Vocoder

A vocoder is a model that generates audio from a Mel spectrogram. HiFiGAN is a generative adversarial network (GAN) model that generates audio from Mel spectrograms. The generator uses transposed convolutions to upsample Mel spectrograms to audio

The following tasks have been implemented for HiFiGAN in the TAO Toolkit:

  • download_specs

  • dataset_convert

  • train

  • infer

  • export

The example specification files for all the tasks associated with the spectrogram generator component of TTS can be downloaded using the following command:

Copy
Copied!
            

tao vocoder download_specs \ -o <target_path> \ -r <results_path>

Required Arguments

  • -o: The target path where the spec files will be stored

  • -r: The results and output log directory

The vocoder for the TAO Toolkit implements the dataset_convert task to convert and prepare datasets that follow the LJSpeech dataset format.

The format and instruction to consume the data in TAO toolkit is identical to the dataset_convert task under Spectrogram Generator.

The spec file for TTS using HiFiGAN includes the trainer, model, training_dataset, validation_dataset, and prior_folder.

The following is a shortened example of a spec file for training HiFiGAN on the LJSpeech dataset.

Copy
Copied!
            

train_dataset: ??? validation_dataset: ??? training_ds: dataset: _target_: "nemo.collections.tts.data.datalayers.AudioDataset" manifest_filepath: ${train_dataset} max_duration: null min_duration: 0.75 n_segments: 8192 trim: false dataloader_params: drop_last: false shuffle: true batch_size: 16 num_workers: 4 validation_ds: dataset: _target_: "nemo.collections.tts.data.datalayers.AudioDataset" manifest_filepath: ${validation_dataset} max_duration: null min_duration: null n_segments: -1 trim: false dataloader_params: drop_last: false shuffle: false batch_size: 16 num_workers: 1 model: preprocessor: _target_: nemo.collections.asr.parts.preprocessing.features.FilterbankFeatures dither: 0.0 frame_splicing: 1 nfilt: 80 highfreq: 8000 log: true log_zero_guard_type: clamp log_zero_guard_value: 1e-05 lowfreq: 0 mag_power: 1.0 n_fft: 1024 n_window_size: 1024 n_window_stride: 256 normalize: null pad_to: 0 pad_value: -11.52 preemph: null sample_rate: 22050 window: hann use_grads: false exact_pad: true generator: _target_: nemo.collections.tts.modules.hifigan_modules.Generator resblock: 1 upsample_rates: [8,8,2,2] upsample_kernel_sizes: [16,16,4,4] upsample_initial_channel: 512 resblock_kernel_sizes: [3,7,11] resblock_dilation_sizes: [[1,3,5], [1,3,5], [1,3,5]] optim: _target_: torch.optim.AdamW lr: 0.0002 betas: [0.8, 0.99] sched: name: CosineAnnealing min_lr: 1e-5 warmup_ratio: 0.02 max_steps: ${trainer.max_steps} l1_loss_factor: 45 denoise_strength: 0.0025 trainer: max_steps: 25000

The specification can be roughly grouped into three categories:

  • Parameters to configure the trainer

  • Parameters that describe the model

  • Pointers to the training and validation dataset

This specification can be used with the tao vocoder train command.

If you would like to change a parameter for your run without changing the specification file itself, you can specify it on the command line directly. For example, if you would like to change the validation batch size, you can add validation_ds.dataloader_params.batch_size=1 to your command, which would override the batch size of 16 in the configuration shown above. An example of this is shown in the training instructions below.

Configuring the Trainer

The following parameter is used to configure the trainer element of the Spectrogram Generator:

Parameter

Datatype

Description

Supported Values

max_steps

int

The maximum number of steps to train the model. This is a field for the trainer parameter. Unlike the FastPitch spectrogram generator, the HiFiGAN trainer uses max_steps to specify the duration of training.

>0

Note

NVIDIA suggests setting trainer.max_steps = 10000 atleast, to train a good model.


Configuring the model

The parameters to help configure the HiFiGAN model are included in the model element. This includes global parameters for the model object and optional parameters to configure the following sub components:

  1. preprocessor

  2. generator

  3. optimizer

  4. scheduler

The global parameters include the following:

Parameter

Datatype

Description

Supported Values

max_steps

int

Specifies the maximum number of steps to train the model. This is a field for the trainer parameter. Unlike the FastPitch spectrogram generator, HiFiGAN trainer uses max_steps to run the duration of training.

Derived from trainer.max_steps

l1_loss_factor

float

The multiplicative factor for L1 loss used in training

denoise_strength

float

The small desnoising factor, currently only used in validation

Preprocessor

Parameter

Datatype

Description

Supported Values

dither

float

The amount of white-noise dithering

frame_splicing

int

The number of spectrogram frames per model step

highfreq

int

The upper bound on the Mel basis in Hz

log

bool

Specifies whether to log the spectrogram

log_zero_guard_type

The guard against taking the log of zero. There are two
options: add and clamp.

low_zero_guard_value

float/str

The guard types require a number to add with or
clamp to. The guard value can be a
float, “tiny”, or “eps”. torch.finfo is used if
“tiny” or “eps” is passed.

