Hubert-discrete symbol-based HiFiGAN with duration predictor (#388)

Co-authored-by: Tomoki Hayashi <hayashi.tomoki@g.sp.m.is.nagoya-u.ac.jp>

README.md

@@ -463,7 +463,8 @@ $ parallel-wavegan-decode \
 
 ## Notes
 
-The terms of use of the pretrained model follow that of each corpus used for the training. Please carefully check by yourself.  
+- The terms of use of the pretrained model follow that of each corpus used for the training. Please carefully check by yourself.  
+- Some codes are derived from ESPnet or Kaldi, which are based on Apache-2.0 licenese.
 
 ## References
 

egs/cvss_c/hubert_voc1/cmd.sh

@@ -0,0 +1,91 @@
+# ====== About run.pl, queue.pl, slurm.pl, and ssh.pl ======
+# Usage: <cmd>.pl [options] JOB=1:<nj> <log> <command...>
+# e.g.
+#   run.pl --mem 4G JOB=1:10 echo.JOB.log echo JOB
+#
+# Options:
+#   --time <time>: Limit the maximum time to execute.
+#   --mem <mem>: Limit the maximum memory usage.
+#   -–max-jobs-run <njob>: Limit the number parallel jobs. This is ignored for non-array jobs.
+#   --num-threads <ngpu>: Specify the number of CPU core.
+#   --gpu <ngpu>: Specify the number of GPU devices.
+#   --config: Change the configuration file from default.
+#
+# "JOB=1:10" is used for "array jobs" and it can control the number of parallel jobs.
+# The left string of "=", i.e. "JOB", is replaced by <N>(Nth job) in the command and the log file name,
+# e.g. "echo JOB" is changed to "echo 3" for the 3rd job and "echo 8" for 8th job respectively.
+# Note that the number must start with a positive number, so you can't use "JOB=0:10" for example.
+#
+# run.pl, queue.pl, slurm.pl, and ssh.pl have unified interface, not depending on its backend.
+# These options are mapping to specific options for each backend and
+# it is configured by "conf/queue.conf" and "conf/slurm.conf" by default.
+# If jobs failed, your configuration might be wrong for your environment.
+#
+#
+# The official documentaion for run.pl, queue.pl, slurm.pl, and ssh.pl:
+#   "Parallelization in Kaldi": http://kaldi-asr.org/doc/queue.html
+# =========================================================~
+
+
+# Select the backend used by run.sh from "local", "stdout", "sge", "slurm", or "ssh"
+cmd_backend="local"
+
+# Local machine, without any Job scheduling system
+if [ "${cmd_backend}" = local ]; then
+
+    # The other usage
+    export train_cmd="utils/run.pl"
+    # Used for "*_train.py": "--gpu" is appended optionally by run.sh
+    export cuda_cmd="utils/run.pl"
+    # Used for "*_recog.py"
+    export decode_cmd="utils/run.pl"
+
+# Local machine, without any Job scheduling system
+elif [ "${cmd_backend}" = stdout ]; then
+
+    # The other usage
+    export train_cmd="utils/stdout.pl"
+    # Used for "*_train.py": "--gpu" is appended optionally by run.sh
+    export cuda_cmd="utils/stdout.pl"
+    # Used for "*_recog.py"
+    export decode_cmd="utils/stdout.pl"
+
+# "qsub" (SGE, Torque, PBS, etc.)
+elif [ "${cmd_backend}" = sge ]; then
+    # The default setting is written in conf/queue.conf.
+    # You must change "-q g.q" for the "queue" for your environment.
+    # To know the "queue" names, type "qhost -q"
+    # Note that to use "--gpu *", you have to setup "complex_value" for the system scheduler.
+
+    export train_cmd="utils/queue.pl"
+    export cuda_cmd="utils/queue.pl"
+    export decode_cmd="utils/queue.pl"
+
+# "sbatch" (Slurm)
+elif [ "${cmd_backend}" = slurm ]; then
+    # The default setting is written in conf/slurm.conf.
+    # You must change "-p cpu" and "-p gpu" for the "partion" for your environment.
+    # To know the "partion" names, type "sinfo".
+    # You can use "--gpu * " by defualt for slurm and it is interpreted as "--gres gpu:*"
+    # The devices are allocated exclusively using "${CUDA_VISIBLE_DEVICES}".
+
+    export train_cmd="utils/slurm.pl"
+    export cuda_cmd="utils/slurm.pl"
+    export decode_cmd="utils/slurm.pl"
+
+elif [ "${cmd_backend}" = ssh ]; then
+    # You have to create ".queue/machines" to specify the host to execute jobs.
+    # e.g. .queue/machines
+    #   host1
+    #   host2
+    #   host3
+    # Assuming you can login them without any password, i.e. You have to set ssh keys.
+
+    export train_cmd="utils/ssh.pl"
+    export cuda_cmd="utils/ssh.pl"
+    export decode_cmd="utils/ssh.pl"
+
+else
+    echo "$0: Error: Unknown cmd_backend=${cmd_backend}" 1>&2
+    return 1
+fi

