Research programs
Learnable signal representations
Wavelets, wavelet packets and fractional superlets as adaptive representations for emotion analysis and speech enhancement.
Research direction
A signal representation determines which details a model can retain. This work brings multiresolution signal processing into trainable neural networks, with filters and denoising operations learned alongside the task.
Related studies include SigWavNet, learnable wavelet-packet models, a bidirectional autoencoder for speech enhancement, and fractional superlets with a spectro-temporal emotion encoder. Papers and code are linked as distinct research outputs.
- Which time-frequency resolutions are useful for a task?
- How can learned denoising preserve meaningful speech structure?