lowfreq

int

The lower bound on the Mel basis in Hz

mag_power

int

The power that the linear spectrogram is rasied to
prior to multiplication with the Mel basis

n_fft

int

The size of window for fft in samples.

n_window_size

int

The size of window for fft in samples.

n_window_stride

int

The stride of the window for fft.

normalize

Normalization can be ‘per_feature’ or ‘all_features’; all
other options disable feature normalization.
all_features normalizes the entire spectrogram
to be mean 0 with std 1.
pre_features normalizes per channel/freq instead.

pad_to

int

Ensures that the output size of the time dimension a mutliple
of pad_to.

pad_value

float

The value that the shorted Mels are padded with.

preemph

The amount of pre-emphasis to add to the audio. This can be
disabled by setting the value to None.

sample_rate

int

The target sample rate to load the audio, in kHz.

window

string

The windowing function for fft, which be one of the following:
‘hann’, ‘hamming’, ‘blackman’, ‘bartlett’.

use_grads

bool

Specifies whether to allow gradients to pass through this
module

exact_pad

bool

Specifies whether to pad the input signal such that the
output length is exactly the input length // 4

Generator

Parameter

Datatype

Description

Supported Values

resblock

int
int

The type of residual block to use. See the hifigan paper
for details.

1, 2

upsample_rates

array of 4 integers

How much each layer upsamples the input, the product of all
numbers in the array must be equal to n_window_stride.

upsample_kernel_sizes

array of 4 integers

The kernel size for each upsampling layer

upsample_initial_channel

int

The first hidden dimension of the layer

resblock_kernel_sizes

array of 3 integers

The kernel sizes of the residual blocks

resblock_dilation_sizes

array of 3 array of 3 integers

The dilation sizes of the residual blocks

Configure the dataset

The datasets that you use should be specified by <xyz>_ds parameters, depending on the use case:

  • For training using tao vocoder train, you should have training_ds to describe your training dataset, and validation_ds to describe your validation dataset.

Each <xyz>_ds config contains two main groups of configuration

  • dataset: The configuration component describing the dataset

  • dataloader: The configuration componenet describing the dataloader

The configurable fields for the dataset field are described in the following table:

Parameter

Datatype

Description

Supported Values

manifest_filepath

string

The filepath to the manifest (.json file) that describes the audio data

Valid filepaths.

n_segments

int

The length of the audio in sample to load. For example, given a sampling rate of 16kHz, and n_segments=16000, a random 1
second of audio from the clip will be loaded. The section will sample randomly every time the audio is batched. This
can be set to -1 to load the entire audio.

> 0

max_duration

float

If audio exceeds this length in seconds, it is filtered from the dataset.

min_duration

float

If the audio is less than this length in seconds, it is filtered from the dataset.

trim

bool

Whether to trim silence from the beginning and end of the audio signal using librosa.effects.trim(). The default
value is False.

True/False

Dataloader

Parameter

Datatype

Description

Supported Values

num_workers

integer

The number of worker threads for loading
the dataset

2

shuffle

bool

Whether to shuffle the data. We recommend True for training data and False for validation.

True/False

batch_size

integer

The training data batch size

drop_last

bool

Specifies whether to drop the last batch if there aren’t enough samples to fill the
batch.

True/False

To train a model from scratch, use the following command:

Copy
Copied!
            

tao vocoder train \ -e <experiment_spec> \ -g <num_gpus> \ -r /path/to/the/results/directory \ -k <encryption_key>

As mentioned above, you can add additional arguments to override configurations from your experiment specification file. This allows you to create valid spec files that leave these fields blank, to be specified as command line arguments at runtime.

For example, the following command can be used to override the training manifest and validation manifest, the number of epochs to train, and the place to save the model checkpoint:

Copy
Copied!
            

tao vocoder train \ -e $SPECS_DIR/vocoder/train.yaml \ -g 1 \ -k $KEY \ -r $RESULTS_DIR/vocoder/train \ train_dataset=$DATA_DIR/ljspeech/ljspeech_train.json \ validation_dataset=$DATA_DIR/ljspeech/ljspeech_val.json \ trainer.max_steps=10000

Required Arguments

  • -e: The experiment specification file to set up training, as in the example given above

  • -r: The path to the results and log directory. Log files, checkpoints, etc. will be stored here

  • -k: The key to encrypt the model

  • Other arguments to override fields in the specification file.

Optional Arguments

  • -g: The number of GPUs to be used in the training in a multi-GPU scenario. The default value is 1.

Training Procedure

At the start of each training experiment, TAO Toolkit will print out a log of the experiment specification, including any parameters added or overridden via the command line. It will also show additional information, such as which GPUs are available, where logs will be saved, how many hours are in each loaded dataset, and how much of each dataset has been filtered.

Copy
Copied!
            

