DX-7 Auto Generating

Updated for final release

---

This is a project aiming on giving an approximate estimation of the synthesizer parameters of mdaDX10.vst to mimic a recorded sound timbre.

A Simple Guide

This project mainly consists of two parts

1. Dataset Generating in folder Dataset_generate

Adjust the configuration in Config.py

run the generate function in dataset_gen.py

2. Network training and testing in folder Regression

Adjust the configuration in sub-folder Configs/

Run the training code (e.g. DX10_Training_random_1k.py)

---

To Teammates

@jeffery @lucas please fork this repo and change the generator written in dataset_gen.py

def param_gen(length, param_num):

    # TODO: make a better version

    # A simple random version
    rand_array = np.random.rand(length, param_num)
    rand_array = rand_array * 100

    # make integer
    rand_array.astype(np.uint8)

    return rand_array

---

Environments

This project is based on Python 3.6

Please install the following packages under pip3

Package Needed	Usage
pickle	data encoding
scipy	scientific computing
numpy	scientific computing
multiprocessing	multi thread computing
tqdm	visualization
tensorflow	machine learning core
keras	machine learning embedding
IPython	debugging

---

Notes

The code for dataset generating part is completed, network is under construction.

The current version encode the dataset into pure binary code (e.g. Datasets/RandomTest/RandomTest-params.pkl)

Two files exists in the dataset folder:

NAME-params.pkl (TRAINING_SET_SIZE * PARAM_NUMBER)

NAME-wavs.pkl (TRAINING_SET_SIZE * FRAMES_PER_FILE)

These can be decoded by pickle

---

Training Data Generating

1. Change the function param_gen in dataset_gen.py

2. Modify Config.py to change the parameters including parameter number, thread number, and training set size

3. Import dataset_gen to any code or in IPython, and call dataset_gen.generate(NAME). A folder called NAME will appear in the Datasets Folder.