Skip to content

Latest commit

 

History

History
101 lines (70 loc) · 2.62 KB

readme.md

File metadata and controls

101 lines (70 loc) · 2.62 KB

Tacotron

An implementation of Tacotron described in the paper using pytorch. Tacotron: Towards End-to-End Speech Synthesis

Published in INTERSPEECH 2017

Requirement

  • torch 1.3.0
  • falcon 1.2.0
  • inflect 0.2.5
  • librosa 0.5.1
  • numpy 1.13.3
  • scipy 1.0.0
  • Unidecode 0.4.21
  • pandas 0.21.0

Datasets

Pre-trained model

Model training

Train using LJ-Speech dataset

python train.py

Train using KSS-dataset

  1. Change options in hyperparams.py
  • cleaners option (26-th line) : from 'english_cleaners' to 'korean_cleaners'
  • dataset option (29-th line) : from 'LJSpeech' to 'KSS'
  • data_path option (30-th line)

  1. Change the sample sentences for generating TTS wav files from english to korean during training. (xx-th line in train.py)

python train.py

Tensorboard

  • You can see the train loss graph.
  • Furthermore, you can listen to generated wav files during training.
Loss wav_files 
tensorboard --logdir=runs

Generate TTS wav files

  1. Download pre-trained model.
  1. Change option in hyperparams.py
  • If you want to generate english wav files, cleaners option (26-th line) should be 'english_cleaners'
  • And if you want to generate korean wav files, cleaners option (26-th line) should be 'korean_cleaners'
  1. Generate TTS wav files
python eval.py --checkpoint_path ./pre_trained_model_path

Experimental Results

Train loss

LJ-Speech KSS

TTS wav files

LJ-results(English)

KSS-results(korean)

Comments

If you have any questions or comments on my codes, please email to me. [email protected]

Reference

[1] https://github.com/soobinseo/Tacotron-pytorch

[2] https://github.com/hccho2/Tacotron-Wavenet-Vocoder-Korean