KAISD

This repository contains the dataset, code, and pretrained models for KAISD (Korean language Alcohol Intoxicated Speech Detector).

This project was done for 2020 Fall KAIST CS470 course and was done by Team 9 (SeokJun Kim, Jeongeon Park, and Sujin Han).

Dataset and Preprocessing

All the audio files for our collected data is in "CS group project wav recordings" folder.

The code for preprocessing (splitting audio files into 8 second segments and converting audio files to mel spectrograms) is in data_preprocessing.ipynb.

Code for Training and Evaluation

The code for training and evaluation is split into two files KAISD_final.ipynb and KAISD_final_with_modifications.ipynb.

KAISD_final contains code for conv_original, pretrained_original, and inception_5. KAISD_final_with_modifications contains code for all other models. KAISD_final_with_modifications also contains code for visaulizing CNN filters and intermediate layers.

Due to the shuffle option in DataLoader (we saved the DataLoader with shuffle=True), the evaluation result using validation set may vary from run to run. Also, since we use two randomly initialized FC layers for Baseline model, the evaluation result for baseline model using held out data also may vary form run to run. However, the changes should be within a small margin.

Pretrained Models

The pretrained models can be found in the "models" folder. The pretrained models for conv_original, pretrained_original, and inception_5 can be found in "models/without_modification" while pretrained models for all other models can be found in models/with_modification.

Link to Google Drive

Only codes (.ipynb files) and audio file labels (.csv files) are uploaded in this Github repository. To access pretrained models and other files necessary to run this code, please go to our google drive provided below.

https://drive.google.com/drive/folders/1e2BpwiHKOx6Zr9nq5CSd0H7dWHD9kX6J?usp=sharing

Instructions for running our code