LPCTron

Tacotron2 + LPCNET for complete End-to-End TTS System

Thanks to https://github.com/MlWoo/ for most of changes ( Check this issue Also MlWoo/LPCNet#4)

Prerequisites

librosa , tqdm , matplotlib, lws , unidecode , inflect, falcon, scipy , numpy, keras

sudo apt-get install python-pyaudio sudo apt-get install portaudio19-dev

Quick Run

Checkout the LPCTron , and run tts.sh it converts text_file ( Tacotron2/text.txt ) to test.wav

Steps of Integration

• Checkout Tacatron2 (https://github.com/alokprasad/LPCTron/blob/master/Tacotron-2/Tacotron2-lpcnet_changes.diff)

Training Tacotron2 specially to be used by LPCNET vocoder ( Instead of Wavenet Vocoder)

python3 preprocess.py --base_dir /media/alok/ws/sandbox/lpc_tacatron2/dataset --dataset LJSpeech-1.1

this will generate /dataset/training_data folder

├── training_data (1)
│   ├── audio (*)
│   ├── linear
│   └── mels
    └── train.txt

*all three folder contain npy array, train.txt contains text for each audio.

• this is replaced by script for training to be used by LPCNET

Training LPCNET with LJPSpeech

Since Tacatron2 is trained with LPSpeech its good idea to train with Same data. LPCNET uses a single Merged PCM for training. So you need to remove wav headers from each file and merge them for training.

===============================================================================

sudo apt-get install python3-tk

pip3 install lws --user

python -m pip uninstall pip && sudo apt install python3-pip --reinstall

python3 train.py --input_dir ../dataset/training_data --tacotron_input ../dataset/training_data/train.txt --model='Tacotron'

python3 synthesize.py --model='Tacotron' --mode='eval'

sudo apt-get install python3-tk

pip3 install lws --user python -m pip uninstall pip && sudo apt install python3-pip --reinstall

Download https://jmvalin.ca/misc_stuff/lpcnet_models/lpcnet15_384_10_G16_100.h5

apt-get install -y google-perftools LD_PRELOAD = /usr/lib/x86_64-linux-gnu/libtcmalloc.so.4

python3 train.py --input_dir ../dataset/training_data --tacotron_input ../dataset/training_data/train.txt --model='Tacotron'

python3 synthesize.py --model='Tacotron' --mode='eval'

./test_lpcnet f32_for_lpcnet.f32 test.s16

ffmpeg -f s16le -ar 16k -ac 1 -i test.s16 test-out.wav