Methods
Learning a multiresolution signal representation
From a waveform to trainable wavelet analysis, learned denoising and a task-specific representation.
A fixed transform encodes prior choices about how to describe a signal. The research behind SigWavNet and the wavelet-packet models asks which parts of that analysis can be learned with the task, while retaining a multiresolution structure.
The decomposition, denoising and downstream network are separate methodological choices. SigWavNet studies learned wavelet analysis; wavelet-packet models decompose approximation and detail branches; NWPA uses learned wavelet-packet representations for speech enhancement. LFST extends the broader research direction through fractional superlets and a spectro-temporal encoder.
Walk through the method
Open each step to inspect the inputs, process and outputs described by its authors.
01 Represent the waveform at several scales
- Input
- Speech waveform
- Output
- Multiresolution coefficients
Learnable filterbanks replace a wholly fixed front end. The decomposition separates signal content across resolutions so that the network can use structures at different temporal and frequency scales.
02 Learn which coefficients to retain
- Input
- Decomposed signal
- Output
- Learned, denoised representation
The wavelet studies use learnable thresholding to attenuate less useful coefficients. The asymmetric hard-thresholding function is trained with the model rather than selected as an independent preprocessing rule.
03 Connect the representation to the task
- Input
- Learned representation
- Output
- Emotion prediction or enhanced speech
For emotion recognition, convolutional and recurrent components learn spectral and temporal structure, with attention in the relevant architectures. For enhancement, the bidirectional autoencoder reconstructs speech. These objectives require different evaluation criteria.
04 Inspect the experimental evidence
- Input
- Model, data and evaluation protocol
- Output
- Task-specific evidence and limitations
Read each paper’s dataset, partitioning, metrics and ablations with its reported results. Follow the linked implementation for the corresponding configuration; emotion-classification scores and speech-enhancement scores describe different tasks.
Projects
- 01
Projects
Adaptive time-frequency learning
A research file connecting learned wavelets, wavelet packets, denoising and fractional superlets.
Research directions
- 01
Research programs
Learnable signal representations
Wavelets, wavelet packets and fractional superlets as adaptive representations for emotion analysis and speech enhancement.