hifigan_vocoder

############################################################################

Model: Unit HiFi-GAN

Tokens: discrete speech units (K-means)

Training: LibriTTS (English)

Authors: Jarod Duret

###################################

Experiment Parameters and setup

################################### local-rank: 0 seed: 4321 __set_seed: !apply:speechbrain.utils.seed_everything [!ref <seed>] output_folder: !ref ./results/hifigan_spk/<seed> save_folder: !ref <output_folder>/save train_log: !ref <output_folder>/train_log.txt progress_sample_path: !ref <output_folder>/samples epochs: 150 keep_checkpoint_interval: 50 use_tensorboard: True

#################################

Data files and pre-processing

################################# data_folder: !PLACEHOLDER # e.g, /datasets/LibriTTS

Data will be automatically downloaded in the data folder

(and resampled to the specified sampling rate)

train_split: [train-clean-100] valid_split: [dev-clean] test_split: [test-clean] train_json: !ref <save_folder>/train.json valid_json: !ref <save_folder>/valid.json test_json: !ref <save_folder>/test.json libritts_subsets: null splits: ["train", "valid", "test"] split_ratio: null skip_prep: False

########################################################

Encoder | HF model

#------------------------------------------------------#

HuBERT | facebook/hubert-large-ll60k

Wav2Vec2 | facebook/wav2vec2-large-960h-lv60-self

WavLM | microsoft/wavlm-large

######################################################## kmeans_folder: speechbrain/SSL_Quantization kmeans_dataset: LibriSpeech-100-360-500 num_clusters: 1000 codes_folder: !ref <save_folder>/codes encoder_type: HuBERT encoder_hub: facebook/hubert-large-ll60k layer: [24] skip_extract: False

speaker_embeddings_folder: !ref <save_folder>/speaker_embeddings speaker_encoder_hub: speechbrain/spkrec-ecapa-voxceleb-mel-spec

################################

Audio Parameters

################################ segment_size: 8960 code_hop_size: 320 sample_rate: 16000 layer_drop: True

hop_length: 256 win_length: 1024 n_mel_channels: 80 n_fft: 1024 mel_fmin: 0.0 mel_fmax: 8000 mel_normalized: False power: 1 norm: "slaney" mel_scale: "slaney" dynamic_range_compression: True

################################

Optimization Hyperparameters

################################ learning_rate: 0.0002 weight_decay: 0.9999 adam_b1: 0.8 adam_b2: 0.99 batch_size: 32

train_dataloader_opts: batch_size: !ref <batch_size> drop_last: False num_workers: 8

valid_dataloader_opts: batch_size: 1 num_workers: 8

test_dataloader_opts: batch_size: 1 num_workers: 8

################################

Model Parameters and model

################################ duration_predictor: False multi_speaker: True normalize_speaker_embeddings: False

embedding params

vocab_size: 6001 # K-means size + 1 for padding 1000x6 embedding_dim: 1024

generator params

in_channels: 1216 out_channels: 1

var_pred_hidden_dim: 128 var_pred_kernel_size: 3 var_pred_dropout: 0.5

###########################################################################################################################################################

version | resblock_type | upsample_kernel_sizes | upsample_factors | resblock_kernel_sizes | upsample_initial_channel | resblock_dilation_sizes

1 | "1" | [16,16,4,4] | [8, 8, 2, 2] | [3, 7, 11] | 512 | [[1, 3, 5], [1, 3, 5], [1, 3, 5]]

2 | "1" | [16,16,4,4] | [8, 8, 2, 2] | [3, 7, 11] | 128 | [[1, 3, 5], [1, 3, 5], [1, 3, 5]]

3 | "2" | [16,16,8] | [8,8,4] | [3,5,7] | 256 | [[1,2], [2,6], [3,12]]

