Measuring benchmark optimization in speech recognition

Recent research reveals that some top-performing open-source speech recognition models may overfit to public benchmarks like VoxPopuli and LibriSpeech, reproducing transcripts even when audio contradicts them. This benchmark optimization inflates scores and misrepresents real-world transcription accuracy.