GPU available: True, used: True TPU available: False, using: 0 TPU cores IPU available: False, using: 0 IPUs [NeMo W 2021-11-02 23:46:59 exp_manager:332] There was no checkpoint folder at checkpoint_dir :/results/vocoder/train_v2/checkpoints. Training from scratch. [NeMo I 2021-11-02 23:46:59 exp_manager:220] Experiments will be logged at /results/vocoder/train_v2 [NeMo W 2021-11-02 23:46:59 exp_manager:823] The checkpoint callback was told to monitor a validation value and trainer's max_steps was set to 10000. Please ensure that max_steps will run for at least 1 epochs to ensure that checkpointing will not error out. [NeMo W 2021-11-02 23:46:59 nemo_logging:349] /opt/conda/lib/python3.8/site-packages/pytorch_lightning/callbacks/model_checkpoint.py:240: LightningDeprecationWarning: `ModelCheckpoint(every_n_val_epochs)` is deprecated in v1.4 and will be removed in v1.6. Please use `every_n_epochs` instead. rank_zero_deprecation( [NeMo I 2021-11-02 23:46:59 features:252] PADDING: 0 [NeMo I 2021-11-02 23:46:59 features:269] STFT using torch [NeMo I 2021-11-02 23:46:59 features:271] STFT using exact pad [NeMo I 2021-11-02 23:46:59 features:252] PADDING: 0 [NeMo I 2021-11-02 23:46:59 features:269] STFT using torch [NeMo I 2021-11-02 23:46:59 features:271] STFT using exact pad [NeMo I 2021-11-02 23:47:01 collections:173] Dataset loaded with 12500 files totalling 22.84 hours [NeMo I 2021-11-02 23:47:01 collections:174] 0 files were filtered totalling 0.00 hours [NeMo I 2021-11-02 23:47:01 collections:173] Dataset loaded with 100 files totalling 0.18 hours [NeMo I 2021-11-02 23:47:01 collections:174] 0 files were filtered totalling 0.00 hours initializing ddp: GLOBAL_RANK: 0, MEMBER: 1/1 Added key: store_based_barrier_key:1 to store for rank: 0 Rank 0: Completed store-based barrier for 1 nodes. ---------------------------------------------------------------------------------------------------- distributed_backend=nccl All DDP processes registered. Starting ddp with 1 processes ----------------------------------------------------------------------------------------------------

You should next see a full printout of the number of parameters in each module and submodule, as well as the total number of trainable and non-trainable parameters in the model.

In the following table, the generator module contains 13.9 million parameters and its submodule generator.conv_pre contains 287,000 parameters. The audio_to_melspec_preprocess is listed with no parameters.

Copy
Copied!
            

