A (Heavily Documented) TensorFlow Implementation of Tacotron: A Fully End-to-End Text-To-Speech Synthesis Model

Major History

• June 21, 2017. Fourth draft.
  • I've updated the code for TF 1.1 to TF 1.2. Turns out that TF 1.2 has a new api for attention wrapper and more detailed options.
  • I've added a sanity check option to the hyperparams.py file. Basically, it's purpose is to find out if our model is able to learn a handful of training data wihtout caring about generalization. For that, the training was done on a single mini-batch (32 samples) over and over again, and sample generation was based on the same text. I observed a quite smooth training curve for as below, and after around 18K global steps it started to generate recognizable sounds. The sample results after 36K steps are available in the logdir_s folder. It took around seven hours on a single gtx 1080. The pretrained files can be downloaded from here. The training curve looks like this.

• June 4, 2017. Third draft.
  • Some people reported they gained promising results, based on my code. Among them are, @ggsonic, @chief7. To check relevant discussions, see this discussion, or their repo.
  • According @ggsonic, instance normalization worked better than batch normalization.
  • @chief7 trained on pavoque data, a German corpus spoken by a single male actor. He said that instance normalization and zero-masking are good choices.
  • Yuxuan, the first author of the paer, advised me to do sanity-check first with small data, and to adjust hyperparemters since our dataset is different from his. I really appreciate his tips, and hope this would help you.
  • Alex's repo, which is another implementation of Tacotron, seems to be successful in getting promising results with some small dataset. He's working on a big one.
• June 2, 2017.
  • Added train_multiple_gpus.py for 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 good results are not guaranteed. I'll be working on debugging this weekend. (Code reviews and/or contributions are more than welcome!)

Requirements

• NumPy >= 1.11.1
• TensorFlow == 1.2
• librosa
• tqdm

Data

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.

File description

• 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 for training.
• eval.py is for sample synthesis.

Training

• STEP 1. Adjust hyper parameters in hyperparams.py if necessary.
• STEP 2. Download and extract the audio data and its text.
• STEP 3. Run train.py. or train_multi_gpus.py if you have more than one gpu.

Sample Synthesis

• Run eval.py to get samples.

Acknowledgements

I would like to show my respect to Dave, the host of www.audiotreasure.com and the reader of the audio files.