visinger2-zh-jp-multisinger-svs

espnet/visinger2-zh-jp-multisinger-svs: Singing Voice Synthesis Model for SingingSDS

This model is espnet/visinger2-zh-jp-multisinger-svs, a multilingual (Chinese and Japanese) singing voice synthesis (SVS) component. It is a key part of the SingingSDS: A Singing-Capable Spoken Dialogue System for Conversational Roleplay Applications project.

• 📚 Paper: SingingSDS: A Singing-Capable Spoken Dialogue System
• 💻 Project Code on GitHub: SingingSDS
• 🚀 Live Demo on Hugging Face Spaces

<div align="center"> <img src="https://github.com/SingingSDS/SingingSDS/raw/main/assets/demo.png" alt="SingingSDS Interface" style="max-width: 100%; height: auto;"/> </div>

⚠️ Notice

This model is for research use only and is strictly prohibited for any commercial use.
Please make sure to comply with this restriction before using the model.

---

ESPnet2 SVS model

Demo: How to use in ESPnet2

Follow the ESPnet installation instructions if you haven't done that already.

cd espnet
git checkout c9a3878a76e1b738ccfc3aa96912d14e0e2d3134
pip install -e .
cd egs2/mixed/svs1
./run.sh --skip_data_prep false --skip_train true --download_model espnet/visinger2-zh-jp-multisinger-svs

