WaveRNN Dataset format about multilingual_text_to_speech

[Tomiinek]

Hello 🙂

No problem!

First, set hyper-parameters in hparams.py ... you might be interested in data_path, voc_mode, bits and mu_law (but please go through issues in the original repository, there are some good hints and conversations about convergence etc.)

Then use the preprocess.py file. To run it, you need a directory with directories for each language, i.e. de, fr, ...

Each of these language-specific directories should contain two directories named wavs and gtas. The wavs directory should contain all the .wav files of the particular language in dataset. The gtas directory should contain ground-truth aligned spectrograms of the corresponding .wav files with the same filename, but .npy extension. These GTA spectrograms can be generated using the gta.py script in this repository.

Then run the preprocess.py where --data_root is the base directory containing language-specific directories, --inputs is a list of names of the language-specific directories, and --output is an output directory. The script will generate mel (GTA spectrograms in a way which is supported by the model) and quant (quantized audio files which are used as targets during training, so it is needed to re-generate files if you change the parameters mentioned above) directories and a metafile dataset.pkl in the output directory.

That is hopefully all 🥳

(I do not remember details precisely, so I am sorry if something I said is not true 😄 )

[sooftware]

Thank you for always giving me a good answer, Tomiinek!!