Task #3699
closed
Task #3680: RA4a - Automatic error prediction
Task #3698: Experiment with one-class clasification for join cost enhancements
Compute features (MFCC, LPC, LPCenv and FFTpow)
Added by Tihelka Dan about 9 years ago.
Updated almost 9 years ago.
Description
For the whole DB compute features like:
- pitch-asynchronous/pitch-synchronous MFCC
- pitch-asynchronous/pitch-synchronous LPC
- pitch-asynchronous/pitch-synchronous FFT power
For pitch-synchronous parameterization use static window size to avoid zeros in FFT. Storem them in ASF files (like candsel features used now).
- % Done changed from 0 to 10
Features computed (not all yet, though, just those following the paper, i.e. 20-20 (20msec windows, 20msec shift) MFCC, LPC and FFTpow). The next step is to compute distance vectors.
- Target version set to RA4: Automatic error prediction and signal modification
- Status changed from Assigned to Resolved
- % Done changed from 10 to 100
Distances computed. Starting to experiment with one-class-classifier.
More details about the whole experiment described in Interspeech 2016 paper, see #3809.
Another set of features computed: 04-20 (20msec windows, 4msec shift), MFCC, LPC, LPCenv and FFTpow. Also distances are going to be computed and the performance of classifiers compared.
Experiments with distances among MFCC, LPC, LPCenv and FFTpow feature vectors, computed for signal framing:
- 20-20 (20msec windows, 20msec shift)
- 04-25 (25msec windows, 4msec shift)
- 12-25 (25msec windows, 12msec shift)
- pm-25 (25msec windows, pitch-synchronous shift)
were submitted to SpeCom 2016 conference (http://specom.nw.ru/).
Results are mixed, we need to focus on error analysis now.
- Subject changed from Compute features to Compute features (MFCC, LPC, LPCenv and FFTpow)
- Status changed from Resolved to Closed
Also available in: Atom
PDF