egs/cvss_c/hubert_voc1/conf/hifigan_hubert_duration.v1.yaml

@@ -0,0 +1,183 @@
+# This configuration is based on HiFiGAN V1, derived
+# from official repository (https://github.com/jik876/hifi-gan).
+
+###########################################################
+#                FEATURE EXTRACTION SETTING               #
+###########################################################
+sampling_rate: 16000     # Sampling rate.
+fft_size: null           # FFT size.
+hop_size: 320            # Hop size.
+win_length: null         # Window length.
+                         # If set to null, it will be the same as fft_size.
+window: null             # Window function.
+num_mels: 1              # Number of mel basis.
+fmin: null               # Minimum freq in mel basis calculation.
+fmax: null               # Maximum frequency in mel basis calculation.
+global_gain_scale: 1.0   # Will be multiplied to all of waveform.
+trim_silence: false      # Whether to trim the start and end of silence.
+trim_threshold_in_db: 20 # Need to tune carefully if the recording is not good.
+trim_frame_size: 1024    # Frame size in trimming.
+trim_hop_size: 256       # Hop size in trimming.
+format: "hdf5"           # Feature file format. "npy" or "hdf5" is supported.
+
+###########################################################
+#         GENERATOR NETWORK ARCHITECTURE SETTING          #
+###########################################################
+generator_type: DiscreteSymbolDurationGenerator
+generator_params:
+    in_channels: 512                      # Number of input channels.
+    out_channels: 1                       # Number of output channels.
+    channels: 512                         # Number of initial channels.
+    num_embs: 500
+    kernel_size: 7                        # Kernel size of initial and final conv layers.
+    upsample_scales: [10, 8, 2, 2]        # Upsampling scales.
+    upsample_kernal_sizes: [20, 16, 4, 4] # Kernel size for upsampling layers.
+    resblock_kernel_sizes: [3, 7, 11]     # Kernel size for residual blocks.
+    resblock_dilations:                   # Dilations for residual blocks.
+        - [1, 3, 5]
+        - [1, 3, 5]
+        - [1, 3, 5]
+    use_additional_convs: true            # Whether to use additional conv layer in residual blocks.
+    bias: true                            # Whether to use bias parameter in conv.
+    nonlinear_activation: "LeakyReLU"     # Nonlinear activation type.
+    nonlinear_activation_params:          # Nonlinear activation paramters.
+        negative_slope: 0.1
+    use_weight_norm: true                 # Whether to apply weight normalization.
+    duration_layers: 2                    # Duration predictor layers
+    duration_chans: 384                   # Duration predictor conv channels
+    duration_kernel_size: 3               # Duration predictor kernel size
+    duration_offset: 1.0                  # Duration predictor offset 
+    duration_dropout_rate: 0.5            # Duration predictor dropout
+    num_spk_embs: 0                       # Do not consider speaker embedding for single spk
+
+###########################################################
+#       DISCRIMINATOR NETWORK ARCHITECTURE SETTING        #
+###########################################################
+discriminator_type: HiFiGANMultiScaleMultiPeriodDiscriminator
+discriminator_params:
+    scales: 3                              # Number of multi-scale discriminator.
+    scale_downsample_pooling: "AvgPool1d"  # Pooling operation for scale discriminator.
+    scale_downsample_pooling_params:
+        kernel_size: 4                     # Pooling kernel size.
+        stride: 2                          # Pooling stride.
+        padding: 2                         # Padding size.
+    scale_discriminator_params:
+        in_channels: 1                     # Number of input channels.
+        out_channels: 1                    # Number of output channels.
+        kernel_sizes: [15, 41, 5, 3]       # List of kernal sizes.
+        channels: 128                      # Initial number of channels.
+        max_downsample_channels: 1024      # Maximum number of channels in downsampling conv layers.
+        max_groups: 16                     # Maximum number of groups in downsampling conv layers.
+        bias: true
+        downsample_scales: [4, 4, 4, 4, 1] # Downsampling scales.
+        nonlinear_activation: "LeakyReLU"  # Nonlinear activation.
+        nonlinear_activation_params:
+            negative_slope: 0.1
+    follow_official_norm: true             # Whether to follow the official norm setting.
+    periods: [2, 3, 5, 7, 11]              # List of period for multi-period discriminator.
+    period_discriminator_params:
+        in_channels: 1                     # Number of input channels.
+        out_channels: 1                    # Number of output channels.