SVS config

<details><summary>expand</summary>

config: conf/tuning/train_visinger2_spk_embed_lang.yaml
print_config: false
log_level: INFO
drop_last_iter: false
dry_run: false
iterator_type: sequence
valid_iterator_type: null
output_dir: exp/svs_train_visinger2_spk_embed_lang_1019
ngpu: 1
seed: 777
num_workers: 0
num_att_plot: 0
dist_backend: nccl
dist_init_method: env://
dist_world_size: null
dist_rank: null
local_rank: 0
dist_master_addr: null
dist_master_port: null
dist_launcher: null
multiprocessing_distributed: false
unused_parameters: true
sharded_ddp: false
use_deepspeed: false
deepspeed_config: null
cudnn_enabled: true
cudnn_benchmark: false
cudnn_deterministic: false
use_tf32: false
collect_stats: false
write_collected_feats: false
max_epoch: 500
patience: null
val_scheduler_criterion:
- valid
- loss
early_stopping_criterion:
- valid
- loss
- min
best_model_criterion:
-   - train
    - total_count
    - max
keep_nbest_models: 15
nbest_averaging_interval: 0
grad_clip: -1
grad_clip_type: 2.0
grad_noise: false
accum_grad: 1
no_forward_run: false
resume: true
train_dtype: float32
use_amp: false
log_interval: 50
use_matplotlib: true
use_tensorboard: true
create_graph_in_tensorboard: false
use_wandb: false
wandb_project: null
wandb_id: null
wandb_entity: null
wandb_name: null
wandb_model_log_interval: -1
detect_anomaly: false
use_adapter: false
adapter: lora
save_strategy: all
adapter_conf: {}
pretrain_path: null
init_param: []
ignore_init_mismatch: false
freeze_param: []
num_iters_per_epoch: 1000
batch_size: 8
valid_batch_size: null
batch_bins: 1000000
valid_batch_bins: null
category_sample_size: 10
train_shape_file:
- exp/svs_stats_raw_phn_None_zh_jp_44100Hz_1019/train/text_shape.phn
- exp/svs_stats_raw_phn_None_zh_jp_44100Hz_1019/train/singing_shape
valid_shape_file:
- exp/svs_stats_raw_phn_None_zh_jp_44100Hz_1019/valid/text_shape.phn
- exp/svs_stats_raw_phn_None_zh_jp_44100Hz_1019/valid/singing_shape
batch_type: sorted
valid_batch_type: null
fold_length:
- 150
- 409600
sort_in_batch: descending
shuffle_within_batch: false
sort_batch: descending
multiple_iterator: false
chunk_length: 500
chunk_shift_ratio: 0.5
num_cache_chunks: 1024
chunk_excluded_key_prefixes: []
chunk_default_fs: null
chunk_max_abs_length: null
chunk_discard_short_samples: true
train_data_path_and_name_and_type:
-   - dump44100_1019/raw/tr_no_dev/text
    - text
    - text
-   - dump44100_1019/raw/tr_no_dev/wav.scp
    - singing
    - sound
-   - dump44100_1019/raw/tr_no_dev/label
    - label
    - duration
-   - dump44100_1019/raw/tr_no_dev/score.scp
    - score
    - score
-   - exp/svs_stats_raw_phn_None_zh_jp_44100Hz_1019/train/collect_feats/pitch.scp
    - pitch
    - npy
-   - exp/svs_stats_raw_phn_None_zh_jp_44100Hz_1019/train/collect_feats/feats.scp
    - feats
    - npy
-   - dump44100_1019/espnet_spk/tr_no_dev/espnet_spk.scp
    - spembs
    - kaldi_ark
-   - dump44100_1019/raw/tr_no_dev/utt2lid
    - lids
    - text_int
valid_data_path_and_name_and_type:
-   - dump44100_1019/raw/dev/text
    - text
    - text
-   - dump44100_1019/raw/dev/wav.scp
    - singing
    - sound
-   - dump44100_1019/raw/dev/label
    - label
    - duration
-   - dump44100_1019/raw/dev/score.scp
    - score
    - score
-   - exp/svs_stats_raw_phn_None_zh_jp_44100Hz_1019/valid/collect_feats/pitch.scp
    - pitch
    - npy
-   - exp/svs_stats_raw_phn_None_zh_jp_44100Hz_1019/valid/collect_feats/feats.scp
    - feats
    - npy
-   - dump44100_1019/espnet_spk/dev/espnet_spk.scp
    - spembs
    - kaldi_ark
-   - dump44100_1019/raw/dev/utt2lid
    - lids
    - text_int
multi_task_dataset: false
allow_variable_data_keys: false
max_cache_size: 0.0
max_cache_fd: 32
allow_multi_rates: false
valid_max_cache_size: null
exclude_weight_decay: false
exclude_weight_decay_conf: {}
optim: adamw
optim_conf:
    lr: 0.0002
    betas:
    - 0.8
    - 0.99
    eps: 1.0e-09
    weight_decay: 0.0
scheduler: exponentiallr
scheduler_conf:
    gamma: 0.998
optim2: adamw
optim2_conf:
    lr: 0.0002
    betas:
    - 0.8
    - 0.99
    eps: 1.0e-09
