How Robust Are Neural Audio Codecs for African Speech? A Multi-Task Benchmark and the Limits of Perceptual Quality
The study benchmarked seven neural audio codecs on African speech datasets to see how well they perform for automatic speech recognition and speaker verification. Findings show that reference-based metrics like STOI are more reliable than perceptual scores for predicting downstream success, and performance varies significantly by task and speech type. Joint codec selection with adaptation is key.