speech_signal_processing
speech_signal_processing copied to clipboard
kleinzcy
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
GMM
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
kleinzcy