waletsoftware/Lesotho

# Lesotho Code-Switched Transcription Analysis Project Overview This project evaluates the performance of two automatic speech recognition (ASR) models, Whisper-Small and Wav2Vec2, on a code-switched Sesotho dataset. It includes: Exploratory Data Analysis (EDA) Model Evaluation (EMA) - Benchmarking WER, CER Error Analysis (EA) - Visualizing errors & linguistic insights Recommendations for Improvement # 1️⃣ Environment Setup You can run this project on Google Colab or locally. If running locally, install dependencies: # A. Using Google Colab 1. Upload the notebook (sesotho_transcription_analysis.ipynb) to Google Colab. 2. Mount Google Drive if required: ```bash from google.colab import drive drive.mount('/content/drive') ``` 3. Install dependencies (required libraries): ```bash !pip install transformers datasets jiwer torchaudio librosa phonemizer seaborn ``` # B. Running Locally 1. Clone this repository ```bash git clone github.com cd sesotho-asr-evaluation ``` 2. Create a virtual environment (recommended): ```bash python -m venv env source env/bin/activate # On Windows: env\Scripts\activate ``` 3. Install required packages: ```bash pip install -r requirements.txt ``` # 2️⃣ Running the Notebook 1. Open Jupyter Notebook: ```bash jupyter notebook ``` 2. Load sesotho_transcription_analysis.ipynb. 3. Run all cells sequentially. # 3️⃣ Dataset The dataset consists of 97 Sesotho code-switched audio samples (.wav) with transcriptions in transcriptions.csv. # Loading Data in Notebook ```bash import pandas as pd # Load transcriptions df = pd.read_csv("path/to/transcriptions.csv") df.head() ``` # 4️⃣ Reproducing Results A. Model Evaluation (EMA) To compare Whisper-Small and Wav2Vec2, we measure: * Character Error Rate (CER) * Word Error Rate (WER) # Running Model Benchmarking ```bash from jiwer import wer, cer wer_score = wer(ground_truth, predictions) cer_score = cer(ground_truth, predictions) print(f"WER: {wer_score:.2f}, C …