Abstract
- Principal Mel-Spectrum Components (Feature)
- Temporal Pooling Functions (Model)
- Single Hidden Layer Neural Network, thus Multi-layer Perceptron (Classifier)
Audio Preprocessing
Feature: PMSC (Principal Mel-Spectrum Components)
- Original Data: 30s, 22.05KHz, mono, wav
- Process Steps:
- DFT (spectral domain)
we compute DFTs over windows of 1024 samples on audio at 22.05 KHz (i.e. roughly 46ms) with a frame step of 512 samples. - Mel-Compression
we run the spectral amplitudes through a set of 256 mel-scaled triangular filters to abtain a set of spectral energy bands. - Principal Component analysis whitening (PCA whitening)
we compute the principal components of a random sub-sample of training set. In order to obtain features with unitary variance, we multiply(乘以) each component by the inverse square of its eigenvalue(特征值平方的倒数). ---- PCA whitening.
Model
PFC (Pooled Features Classifier)
- Pooling Operation
the model applies a given set of pooling functions (how many?) to the PMSC features, and sends the pooled features to a classifier(MLP, with hidden layer of 2000 units, sigmoid activation, L2 weight decay and cross-entropy cost). - Classify
each pooling window is considered as a training example for the classifier, and average the predictions of the classifier over all the windows of a given clip to obtain the final classification (what is the rule?).
Tasks
- Classification (train/test task)
the MLP outputs an affinity prediction for each class (pooling functions tread each pooling window as a training example). - Tagging
- Affinity
the affinity scores for a song is thus directly the output of the MLP. - Binary Classification
choose the threshold that optimizes the F1-score on the validation set.
Tools