speech_signal_processing

Any question, you can pull a issue or email me.

Description

• VAD.py is the first project.

• MFCC_DTW.py is the second project.

• GMM_UBM.py is the third project, and GUI.py is the GUI of this project.

• d_vector.py is final project, and Final_GUI.py is the GUI of this project.

• feature dir saved model and feature file. you can download it from here, code is iwmf.

• All the reports are in report dir.

Requirement

python 3.x,windows

Any other package, run code below

pip install -r requirements.txt

or

pip install dtw librosa fastdtw tqdm sidekit tensorflow keras numpy scipy pyqt sklearn

NOTE:you can use mirror to speed up，refer blog

dataset

download dataset for d-vector from voxceleb.

Experiment Log

MFCC+DTW

DTW	Time(s)	Acc(%)
accelerated_dtw	92	83.72
accelerated_dtw+pre-emphasis	105	74.42
fastdtw	71	60.47
fastdtw+pre-emphasis	79	65.12

Summury:The results of fastdtw is bad than accelerated_dtw, so I suggest you to use accelerated rather than fastdtw if you prefer more on accuracy.

MFCC

blog

GMM

blog

paper

scikit-learn

SIDEKIT

MFCC+GMM

Please read report for more details.

d-vector

we train our model on voxceleb dataset, more details, please read report.

model	time(s)	train_acc	valid_acc	epoch	test_acc	test_time
nn	56s	0.5321	0.4672	50	0.3682	11.92
lstm	2906	0.788	0.5472	100	0.4371	49.53
gru	2977	0.9385	0.7484	30	0.3766	70.05

inference:paper

inference_gru:paper

inference_lstm:paper

Reference

1. audio-mnist-with-person-detection

2. dVectorSpeakerRecognition

3. speaker-verification

4. voxceleb