Jaehyeon Kim, Sungwon Kim, Jungil Kong, and Sungroh Yoon
In our recent paper, we propose Glow-TTS: A Generative Flow for Text-to-Speech via Monotonic Alignment Search.
Recently, text-to-speech (TTS) models such as FastSpeech and ParaNet have been proposed to generate mel-spectrograms from text in parallel. Despite the advantages, the parallel TTS models cannot be trained without guidance from autoregressive TTS models as their external aligners. In this work, we propose Glow-TTS, a flow-based generative model for parallel TTS that does not require any external aligner. We introduce Monotonic Alignment Search (MAS), an internal alignment search algorithm for training Glow-TTS. By leveraging the properties of flows, MAS searches for the most probable monotonic alignment between text and the latent representation of speech. Glow-TTS obtains an order-of-magnitude speed-up over the autoregressive TTS model, Tacotron 2, at synthesis with comparable speech quality, requiring only 1.5 seconds to synthesize one minute of speech in end-to-end. We further show that our model can be easily extended to a multi-speaker setting.
|Glow-TTS at training||Glow-TTS at inference|
1. Environments we use
For Mixed-precision training, we use apex; commit: 37cdaf4
a) Download and extract the LJ Speech dataset, then rename or create a link to the dataset folder:
ln -s /path/to/LJSpeech-1.1/wavs DUMMY
b) Initialize WaveGlow submodule:
git submodule init; git submodule update
Don't forget to download pretrained WaveGlow model and place it into the waveglow folder.
c) Build Monotonic Alignment Search Code (Cython):
cd monotonic_align; python setup.py build_ext --inplace
3. Training Example
sh train_ddi.sh configs/base.json base
4. Inference Example