########################################################################################################################################################### resblock_type: "1" resblock_dilation_sizes: [[1, 3, 5], [1, 3, 5], [1, 3, 5]] resblock_kernel_sizes: [3, 7, 11] upsample_kernel_sizes: [11, 8, 8, 4, 4] upsample_initial_channel: 512 upsample_factors: [5, 4, 4, 2, 2]

inference_padding: 5 cond_channels: 0 conv_post_bias: True

mel_spectogram: !name:speechbrain.lobes.models.HifiGAN.mel_spectogram sample_rate: !ref <sample_rate> hop_length: !ref <hop_length> win_length: !ref <win_length> n_fft: !ref <n_fft> n_mels: !ref <n_mel_channels> f_min: !ref <mel_fmin> f_max: !ref <mel_fmax> power: !ref <power> normalized: !ref <mel_normalized> norm: !ref <norm> mel_scale: !ref <mel_scale> compression: !ref <dynamic_range_compression>

generator: !new:speechbrain.lobes.models.HifiGAN.UnitHifiganGenerator in_channels: !ref <in_channels> out_channels: !ref <out_channels> resblock_type: !ref <resblock_type> resblock_dilation_sizes: !ref <resblock_dilation_sizes> resblock_kernel_sizes: !ref <resblock_kernel_sizes> upsample_kernel_sizes: !ref <upsample_kernel_sizes> upsample_initial_channel: !ref <upsample_initial_channel> upsample_factors: !ref <upsample_factors> inference_padding: !ref <inference_padding> cond_channels: !ref <cond_channels> conv_post_bias: !ref <conv_post_bias> vocab_size: !ref <vocab_size> embedding_dim: !ref <embedding_dim> duration_predictor: !ref <duration_predictor> var_pred_hidden_dim: !ref <var_pred_hidden_dim> var_pred_kernel_size: !ref <var_pred_kernel_size> var_pred_dropout: !ref <var_pred_dropout> multi_speaker: !ref <multi_speaker> normalize_speaker_embeddings: !ref <normalize_speaker_embeddings>

discriminator: !new:speechbrain.lobes.models.HifiGAN.HifiganDiscriminator

modules: generator: !ref <generator> discriminator: !ref <discriminator>

#generator loss stft_loss: null mseg_loss: !new:speechbrain.lobes.models.HifiGAN.MSEGLoss feat_match_loss: !new:speechbrain.lobes.models.HifiGAN.MelganFeatureLoss l1_spec_loss: !new:speechbrain.lobes.models.HifiGAN.L1SpecLoss sample_rate: !ref <sample_rate> hop_length: !ref <hop_length> win_length: !ref <win_length> n_mel_channels: !ref <n_mel_channels> n_fft: !ref <n_fft> n_stft: !ref <n_fft> // 2 + 1 mel_fmin: !ref <mel_fmin> mel_fmax: null mel_normalized: !ref <mel_normalized> power: !ref <power> dynamic_range_compression: !ref <dynamic_range_compression> mseg_dur_loss: False

generator_loss: !new:speechbrain.lobes.models.HifiGAN.GeneratorLoss stft_loss: !ref <stft_loss> stft_loss_weight: 0 mseg_loss: !ref <mseg_loss> mseg_loss_weight: 1 feat_match_loss: !ref <feat_match_loss> feat_match_loss_weight: 10 l1_spec_loss: !ref <l1_spec_loss> l1_spec_loss_weight: 45 mseg_dur_loss: !ref <mseg_dur_loss> mseg_dur_loss_weight: 1

#discriminator loss msed_loss: !new:speechbrain.lobes.models.HifiGAN.MSEDLoss

discriminator_loss: !new:speechbrain.lobes.models.HifiGAN.DiscriminatorLoss msed_loss: !ref <msed_loss>

#optimizer opt_class_generator: !name:torch.optim.AdamW lr: !ref <learning_rate> betas: [!ref <adam_b1>, !ref <adam_b2>]

opt_class_discriminator: !name:torch.optim.AdamW lr: !ref <learning_rate> betas: [!ref <adam_b1>, !ref <adam_b2>]

sch_class_generator: !name:torch.optim.lr_scheduler.ExponentialLR gamma: !ref <weight_decay> last_epoch: -1

sch_class_discriminator: !name:torch.optim.lr_scheduler.ExponentialLR gamma: !ref <weight_decay> last_epoch: -1

#epoch object epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter limit: !ref <epochs>

train_logger: !new:speechbrain.utils.train_logger.FileTrainLogger save_file: !ref <train_log>

#checkpointer checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer checkpoints_dir: !ref <save_folder> recoverables: generator: !ref <generator> discriminator: !ref <discriminator> counter: !ref <epoch_counter>