Voice Conversion with Non-Parallel Data

Using your own dataset

How to make it work:

First, I set up a docker container with this docker container run -dit -v /mas:/mas -v /u:/u -v /dtmp:/dtmp -v /tmp:/tmp --name $(whoami)-deep-voice-changer tensorflow/tensorflow:latest-gpu. I did everything specified by Necsys plus these aditional steps:

1. apt-get install libsndfile1
2. apt-get install ffmpeg libavcodec-extra
3. apt-get install build-essential libcap-dev
4. apt-get install nano
5. apt-get install screen
6. apt-get install virtualenv (or python-virtualenv)
7. apt-get install git
8. apt install libcudnn7 libcudnn7-dev

After that, I sudo'd and cloned the repository by andabi in /u/$(whoami). Then, make an environment and pip install following packages:

• librosa == 0.6.2 (make sure that numba == 0.48 and llvmlite==0.32.1)
• tensorflow-gpu==1.15 (makse sure you pip uninstall tensorflow before, seems to help)
• tensorpack==0.9.0.1
• pydub
• soundfile
• git+https://github.com/wookayin/tensorflow-plot.git@master
• joblib
• pyyaml
• gdown

Once you have these packages, you have to modify two lines in the tensorpack file. Do nano $(nameOfEnv)/lib/python3.6/site-packages/tensorpack/graph_builder/utils.py and change "from tensorflow.contrib import nccl" to "from tensorflow.python.ops.nccl_ops import all_sum" and change "summed = nccl.all_sum(grads)" to "summed = all_sum(grads)". These lines are in the function allreduce_grads.

Finished that? Good, almost there! Now do nano hparams/default.yaml and change "logdir_path: '/u/$(WHOAMI)/deep-voice-conversion/logdir'" (remember to make the folder 'logdir' and also make a dir inside it called "1"). Then, in the same thing change data_path: '/u/$(WHOAMI)/deep-voice-conversion/datasets/arctic/slt/*.wav', which should be under train2.

One last thing is missing, you have to install the pretrained model for train1. Run the following commands:

• gdown https://drive.google.com/uc?id=1yC3G3V03X3s8mKJ1J6bMkOqDT8r-TBb8
• python
• In the python terminal thing, write:

import zipfile
with zipfile.ZipFile("train1.zip", 'r') as zip_ref:
    zip_ref.extractall("logdir/1")

Now you should be good to run train2 by doing python train2.py "1" "2"!

The last bug I ran into was to modify the trailing silences in the audio (because they are already removed in my dataset, not necessary if using the arctic dataset) and also a cuda version incompatibility (Something along these lines: Loaded runtime CuDNN library: 7.1.2 but source was compiled with: 7.3.1. CuDNN library major and minor version needs to match or have higher minor version in case of CuDNN 7.0 or later version. tensorflow/tensorflow#23715)

To use convert.py, you have to add the list thingy to the audio and the other thing. Remember to fix this.