| Name | Type | Params -------------------------------------------------------------------------------- 0 | audio_to_melspec_precessor | FilterbankFeatures | 0 1 | trg_melspec_fn | FilterbankFeatures | 0 2 | generator | Generator | 13.9 M 3 | generator.conv_pre | Conv1d | 287 K 4 | generator.ups | ModuleList | 2.7 M 5 | generator.ups.0 | ConvTranspose1d | 2.1 M 6 | generator.ups.1 | ConvTranspose1d | 524 K 7 | generator.ups.2 | ConvTranspose1d | 33.0 K 8 | generator.ups.3 | ConvTranspose1d | 8.3 K 9 | generator.resblocks | ModuleList | 11.0 M 10 | generator.resblocks.0 | ModuleList | 8.3 M 11 | generator.resblocks.0.0 | ResBlock1 | 1.2 M 12 | generator.resblocks.0.0.convs1 | ModuleList | 591 K 13 | generator.resblocks.0.0.convs1.0 | Conv1d | 197 K 14 | generator.resblocks.0.0.convs1.1 | Conv1d | 197 K 15 | generator.resblocks.0.0.convs1.2 | Conv1d | 197 K 16 | generator.resblocks.0.0.convs2 | ModuleList | 591 K 17 | generator.resblocks.0.0.convs2.0 | Conv1d | 197 K 18 | generator.resblocks.0.0.convs2.1 | Conv1d | 197 K 19 | generator.resblocks.0.0.convs2.2 | Conv1d | 197 K 20 | generator.resblocks.0.1 | ResBlock1 | 2.8 M 21 | generator.resblocks.0.1.convs1 | ModuleList | 1.4 M 22 | generator.resblocks.0.1.convs1.0 | Conv1d | 459 K 23 | generator.resblocks.0.1.convs1.1 | Conv1d | 459 K 24 | generator.resblocks.0.1.convs1.2 | Conv1d | 459 K 25 | generator.resblocks.0.1.convs2 | ModuleList | 1.4 M 26 | generator.resblocks.0.1.convs2.0 | Conv1d | 459 K 27 | generator.resblocks.0.1.convs2.1 | Conv1d | 459 K 28 | generator.resblocks.0.1.convs2.2 | Conv1d | 459 K 29 | generator.resblocks.0.2 | ResBlock1 | 4.3 M 30 | generator.resblocks.0.2.convs1 | ModuleList | 2.2 M 31 | generator.resblocks.0.2.convs1.0 | Conv1d | 721 K 32 | generator.resblocks.0.2.convs1.1 | Conv1d | 721 K 33 | generator.resblocks.0.2.convs1.2 | Conv1d | 721 K 34 | generator.resblocks.0.2.convs2 | ModuleList | 2.2 M 35 | generator.resblocks.0.2.convs2.0 | Conv1d | 721 K 36 | generator.resblocks.0.2.convs2.1 | Conv1d | 721 K 37 | generator.resblocks.0.2.convs2.2 | Conv1d | 721 K 38 | generator.resblocks.1 | ModuleList | 2.1 M 39 | generator.resblocks.1.0 | ResBlock1 | 296 K 40 | generator.resblocks.1.0.convs1 | ModuleList | 148 K 41 | generator.resblocks.1.0.convs1.0 | Conv1d | 49.4 K 42 | generator.resblocks.1.0.convs1.1 | Conv1d | 49.4 K 43 | generator.resblocks.1.0.convs1.2 | Conv1d | 49.4 K 44 | generator.resblocks.1.0.convs2 | ModuleList | 148 K 45 | generator.resblocks.1.0.convs2.0 | Conv1d | 49.4 K 46 | generator.resblocks.1.0.convs2.1 | Conv1d | 49.4 K 47 | generator.resblocks.1.0.convs2.2 | Conv1d | 49.4 K 48 | generator.resblocks.1.1 | ResBlock1 | 689 K 49 | generator.resblocks.1.1.convs1 | ModuleList | 344 K 50 | generator.resblocks.1.1.convs1.0 | Conv1d | 114 K 51 | generator.resblocks.1.1.convs1.1 | Conv1d | 114 K 52 | generator.resblocks.1.1.convs1.2 | Conv1d | 114 K 53 | generator.resblocks.1.1.convs2 | ModuleList | 344 K 54 | generator.resblocks.1.1.convs2.0 | Conv1d | 114 K 55 | generator.resblocks.1.1.convs2.1 | Conv1d | 114 K 56 | generator.resblocks.1.1.convs2.2 | Conv1d | 114 K 57 | generator.resblocks.1.2 | ResBlock1 | 1.1 M 58 | generator.resblocks.1.2.convs1 | ModuleList | 541 K 59 | generator.resblocks.1.2.convs1.0 | Conv1d | 180 K 60 | generator.resblocks.1.2.convs1.1 | Conv1d | 180 K 61 | generator.resblocks.1.2.convs1.2 | Conv1d | 180 K 62 | generator.resblocks.1.2.convs2 | ModuleList | 541 K 63 | generator.resblocks.1.2.convs2.0 | Conv1d | 180 K 64 | generator.resblocks.1.2.convs2.1 | Conv1d | 180 K 65 | generator.resblocks.1.2.convs2.2 | Conv1d | 180 K 66 | generator.resblocks.2 | ModuleList | 518 K 67 | generator.resblocks.2.0 | ResBlock1 | 74.5 K 68 | generator.resblocks.2.0.convs1 | ModuleList | 37.2 K 69 | generator.resblocks.2.0.convs1.0 | Conv1d | 12.4 K 70 | generator.resblocks.2.0.convs1.1 | Conv1d | 12.4 K 71 | generator.resblocks.2.0.convs1.2 | Conv1d | 12.4 K 72 | generator.resblocks.2.0.convs2 | ModuleList | 37.2 K 73 | generator.resblocks.2.0.convs2.0 | Conv1d | 12.4 K 74 | generator.resblocks.2.0.convs2.1 | Conv1d | 12.4 K 75 | generator.resblocks.2.0.convs2.2 | Conv1d | 12.4 K 76 | generator.resblocks.2.1 | ResBlock1 | 172 K 77 | generator.resblocks.2.1.convs1 | ModuleList | 86.4 K 78 | generator.resblocks.2.1.convs1.0 | Conv1d | 28.8 K 79 | generator.resblocks.2.1.convs1.1 | Conv1d | 28.8 K 80 | generator.resblocks.2.1.convs1.2 | Conv1d | 28.8 K 81 | generator.resblocks.2.1.convs2 | ModuleList | 86.4 K 82 | generator.resblocks.2.1.convs2.0 | Conv1d | 28.8 K 83 | generator.resblocks.2.1.convs2.1 | Conv1d | 28.8 K 84 | generator.resblocks.2.1.convs2.2 | Conv1d | 28.8 K 85 | generator.resblocks.2.2 | ResBlock1 | 271 K 86 | generator.resblocks.2.2.convs1 | ModuleList | 135 K 87 | generator.resblocks.2.2.convs1.0 | Conv1d | 45.2 K 88 | generator.resblocks.2.2.convs1.1 | Conv1d | 45.2 K 89 | generator.resblocks.2.2.convs1.2 | Conv1d | 45.2 K 90 | generator.resblocks.2.2.convs2 | ModuleList | 135 K 91 | generator.resblocks.2.2.convs2.0 | Conv1d | 45.2 K 92 | generator.resblocks.2.2.convs2.1 | Conv1d | 45.2 K 93 | generator.resblocks.2.2.convs2.2 | Conv1d | 45.2 K 94 | generator.resblocks.3 | ModuleList | 130 K 95 | generator.resblocks.3.0 | ResBlock1 | 18.8 K 96 | generator.resblocks.3.0.convs1 | ModuleList | 9.4 K 97 | generator.resblocks.3.0.convs1.0 | Conv1d | 3.1 K 98 | generator.resblocks.3.0.convs1.1 | Conv1d | 3.1 K 99 | generator.resblocks.3.0.convs1.2 | Conv1d | 3.1 K 100 | generator.resblocks.3.0.convs2 | ModuleList | 