mio/amadeus

ESPnet2 TTS model

mio/amadeus

This model was trained by mio using amadeus recipe in espnet.

Demo: How to use in ESPnet2

Follow the ESPnet installation instructions
if you haven’t done that already.
cd espnet<br /> git checkout d5b5ec7b2e77bd3e10707141818b7e6c57ac6b3f<br /> pip install -e .<br /> cd egs2/amadeus/tts1<br /> ./run.sh --skip_data_prep false --skip_train true --download_model mio/amadeus<br />

TTS config

expand

config: conf/tuning/finetune_vits.yaml<br /> print_config: false<br /> log_level: INFO<br /> dry_run: false<br /> iterator_type: sequence<br /> output_dir: exp/tts_amadeus_vits_finetune_from_jsut_32_sentence<br /> ngpu: 1<br /> seed: 777<br /> num_workers: 4<br /> num_att_plot: 3<br /> dist_backend: nccl<br /> dist_init_method: env://<br /> dist_world_size: null<br /> dist_rank: null<br /> local_rank: 0<br /> dist_master_addr: null<br /> dist_master_port: null<br /> dist_launcher: null<br /> multiprocessing_distributed: false<br /> unused_parameters: true<br /> sharded_ddp: false<br /> cudnn_enabled: true<br /> cudnn_benchmark: false<br /> cudnn_deterministic: false<br /> collect_stats: false<br /> write_collected_feats: false<br /> max_epoch: 2000<br /> patience: null<br /> val_scheduler_criterion:<br /> - valid<br /> - loss<br /> early_stopping_criterion:<br /> - valid<br /> - loss<br /> - min<br /> best_model_criterion:<br /> - - train<br /> - total_count<br /> - max<br /> keep_nbest_models: 3<br /> nbest_averaging_interval: 0<br /> grad_clip: -1<br /> grad_clip_type: 2.0<br /> grad_noise: false<br /> accum_grad: 1<br /> no_forward_run: false<br /> resume: true<br /> train_dtype: float32<br /> use_amp: false<br /> log_interval: 50<br /> use_matplotlib: true<br /> use_tensorboard: true<br /> create_graph_in_tensorboard: false<br /> use_wandb: true<br /> wandb_project: amadeus<br /> wandb_id: null<br /> wandb_entity: null<br /> wandb_name: null<br /> wandb_model_log_interval: -1<br /> detect_anomaly: false<br /> pretrain_path: null<br /> init_param:<br /> - downloads/f3698edf589206588f58f5ec837fa516/exp/tts_train_vits_raw_phn_jaconv_pyopenjtalk_accent_with_pause/train.total_count.ave_10best.pth:tts:tts<br /> ignore_init_mismatch: false<br /> freeze_param: []<br /> num_iters_per_epoch: null<br /> batch_size: 20<br /> valid_batch_size: null<br /> batch_bins: 5000000<br /> valid_batch_bins: null<br /> train_shape_file:<br /> - exp/tts_stats_raw_linear_spectrogram_phn_jaconv_pyopenjtalk_accent_with_pause/train/text_shape.phn<br /> - exp/tts_stats_raw_linear_spectrogram_phn_jaconv_pyopenjtalk_accent_with_pause/train/speech_shape<br /> valid_shape_file:<br /> - exp/tts_stats_raw_linear_spectrogram_phn_jaconv_pyopenjtalk_accent_with_pause/valid/text_shape.phn<br /> - exp/tts_stats_raw_linear_spectrogram_phn_jaconv_pyopenjtalk_accent_with_pause/valid/speech_shape<br /> batch_type: numel<br /> valid_batch_type: null<br /> fold_length:<br /> - 150<br /> - 204800<br /> sort_in_batch: descending<br /> sort_batch: descending<br /> multiple_iterator: false<br /> chunk_length: 500<br /> chunk_shift_ratio: 0.5<br /> num_cache_chunks: 1024<br /> train_data_path_and_name_and_type:<br /> - - dump/22k/raw/train/text<br /> - text<br /> - text<br /> - - dump/22k/raw/train/wav.scp<br /> - speech<br /> - sound<br /> valid_data_path_and_name_and_type:<br /> - - dump/22k/raw/dev/text<br /> - text<br /> - text<br /> - - dump/22k/raw/dev/wav.scp<br /> - speech<br /> - sound<br /> allow_variable_data_keys: false<br /> max_cache_size: 0.0<br /> max_cache_fd: 32<br /> valid_max_cache_size: null<br /> optim: adamw<br /> optim_conf:<br /> lr: 0.0001<br /> betas:<br /> - 0.8<br /> - 0.99<br /> eps: 1.0e-09<br /> weight_decay: 0.0<br /> scheduler: exponentiallr<br /> scheduler_conf:<br /> gamma: 0.999875<br /> optim2: adamw<br /> optim2_conf:<br /> lr: 0.0001<br /> betas:<br /> - 0.8<br /> - 0.99<br /> eps: 1.0e-09<br /> weight_decay: 0.0<br /> scheduler2: exponentiallr<br /> scheduler2_conf:<br /> gamma: 0.999875<br /> generator_first: false<br /> token_list:<br /> - <blank><br /> - <unk><br /> - '1'<br /> - '2'<br /> - '0'<br /> - '3'<br /> - '4'<br /> - '-1'<br /> - '5'<br /> - a<br /> - o<br /> - '-2'<br /> - i<br /> - '-3'<br /> - u<br /> - e<br /> - k<br /> - n<br /> - t<br /> - '6'<br /> - r<br /> - '-4'<br /> - s<br /> - N<br /> - m<br /> - pau<br /> - '7'<br /> - sh<br /> - d<br /> - g<br /> - w<br /> - '8'<br /> - U<br /> - '-5'<br /> - I<br /> - cl<br /> - h<br /> - y<br /> - b<br /> - '9'<br /> - j<br /> - ts<br /> - ch<br /> - '-6'<br /> - z<br /> - p<br /> - '-7'<br /> - f<br /> - ky<br /> - ry<br /> - '-8'<br /> - gy<br /> - '-9'<br /> - hy<br /> - ny<br /> - '-10'<br /> - by<br /> - my<br /> - '-11'<br /> - '-12'<br /> - '-13'<br /> - py<br /> - '-14'<br /> - '-15'<br /> - v<br /> - '10'<br /> - '-16'<br /> - '-17'<br /> - '11'<br /> - '-21'<br /> - '-20'<br /> - '12'<br /> - '-19'<br /> - '13'<br /> - '-18'<br /> - '14'<br /> - dy<br /> - '15'<br /> - ty<br /> - '-22'<br /> - '16'<br /> - '18'<br /> - '19'<br /> - '17'<br /> - <sos/eos><br /> odim: null<br /> model_conf: {}<br /> use_preprocessor: true<br /> token_type: phn<br /> bpemodel: null<br /> non_linguistic_symbols: null<br /> cleaner: jaconv<br /> g2p: pyopenjtalk_accent_with_pause<br /> feats_extract: linear_spectrogram<br /> feats_extract_conf:<br /> n_fft: 1024<br /> hop_length: 256<br /> win_length: null<br /> normalize: null<br /> normalize_conf: {}<br /> tts: vits<br /> tts_conf:<br /> generator_type: vits_generator<br /> generator_params:<br /> hidden_channels: 192<br /> spks: -1<br /> global_channels: -1<br /> segment_size: 32<br /> text_encoder_attention_heads: 2<br /> text_encoder_ffn_expand: 4<br /> text_encoder_blocks: 6<br /> text_encoder_positionwise_layer_type: conv1d<br /> text_encoder_positionwise_conv_kernel_size: 3<br /> text_encoder_positional_encoding_layer_type: rel_pos<br /> text_encoder_self_attention_layer_type: rel_selfattn<br /> text_encoder_activation_type: swish<br /> text_encoder_normalize_before: true<br /> text_encoder_dropout_rate: 0.1<br /> text_encoder_positional_dropout_rate: 0.0<br /> text_encoder_attention_dropout_rate: 0.1<br /> use_macaron_style_in_text_encoder: true<br /> use_conformer_conv_in_text_encoder: false<br /> text_encoder_conformer_kernel_size: -1<br /> decoder_kernel_size: 7<br /> decoder_channels: 512<br /> decoder_upsample_scales:<br /> - 8<br /> - 8<br /> - 2<br /> - 2<br /> decoder_upsample_kernel_sizes:<br /> - 16<br /> - 16<br /> - 4<br /> - 4<br /> decoder_resblock_kernel_sizes:<br /> - 3<br /> - 7<br /> - 11<br /> decoder_resblock_dilations:<br /> - - 1<br /> - 3<br /> - 5<br /> - - 1<br /> - 3<br /> - 5<br /> - - 1<br /> - 3<br /> - 5<br /> use_weight_norm_in_decoder: true<br /> posterior_encoder_kernel_size: 5<br /> posterior_encoder_layers: 16<br /> posterior_encoder_stacks: 1<br /> posterior_encoder_base_dilation: 1<br /> posterior_encoder_dropout_rate: 0.0<br /> use_weight_norm_in_posterior_encoder: true<br /> flow_flows: 4<br /> flow_kernel_size: 5<br /> flow_base_dilation: 1<br /> flow_layers: 4<br /> flow_dropout_rate: 0.0<br /> use_weight_norm_in_flow: true<br /> use_only_mean_in_flow: true<br /> stochastic_duration_predictor_kernel_size: 3<br /> stochastic_duration_predictor_dropout_rate: 0.5<br /> stochastic_duration_predictor_flows: 4<br /> stochastic_duration_predictor_dds_conv_layers: 3<br /> vocabs: 85<br /> aux_channels: 513<br /> discriminator_type: hifigan_multi_scale_multi_period_discriminator<br /> discriminator_params:<br /> scales: 1<br /> scale_downsample_pooling: AvgPool1d<br /> scale_downsample_pooling_params:<br /> kernel_size: 4<br /> stride: 2<br /> padding: 2<br /> scale_discriminator_params:<br /> in_channels: 1<br /> out_channels: 1<br /> kernel_sizes:<br /> - 15<br /> - 41<br /> - 5<br /> - 3<br /> channels: 128<br /> max_downsample_channels: 1024<br /> max_groups: 16<br /> bias: true<br /> downsample_scales:<br /> - 2<br /> - 2<br /> - 4<br /> - 4<br /> - 1<br /> nonlinear_activation: LeakyReLU<br /> nonlinear_activation_params:<br /> negative_slope: 0.1<br /> use_weight_norm: true<br /> use_spectral_norm: false<br /> follow_official_norm: false<br /> periods:<br /> - 2<br /> - 3<br /> - 5<br /> - 7<br /> - 11<br /> period_discriminator_params:<br /> in_channels: 1<br /> out_channels: 1<br /> kernel_sizes:<br /> - 5<br /> - 3<br /> channels: 32<br /> downsample_scales:<br /> - 3<br /> - 3<br /> - 3<br /> - 3<br /> - 1<br /> max_downsample_channels: 1024<br /> bias: true<br /> nonlinear_activation: LeakyReLU<br /> nonlinear_activation_params:<br /> negative_slope: 0.1<br /> use_weight_norm: true<br /> use_spectral_norm: false<br /> generator_adv_loss_params:<br /> average_by_discriminators: false<br /> loss_type: mse<br /> discriminator_adv_loss_params:<br /> average_by_discriminators: false<br /> loss_type: mse<br /> feat_match_loss_params:<br /> average_by_discriminators: false<br /> average_by_layers: false<br /> include_final_outputs: true<br /> mel_loss_params:<br /> fs: 22050<br /> n_fft: 1024<br /> hop_length: 256<br /> win_length: null<br /> window: hann<br /> n_mels: 80<br /> fmin: 0<br /> fmax: null<br /> log_base: null<br /> lambda_adv: 1.0<br /> lambda_mel: 45.0<br /> lambda_feat_match: 2.0<br /> lambda_dur: 1.0<br /> lambda_kl: 1.0<br /> sampling_rate: 22050<br /> cache_generator_outputs: true<br /> pitch_extract: null<br /> pitch_extract_conf: {}<br /> pitch_normalize: null<br /> pitch_normalize_conf: {}<br /> energy_extract: null<br /> energy_extract_conf: {}<br /> energy_normalize: null<br /> energy_normalize_conf: {}<br /> required:<br /> - output_dir<br /> - token_list<br /> version: '202207'<br /> distributed: false<br />

