DiffSinger-KR

This code is an implementation of DiffSinger for Korean. The algorithm is based on the following papers:

• Liu, J., Li, C., Ren, Y., Chen, F., & Zhao, Z. (2022, June). Diffsinger: Singing voice synthesis via shallow diffusion mechanism. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 36, No. 10, pp. 11020-11028).
• Xiao, Y., Wang, X., He, L., & Soong, F. K. (2022, May). Improving Fastspeech TTS with Efficient Self-Attention and Compact Feed-Forward Network. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (pp. 7472-7476). IEEE.

Structure

• Structure is based on the DiffSinger, but I made some minor changes.
  • The multi-head attention is changed to linearized attention in FFT Block.
    • Positional encoding is removed.
  • Duration embedding is added.
    • It is based on the scaled positional encoding with very low initial scale.
  • Aux decoder and Diffusion are learned at the same time, not two stage.
• I changed several hyper parameters and data type
  • One of mel or spectrogram is can be selected as a feature type.
  • Token type is changed from phoneme to grapheme.
  • Becuase of the supported vocoder, I changed the sample rate of model to 22050Hz.

Supported dataset

• I fixed some midi score to matching between note and wav F0.
• CSD dataset is used for the training of shared checkpoint.
• Pattern generator.py supports the AIHub Dataset, but I did not used for the training of shared checkpoint.

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 Lyric token.

• Notes

  • The highest note value for embedding.

• Durations

  • The highest duration value for embedding.

• Genres

  • Setting the number of genres.

• Singers

  • Setting the number of singers.

• Duration

  • Min duration is used at pattern generating only.
  • Max duration is decided the maximum time step of model. MLP mixer always use the maximum time step.
  • Equality set the strategy about syllable to grapheme.
    • When True, onset, nucleus, and coda have same length or ±1 difference.
    • When False, onset and coda have Consonant_Duration length, and nucleus has duration - 2 * Consonant_Duration.

• Feature_Type

  • Setting the feature type (Mel or Spectrogram).

• Encoder

  • Setting the encoder(embedding).

• Diffusion

  • Setting the Diffusion denoiser.

• 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

• -csd
  • The path of children's song dataset
• -am
  • The path of AIHub multi-singer song dataset
• -step
  • The note step that is explored when generating patterns.
  • The smaller step is, the more patterns are created in one song.
• -hp
  • The path of hyperparameter.

Training

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.

Inference

• Please check Inference.ipynb

Checkpoint

• Please check Huggingface Space

TODO

• Multi singer version version training with AIHub Multi-Singer Song Dataset