9.4 K 101 | generator.resblocks.3.0.convs2.0 | Conv1d | 3.1 K 102 | generator.resblocks.3.0.convs2.1 | Conv1d | 3.1 K 103 | generator.resblocks.3.0.convs2.2 | Conv1d | 3.1 K 104 | generator.resblocks.3.1 | ResBlock1 | 43.4 K 105 | generator.resblocks.3.1.convs1 | ModuleList | 21.7 K 106 | generator.resblocks.3.1.convs1.0 | Conv1d | 7.2 K 107 | generator.resblocks.3.1.convs1.1 | Conv1d | 7.2 K 108 | generator.resblocks.3.1.convs1.2 | Conv1d | 7.2 K 109 | generator.resblocks.3.1.convs2 | ModuleList | 21.7 K 110 | generator.resblocks.3.1.convs2.0 | Conv1d | 7.2 K 111 | generator.resblocks.3.1.convs2.1 | Conv1d | 7.2 K 112 | generator.resblocks.3.1.convs2.2 | Conv1d | 7.2 K 113 | generator.resblocks.3.2 | ResBlock1 | 68.0 K 114 | generator.resblocks.3.2.convs1 | ModuleList | 34.0 K 115 | generator.resblocks.3.2.convs1.0 | Conv1d | 11.3 K 116 | generator.resblocks.3.2.convs1.1 | Conv1d | 11.3 K 117 | generator.resblocks.3.2.convs1.2 | Conv1d | 11.3 K 118 | generator.resblocks.3.2.convs2 | ModuleList | 34.0 K 119 | generator.resblocks.3.2.convs2.0 | Conv1d | 11.3 K 120 | generator.resblocks.3.2.convs2.1 | Conv1d | 11.3 K 121 | generator.resblocks.3.2.convs2.2 | Conv1d | 11.3 K 122 | generator.conv_post | Conv1d | 226 123 | mpd | MultiPeriodDiscriminator | 41.1 M 124 | mpd.discriminators | ModuleList | 41.1 M 125 | mpd.discriminators.0 | DiscriminatorP | 8.2 M 126 | mpd.discriminators.0.convs | ModuleList | 8.2 M 127 | mpd.discriminators.0.convs.0 | Conv2d | 224 128 | mpd.discriminators.0.convs.1 | Conv2d | 20.7 K 129 | mpd.discriminators.0.convs.2 | Conv2d | 328 K 130 | mpd.discriminators.0.convs.3 | Conv2d | 2.6 M 131 | mpd.discriminators.0.convs.4 | Conv2d | 5.2 M 132 | mpd.discriminators.0.conv_post | Conv2d | 3.1 K 133 | mpd.discriminators.1 | DiscriminatorP | 8.2 M 134 | mpd.discriminators.1.convs | ModuleList | 8.2 M 135 | mpd.discriminators.1.convs.0 | Conv2d | 224 136 | mpd.discriminators.1.convs.1 | Conv2d | 20.7 K 137 | mpd.discriminators.1.convs.2 | Conv2d | 328 K 138 | mpd.discriminators.1.convs.3 | Conv2d | 2.6 M 139 | mpd.discriminators.1.convs.4 | Conv2d | 5.2 M 140 | mpd.discriminators.1.conv_post | Conv2d | 3.1 K 141 | mpd.discriminators.2 | DiscriminatorP | 8.2 M 142 | mpd.discriminators.2.convs | ModuleList | 8.2 M 143 | mpd.discriminators.2.convs.0 | Conv2d | 224 144 | mpd.discriminators.2.convs.1 | Conv2d | 20.7 K 145 | mpd.discriminators.2.convs.2 | Conv2d | 328 K 146 | mpd.discriminators.2.convs.3 | Conv2d | 2.6 M 147 | mpd.discriminators.2.convs.4 | Conv2d | 5.2 M 148 | mpd.discriminators.2.conv_post | Conv2d | 3.1 K 149 | mpd.discriminators.3 | DiscriminatorP | 8.2 M 150 | mpd.discriminators.3.convs | ModuleList | 8.2 M 151 | mpd.discriminators.3.convs.0 | Conv2d | 224 152 | mpd.discriminators.3.convs.1 | Conv2d | 20.7 K 153 | mpd.discriminators.3.convs.2 | Conv2d | 328 K 154 | mpd.discriminators.3.convs.3 | Conv2d | 2.6 M 155 | mpd.discriminators.3.convs.4 | Conv2d | 5.2 M 156 | mpd.discriminators.3.conv_post | Conv2d | 3.1 K 157 | mpd.discriminators.4 | DiscriminatorP | 8.2 M 158 | mpd.discriminators.4.convs | ModuleList | 8.2 M 159 | mpd.discriminators.4.convs.0 | Conv2d | 224 160 | mpd.discriminators.4.convs.1 | Conv2d | 20.7 K 161 | mpd.discriminators.4.convs.2 | Conv2d | 328 K 162 | mpd.discriminators.4.convs.3 | Conv2d | 2.6 M 163 | mpd.discriminators.4.convs.4 | Conv2d | 5.2 M 164 | mpd.discriminators.4.conv_post | Conv2d | 3.1 K 165 | msd | MultiScaleDiscriminator | 29.6 M 166 | msd.discriminators | ModuleList | 29.6 M 167 | msd.discriminators.0 | DiscriminatorS | 9.9 M 168 | msd.discriminators.0.convs | ModuleList | 9.9 M 169 | msd.discriminators.0.convs.0 | Conv1d | 2.0 K 170 | msd.discriminators.0.convs.1 | Conv1d | 168 K 171 | msd.discriminators.0.convs.2 | Conv1d | 84.2 K 172 | msd.discriminators.0.convs.3 | Conv1d | 336 K 173 | msd.discriminators.0.convs.4 | Conv1d | 1.3 M 174 | msd.discriminators.0.convs.5 | Conv1d | 2.7 M 175 | msd.discriminators.0.convs.6 | Conv1d | 5.2 M 176 | msd.discriminators.0.conv_post | Conv1d | 3.1 K 177 | msd.discriminators.1 | DiscriminatorS | 9.9 M 178 | msd.discriminators.1.convs | ModuleList | 9.9 M 179 | msd.discriminators.1.convs.0 | Conv1d | 2.2 K 180 | msd.discriminators.1.convs.1 | Conv1d | 168 K 181 | msd.discriminators.1.convs.2 | Conv1d | 84.5 K 182 | msd.discriminators.1.convs.3 | Conv1d | 336 K 183 | msd.discriminators.1.convs.4 | Conv1d | 1.3 M 184 | msd.discriminators.1.convs.5 | Conv1d | 2.7 M 185 | msd.discriminators.1.convs.6 | Conv1d | 5.2 M 186 | msd.discriminators.1.conv_post | Conv1d | 3.1 K 187 | msd.discriminators.2 | DiscriminatorS | 9.9 M 188 | msd.discriminators.2.convs | ModuleList | 9.9 M 189 | msd.discriminators.2.convs.0 | Conv1d | 2.2 K 190 | msd.discriminators.2.convs.1 | Conv1d | 168 K 191 | msd.discriminators.2.convs.2 | Conv1d | 84.5 K 192 | msd.discriminators.2.convs.3 | Conv1d | 336 K 193 | msd.discriminators.2.convs.4 | Conv1d | 1.3 M 194 | msd.discriminators.2.convs.5 | Conv1d | 2.7 M 195 | msd.discriminators.2.convs.6 | Conv1d | 5.2 M 196 | msd.discriminators.2.conv_post | Conv1d | 3.1 K 197 | msd.meanpools | ModuleList | 0 198 | msd.meanpools.0 | AvgPool1d | 0 199 | msd.meanpools.1 | AvgPool1d | 0 200 | feature_loss | FeatureMatchingLoss | 0 201 | discriminator_loss | DiscriminatorLoss | 0 202 | generator_loss | GeneratorLoss | 0 -------------------------------------------------------------------------------- 84.7 M Trainable params 0 Non-trainable params 84.7 M Total params 338.643 Total estimated model params size (MB)

