We use 2 external tools for waveform inversion.
- https://github.com/nii-yamagishilab/project-CURRENNT-public (extended CURRENT toolkit)
- https://github.com/nii-yamagishilab/project-CURRENNT-scripts (recipes for various models using CURRENT toolkit)
Follow the instruction on the README and compile CURRENT toolkit.
We use project-CURRENNT-scripts/waveform-modeling/project-WaveNet to train a WaveNet model.
Read the README and follow the instruction to train a WaveNet model.
There are a few important settings you have to configure.
- Configure
upsampling_rate=200 - Configure the number utterances for training
- Configure training data directory
Tacotron uses larger frame shift, 12.5ms. In this case upsampling rate is 16000Hz / 1000 * 12.5ms = 200. You can configure it here.
The number of utterances for training is 1000 by default. It is too small so set an appropriate value for the dataset you use. You can configure it here.
By default training data location is project-CURRENNT-scripts/waveform-modeling/DATA/.
You can configure the location here.
There are two subdirectories, mfbsp is for mel-spectrogram, and wav32k is for wav files.
You can prepare training data by running the following command in this repository, as an example of LJSpeech.
bazel run preprocess_ljspeech_wavenet -- --hparam-json-file=self-attention-tacotron/examples/ljspeech/self-attention-tacotron.json path/to/LJSpeech-1.1 /path/to/mel/output/dir /path/to/wav/output/dir
Please set the location /path/to/mel/output/dir and /path/to/wav/output/dir to the location described above.
You can start pre-processing and training by the following command.
./00_run.sh
You can find how to predict mel-spectrogram with trained Tacotron model here.
Please configure the output directory of the mel-spectrogram as an input directory for WaveNet by the following environment variable.
export TEMP_WAVEFORM_MODEL_INPUT_DIRS=/path/to/mel-spectrogram/dir
Waveform generation can be started by the following command.
./01_gen.sh