Hi, I am a DSP engineer working on audio encoders and decoders .We are trying to implement a low cost psychomodel for a fixed point DSP. I request the group members to provide a few ideas about implementing low cost psychoacoustic model.What are the most important things needed to improve the quality of encoder( We have a typical 24 bit fixed point DSP and we need the psychomodel in around 20 MIPS)
We are trying to determine the cheapest way of getting the perceptual entropy for a frame. In lame the algorithm for getting PE is as follows: ++++++++++++++++++++++++++++++++++++++++++ 1) Determine PE assuming long blocks 2) Determine the maximum and minimum energy in the three short blocks 3) if(PE > 1000) and (max energy > min energy) override PE value by following 400 *log(mx/ mx + epsilon) I want to modify the algorithm as follows +++++++++++++++++++++++++++++++++++ 1 Since SMR values ar not required by me I wont calculate PE for long blocks .Instead I will set it to a higher nominal value (say 900) instead of 700. For CBR case it wont affect the size of frame. 2)If (max energy > min energy) override PE value by following 400 *log(mx/ mx + epsilon) This saves me from computing long block PE (means I can remove long FFT,unpred calc,spread fn convolution,Tonality calc,pre-echo control etc) And for long blocks PE would anyway be around 1000 as that would be my default value i wish to have some comments on this idea of mine.Specifically what do you guys think about increasing the nominal value of PE from 700 to 900?? Thanks and regards Manjunath _______________________________________________ mp3encoder mailing list [EMAIL PROTECTED] http://minnie.tuhs.org/mailman/listinfo/mp3encoder