As the model starts training, you should see a progress bar per epoch.

Copy
Copied!
            

Epoch 0: 4%|▋ | 35/789 [00:37<13:05, 1.04s/it, g_l1_loss=1.240]

At the end of training, TAO Toolkit will save the last checkpoint at the path specified by the experiment spec file before finishing.

Copy
Copied!
            

[NeMo I 2021-01-20 22:38:48 train:120] Experiment logs saved to '$RESULTS_DIR/vocoder/train' [NeMo I 2021-01-20 22:38:48 train:123] Trained model saved to '$RESULTS_DIR/vocoder/train/checkpoints/trained-model.tlt' INFO: Internal process exited


Current Limitations

  • Currently, only .wav audio files are supported.

  • The vocoder can only be trained from scratch.

To perform inference on individual text lines, use the following command:

Copy
Copied!
            

tao vocoder infer -e <experiment_spec> \ -m <model_checkpoint.tlt> \ -g <num_gpus> \ -k $KEY \ -r </path/to/results/directory/for/logs> \ output_path=</path/to/result/directory/for/vocoder/inference> \ input_path=</path/to/the/input/spectrogram/from/spectro_gen/infer

Required Arguments

  • -e: The experiment specification file to set up inference. This spec file only needs a file_paths parameter that contains a list of individual file paths.

  • -m: The path to the model checkpoint, which should be a .tlt file.

  • -k: The key to encrypt the model

Optional Arguments

  • -g: The number of GPUs to use for inference in a multi-GPU scenario. The default value is 1.

  • -r: The path to the results and log directory. Log files, checkpoints, etc. will be stored here.

  • Other arguments to override fields in the specification file.

Inference Procedure

At the start of inference, TAO Toolkit will print out the experiment specification, including the audio filepaths on which inference will be performed.

When restoring from the checkpoint, it will then log the original datasets that the checkpoint model was trained and evaluated on. This will show the vocabulary that the model was trained on.

Copy
Copied!
            

