Support my efforts: https://www.buymeacoffee.com/tinylucas
This repo is the implementation of Tacotron-2 model for Korean voice.
For someone (like me) wanna check demo before go futher, I posted synthesized files in samples folder.
There are some requirement must be met to install requirements.txt some of them cannot be install via pip command:
such as pyaudio: download here https://www.lfd.uci.edu/~gohlke/pythonlibs/#pyaudio and use
pip install *.whlto install for windows, for linux:sudo apt-get install libasound-dev portaudio19-dev libportaudio2 libportaudiocpp0 sudo apt-get install ffmpeg libav-tools sudo pip install pyaudioOr such as pytorch: download here https://pytorch.org/ and install as home page instruction...
1 - Download traindata from: https://www.kaggle.com/bryanpark/korean-single-speaker-speech-dataset
The downloaded folder has 4 subfolder named from 1~4. For simplicity, I remove subfolder and put all file in a folder called "wavs" - take a look at Train_data_folder_structure.png
And in case of the dataset author changes the file format, I have captured the screenshoot of transcript.txt file format as transcript_file_format.png
put downloaded pretrained model as structure as in pretrained_folder_structure.PNG
open terminal (or cmd interface in Windows), activate virtual environment you wanna use, or use your system environment. If you do not know how to do, take a cup of coffee and visit https://packaging.python.org/guides/installing-using-pip-and-virtual-environments/ there are some packages will require some other apps or libraries (such as pyaudio), or could not installed using pip (such as pytorch), you need using Google to find out how to install them.
In case you are using different OS from me, go to Utils\AudioPreprocess\AudioPreprocess.py file, edit the line 349 like this:
# folder_id = input_dir.split('/')[-1] ### get speaker folder name (in linux)
folder_id = input_dir.split('\\')[-1] ### get speaker folder name (in windows)
##Start Preprocessing process:
open terminal (cmd) run: python Utils\AudioProcessing\AudioPreprocess.py --> this step will take input audio, and generate mels spectrogram, wav file, as well as linear spectrogram and put them in Tacotron_input folder.
This should take less than 2 minutes.
Preprocessing and pretrained model need sharing same hyper parameters, so, if you want to use pretrained model, do not change hparams.py (of course you can change some, but this will need deep understanding, not recommend) file in Utils folder.
after preprocessing stage, run python TacotronModel\train.py to start training with default parameters, or add -h param to see parameters info
This will continuosly train model, using pretrained model, and after checkpoint_interval steps, it will automatically save in pretrained folder.
If you want train model from scratch, use restore parameter, or just delete pretrained model folder.
with a well trained model, using python TacotronModel\synthesize.py command to synthesize. This step will take data from Tacotron_input folder, predict and save result in tacotron_output folder.
The output of synthesizing process, will be input of Wavenet vocoder model training stage.
You can totally synthesize audio without Wavenet vocoder (may be the difference is about audio quality). using: python TacotronModel\synthesize.py --mode=inference
This will generate output using texts (from sentences.txt) and save in tacotron_output/inference, the output audio, plot are in tacotron_output/log-inference folder. (as synthesizing process, the output in tacotron_output/inference folder will be input of Wavenet vocoder inference process.
Because we are working on Flowavenet model for realtime synthesis, I will not post how to train and use Wavenet model here
Waiting for Flowavenet.