Translatotron(without wavenet and auxilary decoder networks)

PyTorch implementation of Direct speech-to-speech translation with a sequence-to-sequence model.

This implementation was on a private Telugu-Hindi dataset which cannot be published but has been tested and verified.

Distributed and Automatic Mixed Precision support relies on NVIDIA's Apex and AMP.

Pre-requisites

1. NVIDIA GPU + CUDA cuDNN

Setup

1. Clone this repo: git clone https://github.com/sam2125/translatotron.git
2. CD into this repo: cd translatotron
3. Initialize submodule: git submodule init; git submodule update
4. Modify the dataset module according to the dataset of your preference. The dataset used was noisy so corresponding filtering was applied to make it fit for training.
5. Install [PyTorch]
6. Install Apex
7. Install python requirements or build docker image
   • Install python requirements: pip install -r requirements.txt

Training

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

Training using a pre-trained model

Training using a pre-trained model can lead to faster convergence

1. Download pretrained published Tacotron 2 model
2. python train.py --output_directory=outdir --log_directory=logdir -c tacotron2_statedict.pt --warm_start

Multi-GPU (distributed) and Automatic Mixed Precision Training

1. python -m multiproc train.py --output_directory=outdir --log_directory=logdir --hparams=distributed_run=True,fp16_run=True

Acknowledgements

This implementation uses code from the following repos: Tacotron 2

Big thanks to the Transalotron paper authors and Tacotron 2 paper authors.