Tacotron2 Implementation (Unofficial)

This repository provides an unofficial implementation of Tacotron2 using PyTorch. For a more detailed and accurate implementation, it is strongly recommended to refer to NVIDIA’s official Tacotron2 repository. This current implementation serves as an acoustic model to generate mel-spectrograms. To generate actual speech, you will need to integrate a vocoder model provided by torchaudio or retrain one. Most of the code in this repository is based on NVIDIA’s Tacotron2 implementation.

1. Getting Started

Prerequisites

• Docker and NVIDIA GPU Drivers (for GPU support)

Setup Instructions

1. Clone the repository:

   git clone https://github.com/Orca0917/Tacotron2.git
   cd Tacotron2
   

2. Build the Docker image:

   docker build -t tacotron2 .
   

3. Run the Docker container:

   docker run -it --name tacotron2-container --gpus all tacotron2
   

4. Train the model inside the container:

   python train.py
   

2. Result

Once the model is trained, you will obtain a mel-spectrogram as the output of the acoustic model, which can be visualized similarly to the image below:

3. Additional Notes

This implementation focuses on generating mel-spectrograms. To complete the text-to-speech pipeline, you will need to use a vocoder (e.g., WaveNet, Griffin-Lim, or a model from torchaudio) to convert the spectrograms into waveform audio.

4. References

[1] Natural TTS Synthesis By Conditioning Wavenet On Mel Spectrogram Predictions. [Link to Paper]

[2] github NVIDIA/Tacotron2 [Link to GitHub]