- install tensorflow (we used version 2.1.0)
pip install -r requirements.txt
download the TIMIT dataset from [this kaggle url] (kaggle.com/mfekadu/darpa-timit-acousticphonetic-continuous-speech) and unzip it into a directory named "data/timit" inside the root of this repository.
python noisy2txt.py
You can adjust options for the run by editing (or adding entries to)
the custom_static_params
dictionary at the bottom of "noisy2txt.py".
You can also run stft_denoise.py
to train a larger denoisng model. A larger dataset need to be downloaded.
The model will be saved under the model
directory and you can evaluate it with sftf_eval.py
.
python timit_loader.py
Note that this allows you to check that the transcripts are cropped
properly and that the noise level is as desired.