Check if value is an outlier.

Convert FBANK to linear spectrogram. Args: feats: FBANK feats, shape: [length, channels] Returns: linear_feats

Compute softmax with input mask.

used to split pinyin into intial and final phonemes

Update results_dictionary from a list of values. Used to update results_dictionary to be returned by parse_values when encountering a clause

Update results_dictionary with a scalar value. Used to update the results_dictionary to be returned by parse_values when encountering a claus

CTC Alignmnet according encoder probs

transform original feature to normalized feature

transform original feature to normalized feature

Applies label smoothing. See https://arxiv.org/abs/1512.00567. Args: inputs: A 3d tensor with shape of [N, T, V], where V is the number of v

Bark to Hertz conversion */

Batch-wise shuffling of the data entries. Each data entry is in the format of (audio_file, file_size, transcript). If epoch_index is

implementation func

clip norm using tf.clip_by_norm

collapse whitespace in text

complex conjugation

complex number div real number

complex multiplication

real and complex mutiplication

compute cmvn for filtered entries

because of memory issue, we used incremental approximation for the calculation of cmvn

compute energy of frame */

compute the logit length

compute variance */

Convert FLAC to WAV and split the transcript. Args: input_dir: the directory which holds the input dataset. source_name: the name

Convert tar.gz to WAV and split the transcript. Args: dataset_dir : the directory which holds the input dataset. subset : the name o

Convert SPH to WAV and split the transcript. Args: dataset_dir : the directory which holds the input dataset. subset : the name of t

Convert rar to WAV and split the transcript. Args: dataset_dir : the directory which holds the input dataset. total_csv_path : the resu

data loader

data loader

data loader

data loader

Removes the hyperparameter with key 'name'. Does nothing if it isn't present. Args: name: Name of the hyperparameter.

delete punctuation tokens

Radix-2 DIT FFT */ isign=-1 ==> FFT, isign=1 ==> IFFT */

do dither */

do pre_emphasis on frame */

Levinson durbin recursion */

Download and extract the given split of dataset. Args: directory: the directory where to extract the tarball. url: the url to dow

Download and extract the given split of dataset. Args: directory: the directory where to extract the tarball. url: the url to down

Download and extract the given dataset. Args: directory: the directory where to extract the tarball. url: the url to download the

Downloads url source Args: dataset_root_dir: local filepath dataset: name of dataset source: url of resource

Forward input chunk by chunk with chunk_size like a streaming fashion Here we should pay special attention to computation cache

evaluate the model

evaluate the model

expand abbreviations in text

expand symbols in text

Extract files from compressed source file Args: dataset_root_dir: local filepath dataset: name of dataset source_file

Extractor mel,f0 and energy and save in output_dir

x:real part y:image part n: fft length

x:real part y:image part n: fft length

Force alignment Args: ctc_probs: hidden state sequence, 3d tensor (B,T, D) y: id sequence tensor 2d tensor (B,L) int bla

Forward just one chunk Args: xs (torch.Tensor): chunk input offset (int): current offset in encoder output time stam

beam search for freeze only support batch=1 Args: samples: the data source to be decoded beam_size: beam size