vits2-vietnam

Follow like this:

Install environment

conda create -y -n whisper python=3.8
conda activate whisper
conda install cudatoolkit=11.3.1 cudnn=8.2.1
pip install -r requirements.txt

Process data

Prepare dataset

• if download from youtube: yt-dlp -x --audio-format wav --batch-file youtube_links.txt
• python Step0_merge_input.py (create each file <= 50k character )
• Upload text to vbee.vn => download
• python Step1_split_audio.py
• python Step2_transcription_whisper.py
• python Step2_transcription_wav2vec.py
• python Step3_map_file.py

Create label

• map predict with ground truth => file map

• check file map with predict

• If use from raw audio from youtube use postprocess.ipynb and download in folder crawl youtube

• pyton Step4_create_label.py

• remember cover audio to mono: python check_type_audio.py

Trainning in vits

git clone https://github.com/p0p4k/vits2_pytorch.git 

cd vit2_pytorch

cd monotonic_align

python setup.py build_ext --inplace

cd ..

• add line "_letters = '0123456789aáảàãạâấẩầẫậăắẳằẵặbcdđeéẻèẽẹêếểềễệfghiíỉìĩịjklmnoóỏòõọôốổồỗộơớởờỡợpqrstuúủùũụưứửừữựvwxyýỷỳỹỵz'" in text/symbols.py

• add line "logging.getLogger('numba').setLevel(logging.WARNING)" in utils.py (ignore warning numba)

• copy file vits2_blv_AQ.json into config

python preprocess.py --text_index 1 --filelists filelists/infore_audio_text_train_filelist.txt filelists/infore_audio_text_val_filelist.txt filelists/infore_audio_text_test_filelist.txt --text_cleaners basic_cleaners

python train.py -c config/vits2_blv_AQ -m blv_AQ

Credits

• ViSV2TTS
• faster_whisper