Function: save the path of .wavs in the root folder to index.txt Params: root: root folder

Return: namespace Params: dict

Return: IS loss Params: output: magnitude spectrogram (>=0), output of RVAE decoder, shape of [bs,T,F] target

Return: KL loss Params: zmean: mean of latent variables, output of RVAE encoder, shape of [bs,T,D] zlogvar: l

Return: convolutional layer sequence Params: config: list of parameters, [input channel, output channel, kernel size, stride, padding

Return: the scale of KL loss Params: cur_step: current training step beta: base scale (default: 1)

Return: GRU module Params: dim_in: input layer dim dim_hidden: hidden layer dim num_layers: number of layers

Copy config.json to the folder Params: config 要保存的字典 config_name 要保存的文件名 path 要保存的文件夹

Return MLP and its feature dim of output Params: dim_in: input feature dim dense: list of layer dims activation_type:

Func: iSTFT Params: same as stft

设置优化器 :params params 优化的目标 config_optim: 配置 :ret optimizer

Func: STFT Params: nfft: nfft hop: hop size win: type of STFT analysis window, str, 'hann' or 'hamming'

Func: EM algorithm

Return: activation function Params: type: 'relu', 'tanh', or 'leakyrelu'

Func: Initialize CTF filter and noise variance Params: Observed_data: observation, shape of [bs, F, T]

Return: sampled latent variables Params: mean: mean of latent variables logvar: log variance of latent variables

Func: From EM output to audio Params: sptm_wo_DC: RVAE-EM output, complex spectrogram others: same as istft

读取音频

读取音频

Class: dataset for training and validating Params: spch_index_txt: clean speech path saved as .txt rir_index_

Class: testset Params: spch_dir: Reverberant speech folder path fs: sample frequency

Class: EM algorithm Params: model: RVAE network EM kparams: parameter dict CTF_len: CTF filte

Params: config: list of parameters, [[input channel, output channel, kernel size, stride, padding]] norm: use batch n

spch_path: 存放干净语音文件名的txt文件地址 rir_path: 存放RIR文件名的txt文件地址 fs: 采样率Hz spch_len: 音频持续时间s norm_mode: 归一化方式

Func: from audio to magnitude spectrogram without DC component Return: magnitude and phase spectrograms Params: audio: audio [bs,

Func: from audio to RVAE-EM input Params: same as aud2sptm_woDC

Initialize parameters in convolutional and transpose convolutional layers (zero-mean Gaussian) Params: mean: mean std: standa

Func: from magnitude spectrogram without DC component to audio Return: Audio Params: mag_woDC: magnitude spectrogram phas

Func: M-step, update CTF filter, shape of [P+1], complex Params: Observed_data: observation in single band [T], complex