    weight_decay: 0.0
scheduler2: exponentiallr
scheduler2_conf:
    gamma: 0.998
generator_first: true
skip_discriminator_prob: 0.0
input_size: null
token_list:
- <blank>
- <unk>
- SP
- i@zh
- e@zh
- d@zh
- y@zh
- uo@zh
- sh@zh
- ian@zh
- ai@zh
- w@zh
- n@zh
- u@zh
- x@zh
- j@zh
- l@zh
- h@zh
- b@zh
- iii@zh
- zh@zh
- uei@zh
- m@zh
- ing@zh
- q@zh
- g@zh
- eng@zh
- z@zh
- a@zh
- an@zh
- en@zh
- iou@zh
- ao@zh
- t@zh
- ou@zh
- iang@zh
- AP
- ong@zh
- ang@zh
- ei@zh
- iao@zh
- ie@zh
- f@zh
- r@zh
- k@zh
- ch@zh
- v@zh
- in@zh
- uan@zh
- c@zh
- uang@zh
- s@zh
- a@jp
- ii@zh
- van@zh
- p@zh
- ve@zh
- o@jp
- i@jp
- ia@zh
- uen@zh
- ua@zh
- iong@zh
- u@jp
- e@jp
- k@jp
- uai@zh
- n@jp
- t@jp
- r@jp
- m@jp
- er@zh
- vn@zh
- s@jp
- d@jp
- w@jp
- o@zh
- sh@jp
- g@jp
- N@jp
- y@jp
- b@jp
- ts@jp
- h@jp
- z@jp
- cl@jp
- ch@jp
- j@jp
- f@jp
- p@jp
- ky@jp
- ry@jp
- gy@jp
- ny@jp
- hy@jp
- my@jp
- by@jp
- py@jp
- ty@jp
- dy@jp
- v@jp
- <sos/eos>
odim: null
model_conf: {}
use_preprocessor: true
token_type: phn
bpemodel: null
non_linguistic_symbols: null
cleaner: null
g2p: null
fs: 44100
postfrontend: null
postfrontend_conf: {}
score_feats_extract: syllable_score_feats
score_feats_extract_conf:
    fs: 44100
    n_fft: 2048
    win_length: 2048
    hop_length: 512
feats_extract: fbank
feats_extract_conf:
    n_fft: 2048
    hop_length: 512
    win_length: 2048
    fs: 44100
    fmin: 80
    fmax: 22050
    n_mels: 80
normalize: global_mvn
normalize_conf:
    stats_file: exp/svs_stats_raw_phn_None_zh_jp_44100Hz_1019/train/feats_stats.npz
svs: vits
svs_conf:
    generator_type: visinger2
    vocoder_generator_type: visinger2
    generator_params:
        hidden_channels: 192
        spk_embed_dim: 192
        langs: 3
        global_channels: 256
        segment_size: 20
        text_encoder_attention_heads: 2
        text_encoder_ffn_expand: 4
        text_encoder_blocks: 6
        text_encoder_positionwise_layer_type: conv1d
        text_encoder_positionwise_conv_kernel_size: 3
        text_encoder_positional_encoding_layer_type: rel_pos
        text_encoder_self_attention_layer_type: rel_selfattn
        text_encoder_activation_type: swish
        text_encoder_normalize_before: true
        text_encoder_dropout_rate: 0.1
        text_encoder_positional_dropout_rate: 0.0
        text_encoder_attention_dropout_rate: 0.1
        use_macaron_style_in_text_encoder: true
        use_conformer_conv_in_text_encoder: false
        text_encoder_conformer_kernel_size: -1
        decoder_kernel_size: 7
        decoder_channels: 256
        decoder_upsample_scales:
        - 8
        - 8
        - 4
        - 2
        decoder_upsample_kernel_sizes:
        - 16
        - 16
        - 8
        - 4
        n_harmonic: 64
        decoder_resblock_kernel_sizes:
        - 3
        - 7
        - 11
        decoder_resblock_dilations:
        -   - 1
            - 3
            - 5
        -   - 1
            - 3
            - 5
        -   - 1
            - 3
            - 5
        use_weight_norm_in_decoder: true
        posterior_encoder_kernel_size: 3
        posterior_encoder_layers: 8
        posterior_encoder_stacks: 1
        posterior_encoder_base_dilation: 1
        posterior_encoder_dropout_rate: 0.0
        use_weight_norm_in_posterior_encoder: true
        flow_flows: -1
        flow_kernel_size: 5
        flow_base_dilation: 1
        flow_layers: 4
        flow_dropout_rate: 0.0
        use_weight_norm_in_flow: true
        use_only_mean_in_flow: true
        use_phoneme_predictor: false
        hubert_channels: null
        vocabs: 102
        aux_channels: 80
        generator_type: visinger2
        vocoder_generator_type: visinger2
        fs: 44100
        hop_length: 512
        win_length: 2048
        n_fft: 2048
    discriminator_type: visinger2
    discriminator_params:
        scales: 1
        scale_downsample_pooling: AvgPool1d
        scale_downsample_pooling_params:
            kernel_size: 4
            stride: 2
            padding: 2
        scale_discriminator_params:
