Reproduce “What is Learnt by the LEArnable Front-end (LEAF)? Adapting Per-Channel Energy Normalisation (PCEN) to Noisy Conditions” on OpenResearch

About this paper

Hanyu Meng, Vidhyasaharan Sethu, Eliathamby Ambikairajah

Empirical analysis revealed that the Per-Channel Energy Normalization (PCEN) layer is the sole component within the LEArnable Front-end (LEAF) that undergoes substantial adaptation during training, while its spectral decomposition and smoothing filters remain stable. This insight led to a targeted PCEN adaptation strategy that efficiently improves LEAF's robustness to stationary and non-stationary noise using minimal noisy data.

Key insights

• Only the Per-Channel Energy Normalization (PCEN) layer within LEAF demonstrates significant parameter learning and adaptation during training across all tested speech tasks.
• The Gabor filterbank for spectral decomposition and the Gaussian low-pass filters for energy smoothing largely retain their initial configurations, showing negligible changes from their pre-set values.
• Targeted adaptation of only the PCEN layer provides an efficient and effective method to enhance a clean-trained LEAF model's robustness to both stationary (Gaussian) and non-stationary (babble) noise.

Summary by alphaXiv.