EkabTawo/African-Accent-Word-Error-Rate-Analysis
Domain:
natural language processingRecord type:
project
Creator:
Eka
Host:
Comparative evaluation of Google Speech Recognition and IBM Watson Speech to Text using Word Error Rate (WER) across African English accents.
# African Accent Word Error Rate Analysis
## Overview
This undergraduate research project evaluates the performance of two commercial **Automatic Speech Recognition (ASR)** systems **Google Speech Recognition** and **IBM Watson Speech to Text**, on English accents from multiple African countries.
Using **Word Error Rate (WER)** as the primary evaluation metric, the study compares transcription accuracy across different African accents and analyses how factors such as **speech recognition engine**, **speaker gender**, and **accent variation** influence recognition performance.
---
## Objectives
- Compare the transcription accuracy of Google Speech Recognition and IBM Watson Speech to Text.
- Evaluate ASR performance across multiple African English accents.
- Investigate the influence of speaker gender on speech recognition accuracy.
- Measure transcription quality using Word Error Rate (WER).
- Identify strengths and limitations of each ASR system when processing diverse African English accents.
---
## Technologies Used
- Python
- Google Speech Recognition API
- IBM Watson Speech to Text
- Natural Language Processing (NLP)
- Word Error Rate (WER)
- Speech Processing
- Microsoft Excel
---
## Repository Structure
```text
African-Accent-Word-Error-Rate-Analysis/
│
├── src/
│ └── word_error_rate_analysis.py
│
├── data/
│
├── docs/
│ └── african_accent_word_error_rate_report.pdf
│
├── presentation/
│ └── african_accent_word_error_rate_presentation.pptx
│
└── README.md
```
---
## Methodology
1. Collect speech recordings from speakers representing selected African countries.
2. Categorise recordings by country and speaker gender.
3. Process each recording using:
- Google Speech Recognition
- IBM Watson Speech to Text
4. Generate automatic transcriptions from both ASR systems.
5. Compare each generated transcript with manually verified reference transcriptions.
6. Calculate Word Error Rate (WER) for every transcription.
7. Compare the performance of both ASR s …
Similar
Non-word error detection in current South African spellcheckersCUPE Universal Phoneme Encoder Substitution for Low-Resource Word Error Rate Reduction in Common VoiceNon-Word Error Correction for LugandaCorrelation Between Representation Quality of Semi-Supervised Models and Downstream Word Error Rate in Low-Resource Languagesrobello2/african-accent-classifierrobello2/african-accent-classifier-test