+        kernel_sizes: [5, 3]               # List of kernal sizes.
+        channels: 32                       # Initial number of channels.
+        downsample_scales: [3, 3, 3, 3, 1] # Downsampling scales.
+        max_downsample_channels: 1024      # Maximum number of channels in downsampling conv layers.
+        bias: true                         # Whether to use bias parameter in conv layer."
+        nonlinear_activation: "LeakyReLU"  # Nonlinear activation.
+        nonlinear_activation_params:       # Nonlinear activation paramters.
+            negative_slope: 0.1
+        use_weight_norm: true              # Whether to apply weight normalization.
+        use_spectral_norm: false           # Whether to apply spectral normalization.
+
+###########################################################
+#                   STFT LOSS SETTING                     #
+###########################################################
+use_stft_loss: false # Whether to use multi-resolution STFT loss.
+use_duration_loss: true # Whether to use duration prediction loss (need for duration predictor)
+duration_loss_params:
+    offset: 1.0
+    reduction: mean
+use_mel_loss: true   # Whether to use Mel-spectrogram loss.
+mel_loss_params:     # Mel-spectrogram loss parameters.
+    fs: 16000
+    fft_size: 1024
+    hop_size: 256
+    win_length: null
+    window: "hann"
+    num_mels: 80
+    fmin: 0
+    fmax: 8000
+    log_base: null   # Log base. If set to null, use natural logarithm.
+generator_adv_loss_params:
+    average_by_discriminators: false # Whether to average loss by #discriminators.
+discriminator_adv_loss_params:
+    average_by_discriminators: false # Whether to average loss by #discriminators.
+use_feat_match_loss: true
+feat_match_loss_params:
+    average_by_discriminators: false # Whether to average loss by #discriminators.
+    average_by_layers: false         # Whether to average loss by #layers in each discriminator.
+    include_final_outputs: true      # Whether to include final outputs in feat match loss calculation.
+
+###########################################################
+#               ADVERSARIAL LOSS SETTING                  #
+###########################################################
+lambda_aux: 45.0       # Loss balancing coefficient for STFT loss.
+lambda_adv: 1.0        # Loss balancing coefficient for adversarial loss.
+lambda_feat_match: 2.0 # Loss balancing coefficient for feat match loss..
+
+###########################################################
+#                  DATA LOADER SETTING                    #
+###########################################################
+batch_size: 16              # Batch size.
+batch_max_steps: 10240      # Length of each audio in batch. Make sure dividable by hop_size.
+pin_memory: true            # Whether to pin memory in Pytorch DataLoader.
+num_workers: 0              # Number of workers in Pytorch DataLoader.
+remove_short_samples: false # Whether to remove samples the length of which are less than batch_max_steps.
+allow_cache: true           # Whether to allow cache in dataset. If true, it requires cpu memory.
+
+###########################################################
+#             OPTIMIZER & SCHEDULER SETTING               #
+###########################################################
+generator_optimizer_type: Adam
+generator_optimizer_params:
+    lr: 2.0e-4
+    betas: [0.5, 0.9]
+    weight_decay: 0.0
+generator_scheduler_type: MultiStepLR
+generator_scheduler_params:
+    gamma: 0.5
+    milestones:
+        - 200000
+        - 400000
+        - 600000
+        - 800000
+generator_grad_norm: -1
+discriminator_optimizer_type: Adam
+discriminator_optimizer_params:
+    lr: 2.0e-4
+    betas: [0.5, 0.9]
+    weight_decay: 0.0
+discriminator_scheduler_type: MultiStepLR
+discriminator_scheduler_params:
+    gamma: 0.5
+    milestones:
+        - 200000
+        - 400000
+        - 600000
+        - 800000
+discriminator_grad_norm: -1
+
+###########################################################
+#                    INTERVAL SETTING                     #
+###########################################################
+generator_train_start_steps: 1     # Number of steps to start to train discriminator.
+discriminator_train_start_steps: 0 # Number of steps to start to train discriminator.
+train_max_steps: 2500000           # Number of training steps.
+save_interval_steps: 50000         # Interval steps to save checkpoint.
+eval_interval_steps: 1000          # Interval steps to evaluate the network.
+log_interval_steps: 100            # Interval steps to record the training log.
+
+###########################################################
+#                     OTHER SETTING                       #
+###########################################################
+num_save_intermediate_results: 4  # Number of results to be saved as intermediate results.