Scribd
Upload
21 views3 pages

CS4622 Machine Learning PROJECT

This document outlines a machine learning project involving speech recognition tasks including speaker, age, gender, and accent classification. The project has two phases: individual modeling using wav2vec features from different layers, with models evaluated via Kaggle competitions; and a group paper combining all layers and improving the joint model, including literature review, explanations, and findings. Participants are evaluated on their individual Kaggle ranks and code quality, and the group paper will undergo blind peer review. Deadlines for the competitions and paper are September 24th and October 8th, respectively.

Uploaded by

Raveen Shamentha
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
21 views3 pages

CS4622 Machine Learning PROJECT

This document outlines a machine learning project involving speech recognition tasks including speaker, age, gender, and accent classification. The project has two phases: individual modeling using wav2vec features from different layers, with models evaluated via Kaggle competitions; and a group paper combining all layers and improving the joint model, including literature review, explanations, and findings. Participants are evaluated on their individual Kaggle ranks and code quality, and the group paper will undergo blind peer review. Deadlines for the competitions and paper are September 24th and October 8th, respectively.

Uploaded by

Raveen Shamentha
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 3

CS4622 - Machine Learning

Project - Speaker, age, gender and accent recognition using


wav2vec base
Dr.R.T.Uthayasanker
August 29, 2023

Project Description
Dataset : AudioMNIST is the dataset used to create the features. Check this link for further
details about the dataset Link.

Below structure is general speech-related task classification models

Figure 1: Overview of speech-related task classification model

This project has two phases:


1. Phase 1: Individual task - Classification model development and Kaggle competition
2. Phase 2: Group submission - 6-page research paper

Phase 1: Individual task


Wav2vec base is commonly used as a feature extraction model. There are 12 transformer layers
in the wav2vec base model. For this project features are extracted from the last 6 transformer
layers (transformer layer 7 to layer 12). For each layer separate kaggle competition is created.
Each train.csv file provided in the competition contains layer features and corresponding speaker,
age, gender, and accent labels.

1
• Label 1 - Speaker

• Label 2 - Age

• Label 3 - Gender

• Label 4 - Accent
Two (2) kaggle competitions will be allocated to each person. Your task is to build classifier
models for predicting all 4 labels individually using features in both training and validation CSV
files provided in the competitions. [Find your competition links here Link]

E.g.
• 1st competition - Layer X

1. Speaker recognition classifier model using layer X


2. Age recognition classifier model using layer X
3. Gender recognition classifier model using layer X
4. Accent recognition classifier model using layer X

• 2nd competition - Layer Y

5. Speaker recognition classifier model using layer Y


6. Age recognition classifier model using layer Y
7. Gender recognition classifier model using layer Y
8. Accent recognition classifier model using layer Y

Do data pre-processing, feature engineering, hyper-parameter tuning, dimensionality reduction,


cross-validation, and other techniques to improve the classifier accuracy. Upload the notebook
and predicted labels as solutions.csv file to Kaggle competition platform created for this project
(More details are provided in the Kaggle competition description, rules sections.)

Phase 2: Group task


Group formation : Maximum 3 people in one group. Only 2 groups can have 2 people.

In your group, the other members would have tried the other two pairs of layers. As a group, your
task is to combine all 6 layers and improve the prediction model, and write a 6-page conference
paper in IEEE format (Link).

For the conference paper writing, do a literature review, do ExplainableAI techniques, and inter-
pret the final model. Include your findings from this project and novel ideas during your feature
engineering and model development stages in the conference paper. Your paper should be up-
loaded in easy-chair. The link will be provided later.

The expected content of your conference paper can be found in this link, for your reference.

2
Evaluation
• Individual task:

– Classifier Model building - 40 marks


∗ Explainability (Interpreting the label predictions and any cross-relations with la-
bels) - 20 marks
∗ Good practice of ML (right evaluation strategy, ensemble methods, feature engi-
neering, etc.) - 10 marks
∗ Git repository (properly documented) - 5 marks
∗ Coding standard - 5 marks
– Kaggle Competition Rank - 20 marks

• Group submission: Conference paper - 40 marks

• We will evaluate your individual task using the ranks from the Kaggle competitions. In
addition, the submitted code (notebook / Git repository) will be evaluated based on above
mentioned criteria.

• We will evaluate your Group-wise conference paper based on a blind review process consid-
ering the quality, findings, interpretations, novelty, etc.

DEADLINES !!
Kaggle competition: 24th September 2023
Paper submission : 8th October 2023

You might also like