Tacotron over MXNet

A tech demo of MXNet capabilities consisting of a Tacotron implementation. This is a work in progress.

This project was made during the 8 weeks from 10-2017 to 12-2017 at the PiCampus AI School in Rome.

List of functionalities and TODOs

• [x] Multithreading data iterator

• [x] DSP tools

• [x] CBHG module for spectrograms

• [x] Basic seq2seq example for string reverse. It we'll be used as Tacotron backbone

• [ ] Encoder with CBHG

• [ ] Attention model

• [ ] Custom decoder for processing r * mel_bands spectrograms frames for each time step during the cell unrolling

• [ ] Switch to MXNet 1.0

• [ ] Switch to Gluon

• [ ] Clean up and organize code for better understanding

Getting Started

• install MXNet: pip install -r requirements.txt

• run: python tacotron.py

Using the default setting, a simple dataset will be used as training. Predictions samples will be generated at the end of the training phase.

If you want to train over a big dataset, Kyubyong has cut and formatted this English bible. You can find his dataset here and the CSV text here .

Prerequisites

This project has been developed on

• MXNet 0.12

• librosa

Authors

This project was developed by Alberto Massidda and Stefano Artuso during Pi School's AI programme in Fall 2017.

Acknowledgments

• Thanks to Roberto Barra Chicote for supporting us

• Thanks to Keith Ito https://github.com/keithito, Kyubyong Park https://github.com/Kyubyong for making us start diving in