Citing ESPnet

@inproceedings{watanabe2018espnet,<br /> 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},<br /> title={{ESPnet}: End-to-End Speech Processing Toolkit},<br /> year={2018},<br /> booktitle={Proceedings of Interspeech},<br /> pages={2207--2211},<br /> doi={10.21437/Interspeech.2018-1456},<br /> url={http://dx.doi.org/10.21437/Interspeech.2018-1456}<br /> }<br /> @inproceedings{hayashi2020espnet,<br /> title={{Espnet-TTS}: Unified, reproducible, and integratable open source end-to-end text-to-speech toolkit},<br /> author={Hayashi, Tomoki and Yamamoto, Ryuichi and Inoue, Katsuki and Yoshimura, Takenori and Watanabe, Shinji and Toda, Tomoki and Takeda, Kazuya and Zhang, Yu and Tan, Xu},<br /> booktitle={Proceedings of IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},<br /> pages={7654--7658},<br /> year={2020},<br /> organization={IEEE}<br /> }<br />

or arXiv:
@misc{watanabe2018espnet,<br /> title={ESPnet: End-to-End Speech Processing Toolkit},<br /> 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},<br /> year={2018},<br /> eprint={1804.00015},<br /> archivePrefix={arXiv},<br /> primaryClass={cs.CL}<br /> }<br />