            in_channels: 1
            out_channels: 1
            kernel_sizes:
            - 15
            - 41
            - 5
            - 3
            channels: 128
            max_downsample_channels: 1024
            max_groups: 256
            bias: true
            downsample_scales:
            - 4
            - 4
            - 4
            - 4
            nonlinear_activation: LeakyReLU
            nonlinear_activation_params:
                negative_slope: 0.1
            use_weight_norm: true
            use_spectral_norm: false
        follow_official_norm: false
        periods:
        - 2
        - 3
        - 5
        - 7
        - 11
        period_discriminator_params:
            in_channels: 1
            out_channels: 1
            kernel_sizes:
            - 5
            - 3
            channels: 32
            downsample_scales:
            - 3
            - 3
            - 3
            - 3
            - 1
            max_downsample_channels: 1024
            bias: true
            nonlinear_activation: LeakyReLU
            nonlinear_activation_params:
                negative_slope: 0.1
            use_weight_norm: true
            use_spectral_norm: false
        multi_freq_disc_params:
            hidden_channels:
            - 256
            - 256
            - 256
            - 256
            - 256
            domain: double
            mel_scale: true
            divisors:
            - 32
            - 16
            - 8
            - 4
            - 2
            - 1
            - 1
            strides:
            - 1
            - 2
            - 1
            - 2
            - 1
            - 2
            - 1
            sample_rate: 44100
            hop_lengths:
            - 110
            - 220
            - 330
            - 441
            - 551
            - 661
    generator_adv_loss_params:
        average_by_discriminators: false
        loss_type: mse
    discriminator_adv_loss_params:
        average_by_discriminators: false
        loss_type: mse
    feat_match_loss_params:
        average_by_discriminators: false
        average_by_layers: false
        include_final_outputs: true
    mel_loss_params:
        fs: 44100
        n_fft: 2048
        hop_length: 512
        win_length: 2048
        window: hann
        n_mels: 80
        fmin: 0
        fmax: 22050
        log_base: null
    lambda_adv: 1.0
    lambda_mel: 45.0
    lambda_feat_match: 2.0
    lambda_dur: 0.1
    lambda_pitch: 10.0
    lambda_phoneme: 1.0
    lambda_kl: 1.0
    sampling_rate: 44100
    cache_generator_outputs: true
pitch_extract: dio
pitch_extract_conf:
    use_token_averaged_f0: false
    use_log_f0: false
    fs: 44100
    n_fft: 2048
    hop_length: 512
    f0max: 810
    f0min: 80
pitch_normalize: null
pitch_normalize_conf:
    stats_file: exp/svs_stats_raw_phn_None_zh_jp_44100Hz_1019/train/pitch_stats.npz
ying_extract: null
ying_extract_conf: {}
energy_extract: null
energy_extract_conf: {}
energy_normalize: null
energy_normalize_conf: {}
required:
- output_dir
- token_list
version: '202412'
distributed: false

</details>

Citing ESPnet

@inproceedings{wu2024muskits-espnet,
author = {Wu, Yuning and Shi, Jiatong and Yu, Yifeng and Tang, Yuxun and Qian, Tao and Lin, Yueqian and Han, Jionghao and Bai, Xinyi and Watanabe, Shinji and Jin, Qin},
title = {Muskits-ESPnet: A Comprehensive Toolkit for Singing Voice Synthesis in New Paradigm},
year = {2024},
booktitle = {Proceedings of the 32nd ACM International Conference on Multimedia},
}

@inproceedings{watanabe2018espnet,
  author={Shinji Watanabe and Takaaki Hori and Shigeki Karita and Tomoki Hayashi and Jiro Nishitoba and Yuya Unno and Nelson Yalta and Jahn Heymann and Matthew Wiesner and Nanxin Chen and Adithya Renduchintala and Tsubasa Ochiai},
  title={{ESPnet}: End-to-End Speech Processing Toolkit},
  year={2018},
  booktitle={Proceedings of Interspeech},
  pages={2207--2211},
  doi={10.21437/Interspeech.2018-1456},
  url={http://dx.doi.org/10.21437/Interspeech.2018-1456}
}