Overview

In this project, we aim to build a machine learning model that can identify the gender of a person from their voice recording. In the process, we use two intermediary data representation format of the audio clips- Mel Spectrogram (Mel) and Mel-Frequency Cepstral Coefficients (MFCC).

Datasets

[MCV] Common Voice by Mozilla.org (https://www.kaggle.com/datasets/mozillaorg/common-voice)

[DLS] Bengali Common Voice Speech Dataset (https://www.kaggle.com/competitions/dlsprint)

Proposed Solution

Mel-Frequency Cepstral Coefficients (MFCC)

Mel Spectrogram

Notebook Details

Training

The training folder contains four notebooks. Each of the notebooks are named as: [Data-Type]_[Dataset]_[Model]. These notebooks are used to train individual models on the train datasets.

└── training
    ├── mel_dls_resnet50_train.ipynb 
    ├── mel_mcv_resnet50.ipynb       
    ├── mfcc_dls_train_resnet50.ipynb
    └── mfcc_mcv_resnet50.ipynb

Evaluation

The evaluation folder contains four notebooks. Each of the notebooks are named as: [Data-Type]_[Datase#1]_on_[Dataset#2]. The models trained on Dataset#1 are used to evaluate Dataset#2.

└── evaluation
    ├── mel_dls_on_mcv.ipynb
    ├── mel_mcv_on_dls.ipynb
    ├── mfcc_dls_on_mcv.ipynb        
    └── mfcc_mcv_on_dls.ipynb

In the report mentioned in the presentation, the comparison between models are shown.

Model Details

• Architecture: ResNet50
• Learning Rate: 0.0001
• Adam Optimizer

Presentation Report

https://docs.google.com/presentation/d/14BWOq6YSmO3GqZHEvCou43Z5A4dlOmKq4pjqUqgZALU/

References

[1] Speaker Gender Recognition Based on Deep Neural Networks and ResNet50 (https://doi.org/10.1155/2022/4444388)

[2] A Machine Learning Approach to Automating Bengali Voice Based Gender Classification (https://ieeexplore.ieee.org/document/9117385)