Pytorch Implementation of Robust and fine-grained prosody control of end-to-end speech synthesis (Unofficial)

This implementation uses the LibriTTS dataset.

1. dev branch: Tacotron2 with multispeaker (speaker embedding). Speaker information is only consumed by Decoder module, and Attention module doesn't see any of it (as authors' intention).
2. text_side branch: Text-side prosody control model implementation.
3. Speech-side prosody control and Prosody normalization are not implemented in current version, but you can simply add them on top of above branches.

1. NVIDIA GPU + CUDA cuDNN

1. Download and extract the LibriTTS dataset
2. Clone this repo: git clone https://github.com/keonlee9420/Robust_Fine_Grained_Prosody_Control.git
3. CD into this repo: cd Robust_Fine_Grained_Prosody_Control
4. Initialize submodule: git submodule init; git submodule update
5. Update .wav paths: sed -i -- 's,/home/keon/speech-datasets/LibriTTS_preprocessed/train-clean-100/,your_libritts_dataset_folder/,g' filelists/*.txt
   • Alternatively, set load_mel_from_disk=True in hparams.py and update mel-spectrogram paths
6. Install PyTorch 1.0
7. Install Apex
8. Install python requirements or build docker image
   • Install python requirements: pip install -r requirements.txt

1. python train.py --output_directory=outdir --log_directory=logdir
2. (OPTIONAL) tensorboard --logdir=outdir/logdir

(TBD)

1. Not supported in current implementation.

1. Single sample: python inference.py -c checkpoint/path -r reference_audio/wav/path -t "synthesize text"
2. Multi samples: python inference_all.py -c checkpoint/path -r reference_audios/dir/path

N.b. When performing Mel-Spectrogram to Audio synthesis, make sure Tacotron 2 and the Mel decoder were trained on the same mel-spectrogram representation.

@misc{lee2021robust_fine_grained_prosody_control,
  author = {Lee, Keon},
  title = {Robust_Fine_Grained_Prosody_Control},
  year = {2021},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/keonlee9420/Robust_Fine_Grained_Prosody_Control}}
}

WaveGlow Faster than real time Flow-based Generative Network for Speech Synthesis

nv-wavenet Faster than real time WaveNet.

This implementation uses code from the following repos: NVIDIA/Tacotron-2, KinglittleQ/GST-Tacotron

We are thankful to the paper authors, specially Younggun Lee, and Taesu Kim.