This repository is refactoring the training code for the first stage model of Bcome Yukarin: Convert your voice to favorite voice.
- Linux OS
- Python 3.6
pip install -r requirements.txt
To run a Python script in this repository, you should set the environment variable PYTHONPATH
to find the yukarin
library.
For example, you can run scripts/foo.py
with the following command:
PYTHONPATH=`pwd` python scripts/foo.py
Put input/target voice data in two directories (ex. input_wav
and target_wav
).
These data should be same file names.
Create input/target acoustic feature files from each voice data.
python scripts/extract_acoustic_feature.py \
-i './input_wav/*' \
-o './input_feature/'
python scripts/extract_acoustic_feature.py \
-i './target_wav/*' \
-o './target_feature/'
Align input and target acoustic features in time direction.
In the following example, create the alignment data between input_feature
and target_feature
into aligned_indexes
.
python scripts/extract_align_indexes.py \
-i1 './input_feature/*.npy' \
-i2 './target_feature/*.npy' \
-o './aligned_indexes/'
Calculate frequency statistics for input and target voice data. Statistics are needed for voice pitch conversion.
python scripts/extract_f0_statistics.py \
-i './input_feature/*.npy' \
-o './input_statistics.npy'
python scripts/extract_f0_statistics.py \
-i './target_feature/*.npy' \
-o './target_statistics.npy'
Modify input_glob
, target_glob
and indexes_glob
in sample_config.json
, then can train.
python train.py \
sample_config.json \
./model_stage1/
Put the test input voice data in a directory (ex. test_wav
), and run voice_change.py
.
python scripts/voice_change.py \
--model_dir './model_stage1' \
--config_path './model_stage1/config.json' \
--input_statistics 'input_statistics.npy' \
--target_statistics 'target_statistics.npy' \
--output_sampling_rate 24000 \
--disable_dataset_test \
--test_wave_dir './test_wav/' \
--output_dir './output/'
Become Yukarin's Second Stage Model can improve the quality of the converted voice.
Train the second stage model referring to Second Stage Model in Become Yukarin.
Put the test input voice data in a directory (ex. test_wav
), and run voice_change_with_second_stage.py
.
python scripts/voice_change_with_second_stage.py \
--voice_changer_model_dir './model_stage1' \
--voice_changer_config './model_stage1/config.json' \
--super_resolution_model './model_stage2/' \
--super_resolution_config './model_stage2/config.json' \
--input_statistics 'input_statistics.npy' \
--target_statistics 'target_statistics.npy' \
--out_sampling_rate 24000 \
--disable_dataset_test \
--dataset_target_wave_dir '' \
--test_wave_dir './test_wav' \
--output_dir './output/'