Investigating Effects of Data Quantityand Augmentation Methods for Speech-To-Text Systems

This repository contains the code to run experiments and evaluations described in the thesis Investigating Effects of Data Quantityand Augmentation Methods for Speech-To-Text.

Prerequisites

Experiments depend on ESPnet. To run the experiments, follow the installation instructions. I recommend using the docker container provided by ESPnet.

After installing ESPnet, copy files from the asr1 folder to the commonvoice asr recipe folder.

cp -r asr1/. <espnet-dir>/egs/commonvoice/asr1/

Running the Experiments

To run experiments, execute the run.sh.

cd <espnet-dir>/egs/commonvoice/asr1/
./run.sh > log 2> err

Note: You may need to activate the conda environment provided with the docker container, before starting the experiments.

conda ./<espnet-dir>/tools/venv/bin/activate

Results

Results will be collected in <asr1>/tensorboard and <asr1>/results.

Results of the experiments are also included in this repository. In case you modify experiments and generate the same outputs, you need to place results into the repository directory under results/raw/ and run run.py.

To generate the outputs the Seaborn and Librosa packages are required.