SpeakerRecognition_tutorial

• A pytorch implementation of d-vector based speaker recognition system.
• ResNet-based feature extractor, global average pooling and softmax layer with cross-entropy loss.
• All the features (log Mel-filterbank features) for training and testing are uploaded.
• Korean manual is included ("2019_LG_SpeakerRecognition_tutorial.pdf").

Requirements

python 3.5+
pytorch 1.0.0
pandas 0.23.4
numpy 1.13.3
pickle 4.0
matplotlib 2.1.0

Datasets

We used the dataset collected through the following task.

• No. 10063424, 'development of distant speech recognition and multi-task dialog processing technologies for in-door conversational robots'

Specification

• Korean read speech corpus (ETRI read speech)
• Clean speech at a distance of 1m and a direction of 0 degrees
• 16kHz, 16bits

We uploaded 40-dimensional log Mel-filterbank energy features extracted from the above dataset.
python_speech_features library is used.

You can extract the features using the code below:

import librosa
import numpy as np
from python_speech_features import fbank

def normalize_frames(m,Scale=True):
    if Scale:
        return (m - np.mean(m, axis=0)) / (np.std(m, axis=0) + 2e-12)
    else:
        return (m - np.mean(m, axis=0))

audio, sr = librosa.load(filename, sr=sample_rate, mono=True)
filter_banks, energies = fbank(audio, samplerate=sample_rate, nfilt=40, winlen=0.025)
filter_banks = 20 * np.log10(np.maximum(filter_banks,1e-5))
feature = normalize_frames(filter_banks, Scale=False)

1. Train

24000 utterances, 240 folders (240 speakers)
Size : 3GB
feat_logfbank_nfilt40 - train

2. Enroll & test

20 utterances, 10 folders (10 speakers)
Size : 11MB
feat_logfbank_nfilt40 - test

Usage

1. Training

Background model (ResNet based speaker classifier) is trained.
You can change settings for training in 'train.py' file.

python train.py

2. Enrollment

Extract the speaker embeddings (d-vectors) using 10 enrollment speech files.
They are extracted from the last hidden layer of the background model.
All the embeddings are saved in 'enroll_embeddings' folder.

python enroll.py

3. Testing

For speaker verification, you can change settings in 'verification.py' file.

python verification.py

For speaker identification, you can change settings in 'identification.py' file.

python identification.py

How to train using your own dataset

1. Modify the line 21 in train.py

train_DB, valid_DB = split_train_dev(c.TRAIN_FEAT_DIR, val_ratio)

• 'c.TRAIN_FEAT_DIR' in configure.py should be the path of your dataset
• 'c.TRAIN_FEAT_DIR' should have the structure as: FEAT_DIR/speaker_folders/features_files.p

2. Modify the line 31 in DB_wav_reader.py

def find_feats(directory, pattern='**/*.p'):
    """Recursively finds all files matching the pattern."""
    return glob(os.path.join(directory, pattern), recursive=True)

• I assumed that all the features are extracted in '.p' format.
• If you want to change the extension, please change line 31 in DB_wav_reader.py
• pattern='**/*.p' should be changed according to your feature format.
• If you didn't extract features yet, please do that using python_speech_features library.
• I didn't upload the code for feature extraction.
• Of course, you can use other libraries.

3. Change the line 12 in SR_Dataset.py

def read_MFB(filename):
    with open(filename, 'rb') as f:
        feat_and_label = pickle.load(f)
        
    feature = feat_and_label['feat'] # size : (n_frames, dim=40)
    label = feat_and_label['label']

• It is assumed that the feature file format is pickle. You need to change the code according to the format.
• You have to change the function 'read_MFB' according to your situation.
• From line 12 to line 16, we load feature (it is assumed the feature is saved using pickle) and label.
• Feature size should be (n_frames, dim) as written in the comment. Label should be the speaker identity in string.
• You can remove from line 20 to 24 because it is assumed that the front and back of the utterance is silence.

4. Change other options

• Be aware that all the settings are set to the small dataset as the training set in this tutorial is very small.
• According to your dataset, you can make the model wider (increase the number of channels) and deeper (change to the ResNet-34) or increase the number of input frames ('NUM_WIN_SIZE' in configure.py).
• More advanced loss function or pooling method (attentive pooling...) also can be used (not implemented here).

Author

Youngmoon Jung ([email protected]) at KAIST, South Korea