utils

Defined in: utils.py

Index

• class DiarizationDataset
  • func __getitem__
  • func read_rttm
• func make_rttm
• func get_metrics
• func plot_annot

class DiarizationDataset()

Defined in utils.py

class DiarizationDataset(dataset_name=None
                 data_dir=None, 
                 sr=16000, 
                 window_len=240, 
                 window_step=120, 
                 transform=None,
                 batch_size_for_ecapa=512,
                 vad_step=4,
                 split='full',
                 use_precomputed_vad= True,
                 use_oracle_vad= False,
                 skip_overlap= True)

Create an abstract class for loading the dataset. This class applies the necessary pre-processing and x-vector feature extraction methods to return the audio file as a bunch of segmented x-vector features to use it directly in the clustering algorithm to predict speaker labels. The module uses the pre-computed X-vectors if available otherwise extract it during the runtime.

Parameters:

Argument	Detail
dataset_name:	str, Name of the pre-existing dataset to use. Options: ami, ami_dev, voxconverse
data_dir:	str, Directory for any dataset other the options specified in dataset_name. Both dataset_name and data_dir cannot be None
sr:	int, Sampling rate of the audio signal
window_len:	int, Window length (in ms) of each of the audio segments to be passed for feature extraction
window_step:	int, Step (in ms) between two windows of audio segments to be passed for feature extraction
transform:	list, List of transforms like mel-transform to be performed on audio while preprocessing, default = None
batch_size_for_ecapa:	int, Batch size of audio segments while performing feature extraction using ECAPA-TDNN
vad_step:	int, Number of windows to split each audio chunk into. Argument used by Silero-VAD module
split:	str, Argument defining type of split of dataset, default = 'full' indicates no split
use_precomputed_vad:	bool, If True, downloads precomputed Voice Activity Detection label output for the dataset. Only available for dataset options specified in dataset_name
use_oracle_vad:	bool, If True, model does Voice Activity Detection directly from groundtruth rttm files bypassing the Silero VAD module.
skip_overlap:	bool, If True, model skips the windows with multiple speakers speaking by inspecting the groundtruth rttm files

Class Functions:

1. __getitem__: def __getitem__(self, idx)

Parameters:

Argument	Detail
idx:	int, Index to the required audio in the list of audio in root directory

Returns:

Variable	Detail
audio_segments:	torch.Tensor, (n_windows, features_len) Tensor of feature vectors of each audio segment window
diarization_segments:	torch.Tensor, (n_windows, n_spks) Tensor containing ground truth of speaker labels, 1 if i-th window has j-th speaker speaking, else 0
audio_segments:	torch.Tensor, (n_windows, features_len) Tensor of feature vectors of each audio segment window
speech_segments:	torch.Tensor, (n_windows,) Tensor with i-th value 1 if VAD returns presence of speech audio in i-th window, else 0
label_path:	str, Path of the rttm file containing labels for the 'idx' wav audio

2. read_rttm: def read_rttm(self, path)

Parameters:

Argument	Detail
path:	str, Path to the RTTM diarization file

Returns:

Variable	Detail
rttm_out:	numpy.ndarray, (..., 3) Array with column 1 holding start time of speaker, column 2 holding end time of speaker, and column 3 holding speaker label

---

def make_rttm()

def make_rttm(out_dir, name, labels, win_step):

Defined in utils.py

Create RTTM Diarization files for non-overlapping speaker labels in var labels. Assumes non-speech part to have value -1 and speech part to have some speaker label (0, 1, 2, ...).

Parameters:

Argument	Detail
out_dir:	str, Directory where the output RTTM diarization files to be saved
name:	str, name for the audio files for which diarization was predicted
labels:	int, Speaker/ Non-speech labels assigned to different audio segments based on the win_step used to extract feature vectors
win_step:	int, Step (in ms) between two windows of audio segments used for feature extraction

Returns:

Variable	Detail
return variable:	str, Path to the saved RTTM diarization file

---

def get_metrics()

def get_metrics(groundtruth_path, hypothesis_path, collar=0.25, skip_overlap=True):

Defined in utils.py

Evaluate the diarization results of all the predicted RTTM files present in hypothesis directory to the grountruth RTTM files present in groundtruth directory.

Parameters:

Argument	Detail
groundtruth_path:	str, directory of groundtruth rttm files
hypothesis_path:	str, directory of hypothesis rttm files
collar:	float, Duration (in seconds) of collars removed from evaluation around boundaries of reference segments
skip_overlap:	bool, If True, calculates Diarization Error Rate ignoring the overlapped region

Returns:

Variable	Detail
metric:	pyannote.metrics, Pyannote metric class having diarization DERs for all the files.

---

def plot_annot()

def plot_annot(name="IS1009a", collar=0.25, skip_overlap=True, groundtruth_path=None, hypothesis_path=None):

Defined in utils.py

Calculate the Diarization Error Rate for filename specified, and print the groundtruth and hypothesis time series plot.

Parameters:

Argument	Detail
name:	str, Name of the file whose time series plot is to be generated. File must be present in the hypothesis_path folder
collar:	float, Duration (in seconds) of collars removed from evaluation around boundaries of reference segments
skip_overlap:	bool, If True, calculates Diarization Error Rate ignoring the overlapped region
groundtruth_path:	str, Directory of groundtruth rttm files
hypothesis_path:	str, Directory of hypothesis rttm files

---