The author explores the limited value of SSL representations from Wav2Vec2 models in improving on-device speech enhancement systems under low-SNR conditions.
