_OF_REPO>
cd speaker_popularity
virtualenv -p python3.8 venv
source venv/bin/activate
pip3 install -r requirements.txt

Unzip modeling_api_results_embeddings/emb_part_1.zip and modeling_api_results_embeddings/emb_part_2.zip and place all files under the modeling_api_results_embeddings directory

unzip modeling_api_results_embeddings/emb_part_1.zip -d modeling_api_results_embeddings/
unzip modeling_api_results_embeddings/emb_part_2.zip -d modeling_api_results_embeddings/

How to run:

1. Generate the dataset including the target labels (popularity + ratings). The generated dataset's name will be metadata/merged_metadata_popularity_std.csv

cd scripts
python3 generate_popularity.py

2. Enhance the dataset with the extracted features (behavioral embeddings + posteriors). The generated dataset's name will be metadata/merged_metadata_popularity_features_std.csv

python3 generate_dataset_with_aggregations.py

3. (Optionally) Generate the text embeddings of the transcripts by running the clean_transcripts_embeddings notebook. An OpenAI API key is required for this step. The generated embeddings dataset will be metadata/embeddings_transcript_clean.csv. You can find the already extracted embeddings in the repo, so this step is optional.

4. Run the experiment.

python3 run_experiment.py

The results wull be saved under the results folder. You can then run the cells in features_analysis notebook, to visualize and explore the results (results/scores.csv)

Additional scripts

• visualize_target_set: Contains visualization regarding the distributions of the popularity and ratings metrics
• eda: Various graphs and aggregations used as part of data exploration
• get_correlations: Prints the correlation matrix between target labels and features