PyTorch implementation of DeepMind's Tacotron-2. A deep neural network architecture described in this paper: Natural TTS: synthesis by conditioning Wavenet on MEL spectogram predictions
A whole TTS system is consisted of front-end and back-end. Tacotron 2 and deep voice 3 are excellent front-ends. And WaveNet and WaveGlow are that back-ends.
I decide to rewrite the Tacotron 2 although Rayhane-mamah and NVidia have implemented it. Because I wish I could master and implement the attention mechanism by this project. And many beginners could benifit from its conciseness to understand Tacotron 2.
It aims to cooperate with the my single-Gaussian wavenet(including autoregressive wavenet and parallel wavenet) and waveglow which are realized with PyTorch.
WARNING: Only support mandarin which is symboled with Pinyin. I plan to support more language later.
Data preprocessing is borrowed from the Tacotron-2 implemented with Tensorflow. It is adapted slightly to meet the requirements of my self-made mandarin datasets. You should specific the index txt and output directory like the script I provides.
#!/bin/bash
python preprocess.py --dataset=mandarin \
--input_txt='metadata.csv' \
--base_dir=./ \
--output_dir='dataset_mandarin'It will create an folder which contains the preprocessed data in the direcroty named base_dir/output_dir.
You can refer to the script.
#!/bin/bash
CUDA_VISIBLE_DEVICES=$1 python train.py --data-root=dataset_mandarin/training_data/ \
--run-name="Tacotron2_beta_profile"The alignment at 20710 step with 32 batch size
