/TTS

Deep learning for Text2Speech

Primary LanguageJupyter NotebookMozilla Public License 2.0MPL-2.0

TTS

This project is a part of Mozilla Common Voice. TTS aims a Text2Speech engine low in cost and high in quality. To begin with, you can hear a sample here.

The model here is highly inspired from Tacotron: A Fully End-to-End Text-To-Speech Synthesis Model however, it has many important updates over the baseline model that make training faster and computationally very efficient. Feel free to experiment new ideas and propose changes.

You can find here a brief note pointing possible TTS architectures and their comparisons.

Requirements and Installation

Highly recommended to use miniconda for easier installation.

  • python 3.6
  • pytorch 0.4
  • librosa
  • tensorboard
  • tensorboardX
  • matplotlib
  • unidecode

Install TTS using setup.py. It will install all of the requirements automatically and make TTS available to all python environment as an ordinary python module. This makes things easier to run your model outside of the project folder.

python setup.py develop

Or you can use requirements.txt to install the requirements only.

pip install -r requirements.txt

Checkpoints and Audio Samples

Checkout here to compare the samples (except the first) below.

Models Commit Audio Sample Details
iter-62410 99d56f7 link First model with plain Tacotron implementation.
iter-170K e00bc66 link More stable and longer trained model.
iter-270K 256ed63 link Stop-Token prediction is added, to detect end of speech.
Best: iter-120K bf7590 link Better for longer sentences

Example Model Outputs

Below you see model state after 16K iterations with batch-size 32.

"Recent research at Harvard has shown meditating for as little as 8 weeks can actually increase the grey matter in the parts of the brain responsible for emotional regulation and learning."

Audio output: https://soundcloud.com/user-565970875/iter16k-f48c3b

example_model_output

Runtime

The most time-consuming part is the vocoder algorithm (Griffin-Lim) which runs on CPU. By setting its number of iterations, you might have faster execution with a small loss of quality. Some of the experimental values are below.

Sentence: "It took me quite a long time to develop a voice, and now that I have it I'm not going to be silent."

Audio length is approximately 6 secs.

Time (secs) System # GL iters
2.00 GTX1080Ti 30
3.01 GTX1080Ti 60

Data

Currently TTS provides data loaders for datasets depicted below. It is also very is to adapt new datasets with few changes.

Training and Fine-tuning

Split metadata.csv into train and validation subsets respectively metadata_train.csv and metadata_val.csv. Note that having a validation split does not work well as oppose to other ML problems since at the validation time model generates spectrogram slices without "Teacher-Forcing" and that leads misalignment between the ground-truth and the prediction. Therefore, validation loss does not really show the model performance. Rather, you might use the all data for training and check the model performance by relying on human inspection.

shuf metadata.csv > metadata_shuf.csv
head -n 12000 metadata_shuf.csv > metadata_train.csv
tail -n 11000 metadata_shuf.csv > metadata_val.csv

To train a new model, you need to define a config.json file (simple template below) and call with the command below.

train.py --config_path config.json

To fine-tune a model, use --restore_path.

train.py --config_path config.json --restore_path /path/to/your/model.pth.tar

If you like to use specific set of GPUs, you need set an environment variable. The code uses automatically all the available GPUs for data parallel training. If you don't specify the GPUs, it uses the all.

CUDA_VISIBLE_DEVICES="0,1,4" train.py --config_path config.json

Each run creates an experiment folder with some meta information, under the folder you set in config.json. Also a copy of config.json is moved under the experiment folder for reproducibility.

In case of any error or intercepted execution, if there is no checkpoint yet under the execution folder, the whole folder is going to be removed.

You can also enjoy Tensorboard, if you point the Tensorboard argument--logdir to the experiment folder.

Testing

Best way to test your pre-trained network is to use Notebooks under notebooks folder.

What is new with TTS

If you train TTS with LJSpeech dataset, you start to hear reasonable results after 12.5K iterations with batch size 32. This is the fastest training with character based methods up to our knowledge. Out implementation is also quite robust against long sentences.

  • Location sensitive attention (ref). Attention is the vital part of text2speech models. Therefore, it is important to use an attention mechanism that suits the diagonal nature of the problem where the output strictly aligns with the text monotonically. Location sensitive attention performs better by looking into the previous alignment vectors and learns diagonal attention more easily. Yet, I believe there is a good space for research at this front to find a better solution.
  • Attention smoothing with sigmoid (ref). Attention weights are computed by normalized sigmoid values instead of softmax for sharper values. That enables the model to pick multiple highly scored inputs for alignments while reducing the noise.
  • Weight decay (ref). After a certain point of the training, you might observe the model over-fitting. That is, model is able to pronounce words probably better but quality of the speech quality gets lower and sometimes attention alignment gets disoriented.
  • Stop token prediction with an additional module. The original Tacotron model does not propose a stop token to stop the decoding process. Therefore, you need to use heuristic measures to stop the decoder. Here, we prefer to use additional layers at the end to decide when to stop.
  • Applying sigmoid to the model outputs. Since the output values are expected to be in the range [0, 1], we apply sigmoid to make things easier to approximate the expected output distribution.

One common question is to ask why we don't use Tacotron2 architecture. According to our ablation experiments, nothing, except Location Sensitive Attention, improves the baseline perfomance of vanilla Tacotron. Please feel free to offer new changes and pull things off. We are happy to discuss and make things better.

Problems waiting to be solved, based on LJSpeech Dataset

  • Punctuations at the end of a sentence affects the pronounciation of the last word in certain cases. Because punctuation sign is attended by the attention module , that forces network to create a voice signal or at least modify the voice signal being generated for neighboring frames. However, punctuation should only affect the general flow (like ? and !) or pauses in the sentence (. or ,).
  • Simpler stop-token prediction. Right now we use RNN to keep the history of the previous frames. However, we never tested, if something simpler would work as well.
  • Train for better mel-specs. Mel-spectrograms are not good enough to be fed to Neural Vocoder. Therefore, we needs better ways to improve the quality. These might be using adversarial training or some other trick ued by image generation methods. At the end, it is partially a text to image problem.
  • irregular words: "minute", "focus", "aren't" etc. Even though, it might be solved by a larger or better dataset, some of irregular words cause network to mis-pronounce. Irregular means in this context is that written form and the pronounciation of the word have a unique difference.

Major TODOs

  • Implement the model.
  • Generate human-like speech on LJSpeech dataset.
  • Generate human-like speech on a different dataset.
  • Adapting Neural Vocoder.
  • Multi-speaker embedding.

References

Precursor implementations