rahelFM/Transcription-Lesotho

## Speech-to-Text Benchmarking on Code-Switched isiZulu-English Dataset This repository contains a notebook demonstrating the benchmarking of two speech-to-text models: Whisper Small and Wav2Vec2, trained on a code-switched isiZulu-English audio dataset. The notebook includes steps followed to do exploratory data analysis, error analysis, model performance evaluation or explorative model analysis with Character Error Rate and Word Error Rate, and incorporates visualization, explainability tools, and human-in-the-loop techniques as we perfomed some of the error analysis by using human intelligence. ## Table of Contents 1. Overview 2. Setup 3. How to Run the Notebook 4. Reproducing Results 5. Model Evaluation & Error Analysis 6. Insights and Suggestions for Improving Performance 7. License ### Steps Involved: - Exploratory Data Analysis (EDA) on the dataset - Error Metrics: Computation of CER and WER - Model Evaluation: Training and evaluation of Whisper Small and Wav2Vec2 - Error Analysis: Focused error analysis on Whisper Small model performance - Explainability: Used substitution, insertion and deletion to have insight - Human-in-the-loop: Suggestions for further performance improvements ## Setup ### Prerequisites Before running the notebook, ensure the following libraries are installed in your environment: | Library | Purpose & Use Case | |------------------------------------|------------------------------------------------------------------------------------| | `os` | Handling file paths and directories, and interacting with the operating system. | | `librosa` | Audio analysis library for feature extraction, signal processing, and transformations.| | `librosa.display` | For visualizing audio signals and spectrograms. | | `matplotlib.pyplot` | Plottin …