[NeMo W 2021-11-02 23:53:51 nemo_logging:349] /opt/conda/lib/python3.8/site-packages/torchaudio-0.7.0a0+42d447d-py3.8-linux-x86_64.egg/torchaudio/backend/utils.py:53: UserWarning: "sox" backend is being deprecated. The default backend will be changed to "sox_io" backend in 0.8.0 and "sox" backend will be removed in 0.9.0. Please migrate to "sox_io" backend. Please refer to https://github.com/pytorch/audio/issues/903 for the detail. warnings.warn( [NeMo W 2021-11-02 23:53:51 experimental:27] Module <class 'nemo.collections.asr.data.audio_to_text_dali._AudioTextDALIDataset'> is experimental, not ready for production and is not fully supported. Use at your own risk. [NeMo W 2021-11-02 23:53:54 nemo_logging:349] /opt/conda/lib/python3.8/site-packages/torchaudio-0.7.0a0+42d447d-py3.8-linux-x86_64.egg/torchaudio/backend/utils.py:53: UserWarning: "sox" backend is being deprecated. The default backend will be changed to "sox_io" backend in 0.8.0 and "sox" backend will be removed in 0.9.0. Please migrate to "sox_io" backend. Please refer to https://github.com/pytorch/audio/issues/903 for the detail. warnings.warn( [NeMo W 2021-11-02 23:53:55 experimental:27] Module <class 'nemo.collections.asr.data.audio_to_text_dali._AudioTextDALIDataset'> is experimental, not ready for production and is not fully supported. Use at your own risk. [NeMo W 2021-11-02 23:53:55 nemo_logging:349] /home/jenkins/agent/workspace/tlt-pytorch-main-nightly/tts/vocoder/scripts/infer.py:90: UserWarning: 'infer.yaml' is validated against ConfigStore schema with the same name. This behavior is deprecated in Hydra 1.1 and will be removed in Hydra 1.2. See https://hydra.cc/docs/next/upgrades/1.0_to_1.1/automatic_schema_matching for migration instructions. [NeMo I 2021-11-02 23:53:55 tlt_logging:20] Experiment configuration: restore_from: /results/vocoder/train/checkpoints/trained-model.tlt exp_manager: task_name: infer explicit_log_dir: /results/vocoder/infer input_path: /results/spectro_gen/infer/spectro output_path: /results/vocoder/infer/wav sample_rate: 22050 encryption_key: '***' [NeMo W 2021-11-02 23:53:55 exp_manager:26] Exp_manager is logging to `/results/vocoder/infer``, but it already exists. [NeMo I 2021-11-02 23:54:06 features:252] PADDING: 0 [NeMo I 2021-11-02 23:54:06 features:269] STFT using torch [NeMo I 2021-11-02 23:54:06 features:271] STFT using exact pad [NeMo I 2021-11-02 23:54:06 features:252] PADDING: 0 [NeMo I 2021-11-02 23:54:06 features:269] STFT using torch [NeMo I 2021-11-02 23:54:06 features:271] STFT using exact pad [NeMo I 2021-11-02 23:54:13 infer:73] The prediction results: [NeMo I 2021-11-02 23:54:14 infer:83] Predicted audio: /results/vocoder/infer/wav/0.wav [NeMo I 2021-11-02 23:54:15 infer:83] Predicted audio: /results/vocoder/infer/wav/1.wav [NeMo I 2021-11-02 23:54:15 infer:83] Predicted audio: /results/vocoder/infer/wav/2.wav [NeMo I 2021-11-02 23:54:15 infer:86] Experiment logs saved to '/results/vocoder/infer' 2021-11-02 16:54:17,240 [INFO] tlt.components.docker_handler.docker_handler: Stopping container.

The path to the Mel spectrograms generated by the infer task is shown in the last lines of the log.

Current Limitations

  • Currently, only .wav audio files are generated.

To fine-tune a model from a checkpoint, use the following command:

Copy
Copied!
            

!tao vocoder finetune \ -e <experiment_spec> \ -g <num_gpus> \ -m <model_checkpoint> \ train_dataset=<train.json> \ validation_dataset=<val.json> \ trainer.max_steps=1000

Required Arguments

  • -e: The experiment specification file to set up fine-tuning.

  • -m: Path to the model checkpoint from which to fine-tune. Should be a .tlt file.

  • train_dataset: The path to the training manifest. Please see the section below on

    finetuning data.

  • validation_dataset: The path to the validation manifest.

  • trainer.max_steps: Number of steps used to finetune the model. We recommend adding 500

    for each minute in the finetuning data.

Optional Arguments

  • -g: The number of GPUs to be used for fine-tuning in a multi-GPU scenario (default: 1).

  • -r: The path to the results and log directory. Log files, checkpoints, etc., will be stored here.

  • Other arguments to override fields in the specification file.

Finetuning Dataset

For best results if using a FastPitch and HiFiGAN combination, finetuning HiFiGAN should be done on the outputs from a finetuned FastPitch model. In order to do this, you must have a finetuned FastPitch model, do inference with the FastPitch model, and update the training .json to have a mel_filepath: attribute for each .wav file.

Let’s do inference with FastPitch first.

Copy
Copied!
            

!tao spectro_gen infer \ -e <experiment_spec> \ -g <num_gpus> \ -m <model_checkpoint> \ output_path=<An empty directory where the specs are saved> \ speaker=1 \ mode="infer_hifigan_ft" \ input_json=<train.json>

The important arguments are:

  • output_path: The directory where the spectrograms are saved

  • input_json: The .json file that contains the finetuning data

  • mode="infer_hifigan_ft": Must be specified

  • speaker: The FastPitch speaker id. Should be 1 in most cases.

