From: Paul L. <Paul.Lamere@Sun.COM> - 2003-10-31 18:01:14
|
I forgot to cc the devel list on this as well, so Iadded it this time. Philip Kwok wrote: > Paul, > > Thanks for writing this up. I think this looks good. My comments are > embedded: > >> Phil: >> >> At the staff meeting we received the action item to figure out a good >> way to get good timing information for the FrontEnd. >> >> Here is a suggestion: >> >> add a time stamp to the Data class. The time stamp would indicate >> the time in milliseconds for when the data in the Data packet was >> collected. >> >> Cepstrum and Feature classes would get their timestamp from the first >> of the Audio packets used to create the cepstrum/Feature. > > > The front-end Windower converts audio into audio windows that are 10ms > (the 'windowShift') apart. I wonder if those audio windows should have > the window-shifted timestamp instead of the original timestamp. The > Windower simply add the windowShift with each new window. That seems fair. If I understand you correctly a window could look like this (excuse the ascii graphics) . ... | ............. | | ..............| |...............|.. ..|...............|..... .....|...............|....... .....|...............|....... A B With some overlap from the previous and next frames. We want the timestamp for this frame to not include the previous frame overlap. That is, the timestamp should for point B, not point A. > >> >> The Data class would be extended to include the timestamp in the >> constructors: >> >> Data(long timeStamp) >> Data(Signal signal, long timeStamp) >> Data(Utterance utterance, long timeStamp) >> >> >> And it would include a method: >> >> long getTimeStamp(); >> >> >> Any thoughts on this? >> >> Paul >> >> > > |