Batched inference of embeddings. `x` is a waveform of shape (bs, T)

critical down_factor LPF of `c` of shape (bs, channels, seq_len)

Downsampling: apply LPF and only keep every `down_factor` input `c` of shape (bs, channels, seq_len).

Logistic discriminator loss. `d_real_outputs` (bs,): D(x), or in our case D(c) `d_generated_outputs` (bs,): D(G(x)), or in our case D(W(z

Logistic generator loss. Assumes input is D(G(x)), or in our case, D(W(z)). `disc_outputs` of shape (bs,)

`c_real` and `c_fake` of shape (bs, dim)

Homebrew ADA accumulation

Numpy implementation of the Frechet Distance. The Frechet distance between two multivariate Gaussians X_1 ~ N(mu_1, C_1) and X_2 ~ N(mu_2, C_2

Numpy implementation of the Frechet Distance. The Frechet distance between two multivariate Gaussians X_1 ~ N(mu_1, C_1) and X_2 ~ N(mu_2, C_2

Assumes `z` of shape (bs, z_dim)

HiFiGAN which works on HuBERT embeddings. Optionally specify `model_dir` as location of hubert checkpoint.

Find % of `closest_n` that belong to `real` or `fake` (bs, c_dim) cs

`c_fake` (bs, seq_len, c_dim)

Reads in audio from `path` using soundfile. Optionally ensure audio is of length `ref_len` samples.

Upsampling: inserts zeros and convolves with LPF (interpolation filter) `c` of shape (bs, channels, seq_len)

`c` is of shape (bs, seq_len, c_dim)

Instantiate a hubert feature reader from checkpoint file at `checkpoint_path`, setup to extract features from `layer` of the model. Optionally

Inference wrapper for generator `g` and `hifigan` vocoder for the conv2 series of models which predict a sequence of hubert vectors, which i

sr=50 for hubert features, as they are every 20ms <-> 50Hz

M must be a power of 2.

Approximates a KL divergence metric between 1D distributions of values from real and fake c embeddings.

Calculate AM Score from `gen_probs` (N1, n_classes) and `train_probs` (N2, n_classes)

Density GAN which generates mel-spectrograms from standard normal vectors and then uses hifigan to vocode them back to the time domain.

Compute FID for features of generated utterances `c_f` (N, dim) and features from training set `c_train` (N, dim).

Recursively flattens a config into a flat dictionary compatible with tensorboard's `add_hparams` function.

`z` of shape (bs, z_dim) Returns `c` of shape (bs, seq_len, c_dim)

`x` of shape (bs, c_in, seq_len), `w` of shape (bs, w_dim)

Get Fourier features (bs, seq_len, channels) from a `w` vec (bs, w_dim)

Adapted from StyleGAN2 official implementation. `x` of shape (bs, channels, seq_len)

`c` of shape (bs, seq_len, c_dim)

`c` of shape (bs, c_dim, seq_len)

Generate waveforms (N, 16000) from latent standard normal `z` (N, z_dim)

Generate waveforms (N, 16000) from W latent space `w` (N, w_dim)

Converts audio in a given `file_path` to hubert features. Optionally directly provide `audio` as 1D float numpy array of correct sam

Facebook HuBERT BASE model

Calculate inception score from `logprobs_gen` (bs, n_classes) of log probabilities. Adapted from https://github.com/HazyResearch/state-

Calculate linear separability with SVM, StyleGAN1-style, given deep classifier class `logits` (N, n_classes) and all the latent vectors `latents`

Calculate modified inception score from `probs_gen` (bs, n_classes) of probabilities. Adapted from https://github.com/HazyResearch/stat

Calculates the number of statistically distinct bins (NDB) from features of real data `feat_data` (N1, dim) and features of generated da

Generate `N` audio samples, returning a tensor of shape (N, 16000)

Generate latent W vectors (N, w_dim) from latent standard normal `z` (N, z_dim)