Re averaging of the speech information, I think we need some sort of update every 20ms, but the amount of information varies depending on the current speech frame.
There is a neat duality that helps us. The most perceptually important frames that need high quality LSP quantisation (high energy voiced speech) change slowly. Unvoiced speech LSP information changes quickly but can be coarsely quantised. That's a useful corpus, I didn't know about that one. I use a training file from the TIMIT database, extracted years ago. The database can affect how well the codec works with speech from different sources. For example the standard samples I use are filtered to roughly telephone bandwidth. One day I would to make the codec work well with input speech that has different noise, HP/LP filtering, spectral tilt etc. Some sort of pre-processor (like used in speex) might be a good way to achieve this, for example apply equalisation, AGC, standard HP/LP filtering and noise reduction prior to encoding. Cheers, David On Sat, 2011-10-29 at 12:13 -0700, Bruce Perens wrote: > How well can we average the rate of key frames (that's an animation > term, meaning full LSP frames in this case) and delta frames over time? > We could probably come up with normal and worst cases for various > intervals by processing a speech corpus. Maybe the VoxForge one at > http://www.repository.voxforge1.org/downloads/SpeechCorpus/Trunk/Audio/Main/16kHz_16bit/ > > Thanks > > Bruce > > > > ------------------------------------------------------------------------------ Get your Android app more play: Bring it to the BlackBerry PlayBook in minutes. BlackBerry App World™ now supports Android™ Apps for the BlackBerry® PlayBook™. Discover just how easy and simple it is! http://p.sf.net/sfu/android-dev2dev _______________________________________________ Freetel-codec2 mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/freetel-codec2
