Tacotron

An implementation of Tacotron speech synthesis in TensorFlow.

Audio Samples

Audio Samples from models trained using this repo.
- The first set was trained for 877K steps on the LJ Speech Dataset
  - Speech started to become intelligble around 20K steps.
  - Although loss continued to decrease, there wasn't much noticable improvement after ~250K steps.
- The second set was trained by @MXGray for 140K steps on the Nancy Corpus.

Background

Earlier this year, Google published a paper, Tacotron: A Fully End-to-End Text-To-Speech Synthesis Model, where they present a neural text-to-speech model that learns to synthesize speech directly from (text, audio) pairs. However, they didn't release their source code or training data. This is an attempt to provide an open-source implementation of the model described in their paper.

The quality isn't as good as Google's demo yet, but hopefully it will get there someday :-). Pull requests are welcome!

Quick Start

Installing dependencies

Install Python 3.
Install the latest version of TensorFlow for your platform. For better performance, install with GPU support if it's available. This code works with TensorFlow 1.3 or 1.4.
Install requirements:
```
pip install -r requirements.txt
```

Using a pre-trained model

Download and unpack a model:

curl http://data.keithito.com/data/speech/tacotron-20170720.tar.bz2 | tar xjC /tmp

Run the demo server:

python3 demo_server.py --checkpoint /tmp/tacotron-20170720/model.ckpt

Point your browser at localhost:9000
- Type what you want to synthesize

Training

Note: you need at least 40GB of free disk space to train a model.

Download a speech dataset.

The following are supported out of the box:
- LJ Speech (Public Domain)
- Blizzard 2012 (Creative Commons Attribution Share-Alike)
You can use other datasets if you convert them to the right format. See TRAINING_DATA.md for more info.

Unpack the dataset into ~/tacotron

After unpacking, your tree should look like this for LJ Speech:

tacotron
  |- LJSpeech-1.0
      |- metadata.csv
      |- wavs

or like this for Blizzard 2012:

tacotron
  |- Blizzard2012
      |- ATrampAbroad
      |   |- sentence_index.txt
      |   |- lab
      |   |- wav
      |- TheManThatCorruptedHadleyburg
          |- sentence_index.txt
          |- lab
          |- wav

Preprocess the data
```
python3 preprocess.py --dataset ljspeech
```
- Use --dataset blizzard for Blizzard data
Train a model
```
python3 train.py
```
Tunable hyperparameters are found in hparams.py. You can adjust these at the command line using the --hparams flag, for example --hparams="batch_size=16,outputs_per_step=2". Hyperparameters should generally be set to the same values at both training and eval time.
Monitor with Tensorboard (optional)
```
tensorboard --logdir ~/tacotron/logs-tacotron
```
The trainer dumps audio and alignments every 1000 steps. You can find these in ~/tacotron/logs-tacotron.
Synthesize from a checkpoint
```
python3 demo_server.py --checkpoint ~/tacotron/logs-tacotron/model.ckpt-185000
```
Replace "185000" with the checkpoint number that you want to use, then open a browser to localhost:9000 and type what you want to speak. Alternately, you can run eval.py at the command line:
```
python3 eval.py --checkpoint ~/tacotron/logs-tacotron/model.ckpt-185000
```
If you set the --hparams flag when training, set the same value here.

Notes and Common Issues

TCMalloc seems to improve training speed and avoids occasional slowdowns seen with the default allocator. You can enable it by installing it and setting LD_PRELOAD=/usr/lib/libtcmalloc.so.
You can train with CMUDict by downloading the dictionary to ~/tacotron/training and then passing the flag --hparams="use_cmudict=True" to train.py. This will allow you to pass ARPAbet phonemes enclosed in curly braces at eval time to force a particular pronunciation, e.g. Turn left on {HH AW1 S S T AH0 N} Street.
If you pass a Slack incoming webhook URL as the --slack_url flag to train.py, it will send you progress updates every 1000 steps.
Occasionally, you may see a spike in loss and the model will forget how to attend (the alignments will no longer make sense). Although it will recover eventually, it may save time to restart at a checkpoint prior to the spike by passing the --restore_step=150000 flag to train.py (replacing 150000 with a step number prior to the spike). Update: a recent fix to gradient clipping by @candlewill may have fixed this.
During eval and training, audio length is limited to max_iters * outputs_per_step * frame_shift_ms milliseconds. With the defaults (max_iters=200, outputs_per_step=5, frame_shift_ms=12.5), this is 12.5 seconds.

If your training examples are longer, you will see an error like this: Incompatible shapes: [32,1340,80] vs. [32,1000,80]

To fix this, you can set a larger value of max_iters by passing --hparams="max_iters=300" to train.py (replace "300" with a value based on how long your audio is and the formula above).

Other Implementations

By Alex Barron: https://github.com/barronalex/Tacotron
By Kyubyong Park: https://github.com/Kyubyong/tacotron

GaoSandy / tacotron