Glow-TTS Official Repository

Glow-TTS: A Generative Flow for Text-to-Speech via Monotonic Alignment Search

• Github https://github.com/jaywalnut310/glow-tts

• Paper https://arxiv.org/abs/2005.11129

• Demo https://jaywalnut310.github.io/glow-tts-demo/index.html

Multispeaker enabled Glow-tts

Glow-tts with korean cleaner, enabled multispeaker training (reffering to some of issues).

This repo recommended to be used as a reference for multispeaker training.

_custom : executed with korean cleaners.

_custom_multi : executed with korean cleaners, for multispeaker training.

Single korean speaker demo with KSS is available. link

Korean cleaner

• from Pitchtron https://github.com/hash2430/pitchtron

Solved issues

• Due to apex(commit: 37cdaf4) dependency, I used pytorch 1.3.0 (instead of 1.2.0)

• For multispeaker setting

  • filelist should be in followed format.

    audio_path(*.wav)|speaker_id|transcript related issue

  • Add n_speakers, gin_channels to config is recommended. related issue

  • (TextMelLoader, TextMelCollate) should be replaced with (TextMelSpeakerLoader, TextMelSpeakerCollate) in init.py, train.py

    Also, edit (x, x_lengths, y, y_lengths) to (x, x_lengths, y, y_lengths, g).

  • Usage of speaker information(g) should be delievered explicitly to FlowGenerator. related issue

    generator(x=x, x_lengths=x_lengths, y=y, y_lengths=y_lengths, g=g, gen=False) (I do not know why)

• 'Gradient overflow' might be caused due to data problem. related issue

1. Environments (edited)

• Python==3.6.9
• pytorch==1.3.0
• cython==0.29.12
• librosa==0.7.1
• numpy==1.16.4
• scipy==1.5.4
• nltk==3.6.5

2. Pre-requisites

Please check official repository.

3. Training Example

sh train_custom_multi_ddi.sh configs/base.json base

4. Inference Example

See inference.ipynb