Skip to content

Official implementation for the paper Fine-grained style control in transformer-based text-to-speech synthesis.

License

Notifications You must be signed in to change notification settings

b04901014/FG-transformer-TTS

Repository files navigation

LST-TTS

Official implementation for the paper Fine-grained style control in transformer-based text-to-speech synthesis. Submitted to ICASSP 2022. Audio samples/demo for our system can be accessed here

  • Mar. 5 2022: Fixed a inference bug of not passing the causal mask, quality of samples should be slightly better. (I have not updated the demos with this.)

Setting up submodules

git submodule update --init --recursive

Get the waveglow vocoder checkpoint from here (This is from the NVIDIA official WaveGlow repo).

Setup environment

See docker/Dockerfile for the packages need to be installed.

Dataset preprocessing

python preprocess_LJSpeech.py --datadir LJSpeechDir --outputdir OutputDir

Get the leading and trailing scilence marks from this repo, and put vctk-silences.0.92.txt in your VCTK dataset directory.

python preprocess_VCTK.py --datadir VCTKDir --outputdir Output_Train_Dir
python preprocess_VCTK.py --datadir VCTKDir --outputdir Output_Test_Dir --make_test_set
  • --make_test_set: specify this flag to process the speakers in the test set, otherwise only process training speakers.

Training

LJSpeech

python train_TTS.py --precision 16 \
                    --datadir FeatureDir \
                    --vocoder_ckpt_path WaveGlowCKPT_PATH \
                    --sampledir SampleDir \
                    --batch_size 128 \
                    --check_val_every_n_epoch 50 \
                    --use_guided_attn \
                    --training_step 250000 \
                    --n_guided_steps 250000 \
                    --saving_path Output_CKPT_DIR \
                    --datatype LJSpeech \
                    [--distributed]
  • --distributed: enable DDP multi-GPU training
  • --batch_size: batch size per GPU, scale down if you train with multi-GPU and want to keep the same batch size
  • --check_val_every_n_epoch: sample and validate every n epoch
  • --datadir: output directory of the preprocess scripts

VCTK

python train_TTS.py --precision 16 \
                    --datadir FeatureDir \
                    --vocoder_ckpt_path WaveGlowCKPT_PATH \
                    --sampledir SampleDir \
                    --batch_size 64 \
                    --check_val_every_n_epoch 50 \
                    --use_guided_attn \
                    --training_step 150000 \
                    --n_guided_steps 150000 \
                    --etts_checkpoint LJSpeech_Model_CKPT \
                    --saving_path Output_CKPT_DIR \
                    --datatype VCTK \
                    [--distributed]
  • --etts_checkpoint: the checkpoint path of pretrained model (on LJ Speech)

Synthesis

We provide examples for synthesis of the system in synthesis.py, you can adjust this script to your own usage. Example to run synthesis.py:

python synthesis.py --etts_checkpoint VCTK_Model_CKPT \
                    --sampledir SampleDir \
                    --datatype VCTK \
                    --vocoder_ckpt_path WaveGlowCKPT_PATH

Pretrained checkpoints

We provide pretrained checkpoints on LJ Speech and VCTK. The model is a little large since it contains all the training and optimizer states.

About

Official implementation for the paper Fine-grained style control in transformer-based text-to-speech synthesis.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published