NaturalSpeech 2

• This code is a unofficial implementation of NaturalSpeech 2.
• The algorithm is based on the following paper:

Shen, K., Ju, Z., Tan, X., Liu, Y., Leng, Y., He, L., ... & Bian, J. (2023). NaturalSpeech 2: Latent Diffusion Models are Natural and Zero-Shot Speech and Singing Synthesizers. arXiv preprint arXiv:2304.09116.

Modifications from Paper

• The structure is derived from NaturalSpeech 2, but I made several modifications.
• The audio codec has been changed to HifiCodec from AcademiCodec.
  • This is done to reduce the time spent training a separate audio codec.
  • The model uses 22.05Khz audio, but no audio resampling is applied.
  • To maintain similarity with the paper, it may be better to apply Google's SoundStream instead of HifiCodec, but I couldn't apply SoundStream to this repository because official pyTorch source code or pretrained model was not provided.
    • Although this repository does not use, there is also a c++ or tflite version of Lyra, which may allow the application of SoundStream using it.
  • Meta's Encodec 24K version was also tested, but it could not be trained.
• About CE-RVQ
  • I have observed that the quality of the model with CE-RVQ applied is decreased, so it has been removed from the current implementation.
  • However, this does not necessarily mean that CE-RVQ is ineffective.
  • It could be due to the change of codec in this repository or the presence of bugs in the implemented module.
  • This aspect requires further validation and investigation.
• Information on the segment length σ of the speech prompt during training was not found in the paper and was arbitrarily set.
  • The σ = 3, 5, and 10 seconds used in the evaluation of paper are too long to apply to both the variance predictor and diffusion during training.
  • To ensure stability in pattern usage, half the length of the shortest pattern used in each training is set as σ for each training.
• The target duration is obtained through Alignment learning framework (ALF), rather than being brought in externally.
  • Using external modules such as Montreal Force Alignment (MFA) may have benefits in terms of training speed or stability, but I prioritized simplifying the training process.
• Padding is applied between tokens like 'A <P> B <P> C ....'
  • I could not verify whether there was a difference in performance depending on its usage.

Supported dataset

• To apply zero-shot reported in the paper, I believe that it is necessary to have as many speakers as possible in the training data, but I were unable to test Multilingual LibriSpeech due to current environment.
• Tested
  • LJ Dataset
  • VCTK Dataset
    • This repository used the VCTK092 from Torchaudio(https://datashare.is.ed.ac.uk/bitstream/handle/10283/3443/VCTK-Corpus-0.92.zip)
• Supported but not tested
  • LibriTTS Dataset
  • Mulitilingual LibriSpeech Dataset
    • Only English language dataset generation is tested.

Hyper parameters

Before proceeding, please set the pattern, inference, and checkpoint paths in Hyper_Parameters.yaml according to your environment.

• Sound

  • Setting basic sound parameters.

• Tokens

  • The number of token.
  • After pattern generating, you can see which tokens are included in the dataset at Token_Path.

• Audio_Codec

  • Setting the audio codec.
  • This repository is using HifiCodec, so only the size of the latents output from HifiCodec's encoder is set for reference in other modules.

• Train

  • Setting the parameters of training.

• Inference_Batch_Size

  • Setting the batch size when inference

• Inference_Path

  • Setting the inference path

• Checkpoint_Path

  • Setting the checkpoint path

• Log_Path

  • Setting the tensorboard log path

• Use_Mixed_Precision

  • Setting using mixed precision

• Use_Multi_GPU

  • Setting using multi gpu
  • By the nvcc problem, Only linux supports this option.
  • If this is True, device parameter is also multiple like 0,1,2,3.
  • And you have to change the training command also: please check multi_gpu.sh.

• Device

  • Setting which GPU devices are used in multi-GPU enviornment.
  • Or, if using only CPU, please set '-1'. (But, I don't recommend while training.)

Generate pattern

python Pattern_Generate.py [parameters]

Parameters

• -lj
  • The path of LJSpeech dataset
• -vctk
  • The path of VCTK dataset
• -libri
  • The path of LbiriTTS dataset
• -hp
  • The path of hyperparameter.

About phonemizer

• To phoneme string generate, this repository uses phonimizer library.
• Please refer here to install phonemizer and backend
• In Windows, you need more setting to use phonemizer.
  • Please refer here
  • In conda enviornment, the following commands are useful.

    conda env config vars set PHONEMIZER_ESPEAK_PATH='C:\Program Files\eSpeak NG'
    conda env config vars set PHONEMIZER_ESPEAK_LIBRARY='C:\Program Files\eSpeak NG\libespeak-ng.dll'

Run

Command

Single GPU

python Train.py -hp <path> -s <int>

• -hp <path>

  • The hyper paramter file path
  • This is required.

• -s <int>

  • The resume step parameter.
  • Default is 0.
  • If value is 0, model try to search the latest checkpoint.

Multi GPU

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 OMP_NUM_THREADS=32 python -m torch.distributed.launch --nproc_per_node=8 Train.py --hyper_parameters Hyper_Parameters.yaml --port 54322

• I recommend to check the multi_gpu.sh.

Checkpoint

• This checkpoint was trained in a single GPU environment (RTX4090 x 1) with the VCTK dataset.
• It has limited quality compared to the official demo, and there are issues with the generation for unseen reference.
• While checking the loss flow, I observed the possibility of loss decreasing as training progresses, but it doesn't guarantee an improvement in quality.
• Unfortunately, I don't have personal resources for testing beyond the current state, so I'm releasing the checkpoint after discontinuation.