# ποΈ CodecEval-Africa
**A comprehensive benchmark for neural speech codecs on African-accented English and conversational audio**
---
## π Table of Contents
- Overview
- Features
- Supported Codecs
- Evaluation Metrics
- Datasets
- Quick Start
- Installation
- Usage
- Configuration
- Results
- Project Structure
- Contributing
- License
---
## π― Overview
**CodecEval-Africa** is a comprehensive benchmarking framework for evaluating state-of-the-art neural speech codecs on African-accented English, conversational audio, and multilingual African speech. This project provides:
- **7 Neural Codecs** - Evaluation of leading speech compression models
- **6 Quality Metrics** - Comprehensive perceptual and objective quality assessment
- **African Speech Datasets** - Focused evaluation on African-accented English and multilingual African languages
- **Multi-Bitrate Analysis** - Performance evaluation across various compression rates
- **GPU Acceleration** - Optimized for high-performance computing environments
---
## β¨ Features
- ποΈ **Multi-Codec Support** - EnCodec, DAC, LanguageCodec, UniCodec, SemantiCodec, FocalCodec, WavTokenizer
- π **Comprehensive Metrics** - NISQA, ViSQOL, DNSMOS, Speaker Similarity, Prosody, ASR WER
- π **African Speech Focus** - Specialized evaluation on African-accented English and multilingual African language datasets
- β‘ **GPU Accelerated** - Optimized batch processing for SLURM clusters
- π **Bitrate Analysis** - Performance evaluation from ultra-low to high bitrates
- π§ **Modular Design** - Easy to extend with new codecs and metrics
- π **Detailed Logging** - Comprehensive output for analysis and visualization
---
## π§ Supported Codecs
| Codec | Bitrates | Features | Hardware |
|-------|----------|----------|----------|
| **EnCodec** | 3.0, 6.0, 12.0, 24.0 kbps | Causal streaming, Meta's production codec | GPU |
| **DAC** | 8, 16, 24 kbps | High-quality residual codec, multiple sampling rates | GPU |
| **LanguageCodec** | β¦