A (Heavily Documented) TensorFlow Implementation of Tacotron: A Fully End-to-End Text-To-Speech Synthesis Model
- June 2, 2017. Added
train_multiple_gpus.pyfor multiple gpus. - June 1, 2017. Second draft. I corrected some mistakes with the help of several contributors (THANKS!), and re-factored source codes so that they are more readable and modular. So far, I couldn't get any promising results.
- May 17, 2017. First draft. You can run it following the steps below, but probably you should get poor results. I'll be working on debugging this weekend. (Code reviews and/or contributions are more than welcome!)
- NumPy >= 1.11.1
- TensorFlow >= 1.1
- librosa
- scipy
Since the original paper was based on their internal data, I use a freely available one, instead.
The World English Bible is a public domain update of the American Standard Version of 1901 into modern English. Its text and audio recordings are freely available here. Unfortunately, however, each of the audio files matches a chapter, not a verse, so is too long for many machine learning tasks. I had someone slice them by verse manually. You can download the audio data and its text from my dropbox.
- hyperparams.py: includes all hyper parameters that are needed.
- prepro.py: loads vocabulary, training/evaluation data.
- data_load.py: loads data and put them in queues so multiple mini-bach data are generated in parallel.
- utils.py: has several custom operational functions.
- modules.py: contains building blocks for encoding/decoding networks.
- networks.py: has three core networks, that is, encoding, decoding, and postprocessing network.
- train.py: is in charge of training.
- eval.py: is in charge of sample synthesis.
- STEP 1. Adjust hyper parameters in
hyperparams.pyif necessary. - STEP 2. Download and extract the audio data and its text.
- STEP 3. Run
train.py. ortrain_multiple_gpus.pyif you have more than one gpu.
- Run
eval.pyto get samples.
I would like to show my respect to Dave, the host of www.audiotreasure.com and the reader of the audio files.