UnOfficial PyTorch implementation of LightSpeech: Lightweight and Fast Text to Speech with Neural Architecture Search. This repo uses the FastSpeech 2 implementation of Espnet as a base. This repo only implements the final version of LightSpeech model not the Neural Architecture Search as mentioned in paper.

But I am able to compress only 3x (from 27 M to 7.99 M trainable parameters) not 15x.

Requirements :

All code written in Python 3.6.2 .

  • Install Pytorch

Before installing pytorch please check your Cuda version by running following command : nvcc --version

pip install torch torchvision

In this repo I have used Pytorch 1.6.0 for torch.bucketize feature which is not present in previous versions of PyTorch.

  • Installing other requirements :
pip install -r requirements.txt
  • To use Tensorboard install tensorboard version 1.14.0 seperatly with supported tensorflow (1.14.0)

For Preprocessing :

filelists folder contains MFA (Motreal Force aligner) processed LJSpeech dataset files so you don’t need to align text with audio (for extract duration) for LJSpeech dataset. For other dataset follow instruction here. For other pre-processing run following command :

python .\ -d path_of_wavs -c configs/default.yaml

For finding the min and max of F0 and Energy

python .\

Update the following in by min and max of F0 and Energy

p_min = Min F0/pitch
p_max = Max F0
e_min = Min energy
e_max = Max energy

For training

 python --outdir etc -c configs/default.yaml -n "name"

For inference


python .\ -c .\configs\default.yaml -p .\checkpoints\first_1\xyz.pyt --out output --text "ModuleList can be indexed like a regular Python list but modules it contains are properly registered."

For TorchScript Export

python -c configs/default.yaml -n fastspeech_scrip --outdir etc

Checkpoint and samples: