DeepFry: Identifying Vocal Fry Using Deep Neural Networks

This repository is for the paper DeepFry: Identifying Vocal Fry Using Deep Neural Networks by Bronya R. Chernyak, Talia Ben Simon, Yael Segal, Jeremy Steffman, Eleanor Chodroff, Jennifer S. Cole, Joseph Keshet.
It contains code for predicting creaky voice, as well as pre-trained models.

We provide two pre-trained models:

• DeepFry - from the paper.
• DeepFry - trained on both the Nuclear and Pre-Nuclear datasets, that were described in the paper.

This repository enables you to identify creaky frames in a given audio, see details below.

Requirements and Installation

Conda (Linux)

conda env create -f environment.yml

Custom Installation

• To run this repoitory, your environment should have Python 3.8.
• You will need Pytorch (1.12.0), but you don't have to use GPU. Please refer to: https://pytorch.org/get-started/locally/ and install the stable pytorch version most suitable for you environment specifications (OS, CUDA version, etc.).
• Finally, you will need the packages specified in the requirements file:

  pip install -r requirements.txt

Identifying Creak (Running the repo)

There are two options to run this repository:

1. Run on a directory with wav files without corresponding annotated textgrids.
2. Run on a directory with wav files and their corresponding textgrids.

• Remove the following argument from the commands below to run on CPU: --cuda
• If you get an error regarding the number of workers used in the test dataloader, due to you computers specs, you can change them by adding the following argument: --workers num_workers
• To run on a custom dataset, the .wav files (and optionally their corresponding TextGrid files) should be located under a folder named 'test' as follows(- See the 'allstar' folder in this repository for an example):

  |-- CustomDataDIR
  |   |-- test
  |   |   |-- file1.wav
  |   |   |-- file1.TextGrid
  |   |   |-- file2.wav
  |   |   |-- file2.TextGrid

  and then you can specify the argument --data_dir CustomDataDir

Identify creak - ALLSTAR dataset

This options allows you to test the repository. In the folder 'allstar' you will find wav files with their corresponding textgrids, which we used to test our model on, as specified in the paper.
Note that the results in the paper were reported for 20ms to have a proper comparison between methods, while our model was trained on 5ms, so the measures here might differ slighly.

Run option #1: Only output measures:

python run.py --data_dir allstar --model_name model_path --cuda

Run option #2: Output measures & Write predictions to a textgrid:

python run.py --data_dir allstar --model_name model_path  --out_dir out_path --cuda

where model_path is the absolute path to the pre-trained model, and out_dir is the path to the directory in which the textgrids will be saved to with the predictions of the model.

Identify Creak - custom dataset - no annotations

python run.py --data_dir data_path --model_name model_path --out_dir out_path --custom --cuda

Where model_path and out_dir is the same as above and data_path is the absolute path to a directory with wav files in which creak should be identified.

Identify Creak - custom dataset - with annotations

• Annotated silences - optional: should be under 'Speaker - word' tier, marked as 'sp' or under 'creak-gold' without a mark.
• Annotated creak - optional: should be under 'creak-gold' tier, marked as 'c'.
  You can refer to the TextGrid files in the "allstar" folder for an example

  python run.py --data_dir data_path --model_name model_path --out_dir out_path --cuda

Where model_path and out_dir is the same as above and data_path is the absolute path to a directory with wav files in which creak should be identified alongside their corresponding textgrids.