After this is done running, inside the output_path: there should be files such as 1.npy, 2.npy, … and so on. For each line inside of input_json:, please add a mel_filepath: attribute that corresponds to the saved spectrograms. For example, line 1 in input_json: should have "mel_filepath": "<PATH_TO_OUTPUT_PATH>/1.npy":.

Now you can run hifigan finetuning using your updated input_json: as train_dataset:.

You can export a trained HiFiGAN model to Riva format, which contains all the model artifacts necessary for deployment to Riva Services. For more details about Riva, see this page.

To export a HiFiGAN model to the Riva format, use the following command:

Copy
Copied!
            

tao vocoder export -e <experiment_spec> \ -m <model_checkpoint> \ -r <results_dir> \ -k <encryption_key> \ export_format=RIVA \ export_to=<filename.riva>

Required Arguments

  • -e: The experiment specification file for export. See the Export Spec File section below for more details.

  • -m: The path to the model checkpoint to be exported, which should be a .tlt file

  • -k: The encryption key

Optional Arguments

  • -r: The path to the directory where results will be stored.

Export Spec File

The following is an example spec file for model export:

Copy
Copied!
            

# Name of the .tlt EFF archive to be loaded/model to be exported. restore_from: trained-model.tlt # Set export format: RIVA export_format: RIVA # Output EFF archive containing model checkpoint and artifacts required for Riva Services export_to: exported-model.riva

Parameter

Datatype

Description

Default

restore_from

string

The path to the pre-trained model to be exported

trained_model.tlt

export_format

string

The export format

N/A

export_to

string

The target path for the export model

exported-model.riva

A successful run of the model export generates the following log:

Copy
Copied!
            

[NeMo W 2021-11-02 23:56:28 nemo_logging:349] /opt/conda/lib/python3.8/site-packages/torchaudio-0.7.0a0+42d447d-py3.8-linux-x86_64.egg/torchaudio/backend/utils.py:53: UserWarning: "sox" backend is being deprecated. The default backend will be changed to "sox_io" backend in 0.8.0 and "sox" backend will be removed in 0.9.0. Please migrate to "sox_io" backend. Please refer to https://github.com/pytorch/audio/issues/903 for the detail. warnings.warn( [NeMo W 2021-11-02 23:56:29 experimental:27] Module <class 'nemo.collections.asr.data.audio_to_text_dali._AudioTextDALIDataset'> is experimental, not ready for production and is not fully supported. Use at your own risk. [NeMo W 2021-11-02 23:56:32 nemo_logging:349] /opt/conda/lib/python3.8/site-packages/torchaudio-0.7.0a0+42d447d-py3.8-linux-x86_64.egg/torchaudio/backend/utils.py:53: UserWarning: "sox" backend is being deprecated. The default backend will be changed to "sox_io" backend in 0.8.0 and "sox" backend will be removed in 0.9.0. Please migrate to "sox_io" backend. Please refer to https://github.com/pytorch/audio/issues/903 for the detail. warnings.warn( [NeMo W 2021-11-02 23:56:32 experimental:27] Module <class 'nemo.collections.asr.data.audio_to_text_dali._AudioTextDALIDataset'> is experimental, not ready for production and is not fully supported. Use at your own risk. [NeMo W 2021-11-02 23:56:33 nemo_logging:349] /home/jenkins/agent/workspace/tlt-pytorch-main-nightly/tts/vocoder/scripts/export.py:85: UserWarning: 'export.yaml' is validated against ConfigStore schema with the same name. This behavior is deprecated in Hydra 1.1 and will be removed in Hydra 1.2. See https://hydra.cc/docs/next/upgrades/1.0_to_1.1/automatic_schema_matching for migration instructions. [NeMo I 2021-11-02 23:56:33 tlt_logging:20] Experiment configuration: restore_from: /results/vocoder/train/checkpoints/trained-model.tlt export_to: vocoder.riva export_format: RIVA exp_manager: task_name: export explicit_log_dir: /results/vocoder/export encryption_key: '**********' [NeMo W 2021-11-02 23:56:33 exp_manager:26] Exp_manager is logging to `/results/vocoder/export``, but it already exists. [NeMo I 2021-11-02 23:56:43 features:252] PADDING: 0 [NeMo I 2021-11-02 23:56:43 features:269] STFT using torch [NeMo I 2021-11-02 23:56:43 features:271] STFT using exact pad [NeMo I 2021-11-02 23:56:43 features:252] PADDING: 0 [NeMo I 2021-11-02 23:56:43 features:269] STFT using torch [NeMo I 2021-11-02 23:56:43 features:271] STFT using exact pad [NeMo I 2021-11-02 23:56:50 export:57] Model restored from '/results/vocoder/train/checkpoints/trained-model.tlt' Removing weight norm... [NeMo I 2021-11-02 23:57:03 export:72] Experiment logs saved to '/results/vocoder/export' [NeMo I 2021-11-02 23:57:03 export:73] Exported model to '/results/vocoder/export/vocoder.riva' [NeMo I 2021-11-02 23:57:04 export:80] Exported model is compliant with Riva


© Copyright 2022, NVIDIA. Last updated on Jun 6, 2022.