Natural-Noise-Generator

Intuition

There are many applications that play natural noise like forest, stove, forest birds singing, sea wave, etc.
But actually they are playing a audio clip again and again, which means they repeat every several minutes.
Sometimes it may be a little bit annoying when noticing.

So we want to create a Natural Noise Generator, basically based on waveGAN, to generate natural noise continously without repeating.

Model

Based on WaveGAN, the Generator takes additional input from last part of previous audio sample as Y. See model.py for details.

Original Wavegan Network

Adaptation of Network

Running

Trainning

python main.py train --data_dir ./data/ --data_fast_wav --verbose

Generator Inferencing

python main.py generate --wav_out_time 150

BU SCC Environment Setup

Since there is no tensorflow=1.12 on SCC, we need to use anaconda to setup environment

1. module load anaconda3
2. conda create -n my_root --clone="/share/pkg/anaconda3/4.4.0/install"
3. source activate my_root
4. conda install -c anaconda tensorflow-gpu
5. pip install --user librosa==0.6.2
   Done!
   Then you can submit the job to SCC using qsub train.sh
   remember to modify project name

Remote Tensorboard

open a terminal
ssh -NfL localhost:16007:localhost:6007 jiaxin@scc1.bu.edu
open another terminal
ssh jiaxin@scc1.bu.edu
navigate to the working directory
module load python/3.6.2 tensorflow/r1.10
tensorboard --logdir=./train --port 6007
in the web browser localhost:16007

Result

Training on Ocean Wave Dataset

This is the spectogram of ground truth, audio generated by WaveGan, and audio generated by our adapted network. From the spectogram, we can see our network increase the continuity of the generated audio.