dinhanhx/automatic_speaker_recognition

A repos for USTH Digital Signal Processing 2020 Group 3 project. It's quite obvious in the title.

Automatic speaker recognition

A repos for USTH Digital Signal Processing 2020 Group 3 project. It's quite obvious in the title.

Introduction

What is speaker recognition

What is digital signal processing

This project harness the power of function mfcc from python_speech_features and model gmm from sklearn.

Read more about Mel frequency cepstrum coefficients and Gaussian Mixture model.

Datasets

This is the datasets. Remember to read AudioInfo.txt in Sunday datasets before processing.

135 .wav files of each person are 135 lines in transcripts/random_sentences.txt.

Note that Friday datasets is just an archive of Sunday datasets. Please use Sunday datasets.

Approach

Each Sunday_datasets/mix, Sunday_datasets/low, Sunday_datasets/high, I take 100 out of 135 .wav files of each person then I fit these files into a model which will represent that person's unique voice features. The rest 35 .wav files of each person are used to test the system of models.

100 .wav files are be shuffled to show that order of files is not important.

Plan:

• Train models with Sunday_datasets/mix folder.
• Train models with Sunday_datasets/low folder.
• Train models with Sunday_datasets/high folder.
• Then test each system of models on Sunday_datasets/mix, Sunday_datasets/low, Sunday_datasets/high folders.

Read our report for more details.

Project structure

To have clear view of folders and files

+--venv/
|
+--transcripts/
|  +--usth.txt
|  +--random_sentences.txt
|
|--datasets/
|  +--mix/
|  |  +AudioInfo.txt
|  |
|  +--low/
|  |  +AudioInfo.txt
|  |
|  +--high/
|     +AudioInfo.txt
|  
|--source_code/
|  +--Friday_script_models/ # Ignorable
|  +--models/ # Where models are saved as binary files
|  +--mfcc_gmm_func.py # Script of functions to call mfcc and gmm
|  +--requirements.txt # pip install -r requirements.txt
|  +--train_models.py
|  +--try_models.py
|
+--LICENSE
+--README.md
+--.gitignore

Group's member

• Vu Dinh Anh
• Ngo Ngoc Duc Huy
• Nguyen Quoc Thong
• Le Huy Quang
• Trinh Quoc Dat
• Luu Hai Nam
• Tran Minh Hieu
• Dinh Gia Luong

automatic_speaker_recognition's Issues

Refractoring code

I find that the code in mfcc_gmm_func.py is hard to read because:

• Overcommenting: some comments are redundant and distracting (e.g. it's quite obvious that wavfile is for handling wav file). In addition, I think comments are more helpful where the imported function/module is used than where it's imported.
• Function docstrings are unconventionally placed: They're put as comments before function declaration

Since everyone has their own tasks now, I'll try to do this without affecting its usage