The Neural Network (NN) based Speech Synthesis System

This repository contains the Neural Network (NN) based Speech Synthesis System
developed at the Centre for Speech Technology Research (CSTR), University of Edinburgh.

Step 0: download data from http://homepages.inf.ed.ac.uk/zwu2/dnn_baseline/dnn_baseline_practice.tar.gz and put it in your working directory

Step 1: Install Theano version 0.6 and above (http://deeplearning.net/software/theano/). You may need other third-party python libraries such as matplotlib.

Step 2: Modify the configuration file 'myconfigfile_mvn_DNN_basic.conf' to fit your working environment. a) Change the following settings: [Paths] work: <YOUR_OWN_WORKING_DIRECTORY> data: <WHERE_IS_YOUR_DATA>

        log_config_file: <CODE_DIRECTORY>/configuration/myloggingconfigfile.conf

        sptk: <WHERE_ARE_SPTK_BINARY_FILES>
        straight: <WHERE_ARE_STRAIGHT_BINARIES>
        
b) By default, the program will run through the following processes if you keep them to be 'True'. You could run step by step by tuning on one process and tuning off all the other processes.
    NORMLAB  : True
    MAKECMP  : True
    NORMCMP  : True
    TRAINDNN : True
    DNNGEN   : True
    GENWAV   : False
    CALMCD   : True

Note: current recipe required C-version STRAIGHT, which cannot be distributed. Please replace the GENWAV step by using your own vocoder.

Step 3: Create myloggingconfigfile.conf based on exampleloggingconfigfile.conf

Step 4: Run './submit.sh ./run_lstm.py ' script. It will lock a GPU automatically. If you do not want to lock GPU, please modify the ./run_dnn.sh script.

If everything goes smoothly, after 3 to 5 hours, the program will give objective results, and synthesised voices if you have your own vocoder.

Feel free to contact us if you have any questions.