I'm reading through the LogMath library trying to better understand exactly what it does.
I was wondering if you could explain the reasoning for choosing a base of 1.0001 or 1.0003. I'm having a hard time undersatading the comment's explanation of this choice of base. Were these numbers found experimentally to give the best 'resolution'?
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hi,
I'm reading through the LogMath library trying to better understand exactly what it does.
I was wondering if you could explain the reasoning for choosing a base of 1.0001 or 1.0003. I'm having a hard time undersatading the comment's explanation of this choice of base. Were these numbers found experimentally to give the best 'resolution'?
Yes, in sphinx3 it's important for the scores to fit 4 bytes integer so you need to balance between resolution and value range.
In sphinx4 logmath is actually not needed at all since it uses floating point.
Do you mean that computation does not need to be done in the log domain or that the LogMath java class is not needed?
Computation must be done in log domain but it can use natural log base, not 1.0001, so LogMath class is not really needed.