Project Speech

Overview

Project Speech was my university dissertation project focused on building and evaluating a speech-to-text model. The main goal was to test whether training with generic speech, impaired speech, or a mixture of both would improve or worsen the model's performance.

Features

• Custom Speech-to-Text Model: Developed and trained using Python and deep learning frameworks.
• Dataset Experimentation: Compared model performance on generic speech, impaired speech, and a mixed dataset.
• Performance Analysis: Evaluated accuracy and robustness across different data types.
• Research Focus: Explored the impact of dataset composition and model architecture on speech recognition.

Findings

• The results showed that model performance depended more on the type of model architecture used and the inclusion of a mixed dataset, rather than just the type of speech data.
• Using a mixture of generic and impaired speech data generally improved robustness.

Technologies Used

• Python
• Deep Learning (e.g., TensorFlow or PyTorch)
• Data preprocessing and augmentation

Challenges & Solutions

• Data Collection: Sourced and prepared both generic and impaired speech datasets.
• Model Selection: Experimented with different architectures to find the best fit.
• Evaluation: Designed fair tests to compare model performance across datasets.