Random Research Notes

This is a directory for random pages describing little incidental research results that perhaps deserve to be documented, but don't fit in anywhere else.

Notes on some ICASSP-2000 papers

Both as a reminder to myself and for any use they may be to others in the group, I've tried to make brief notes on the papers I noticed at the just-passed ICASSP-2000.

Label-conditional spectral feature averages

Following Hynek Hermansky's presentation of Sangita Sharma's research on using very long time windows, I calculated the average time-frequency planes for the different labels over a 70 hour data set. We also see a wide influence of several hundred milliseconds.

Temporal structure of Modulation Spectrogram features

The plots of the Modulation-filtered Spectrogram MSG1 features (that we've been using in Broadcast News) are quite surprising.

Comparison of different feature types

I did a series of tests of basic features with and without rasta processing. For the NUMBERS95 task I used, all variants gave more or less the same result, which was odd.

Frame-rate adaptive RASTA transfer function

The RASTA filter is partially defined in terms of its modulation-frequency response, so I modified the code to take a stab at maintaining the same response even when the steptime is not the canonical 10ms. This plot compares the responses for 10ms and 16ms frame rates.

Examples of phone-less words from STP

This is Eric's collection of Switchboard examples with audible words that appear to lack any acoustic correlates.

Correlation of different feature types

Florian Schiel asked Morgan for a reference to the 'received wisdom' that the cepstral transform 'decorrelates' spectral coefficients. I did some measurements comparing the two to see how much difference there really was.