Speech:Exps 0160

From Openitware
Jump to: navigation, search


Author: David Meehan

Date: 2/12/2014

Purpose: Run a train using the first_5hr data set with 2000 senones.

Details: After reading more of Eric's experiment logs, I found a very interesting website he linked to. The page was part of the Sphinx FAQ, and contained a chart of senone's compared to hours. The chart suggests that 4-6 hours = 1000-2500 senones. This seems to contradict the Sphinx guide on training which suggests to use 200 for 5 hours. But perhaps this discrepancy explains why using 200 got such poor results when I tried 200. I will rerun the first_5hr train with 2000 senones instead. Eric's baseline was a 30% error rate. I would like to recreate this as closely as possible. My rate may be higher as he reports the first five hour test having worse results than the last (which he used). But for the sake of comparing my results with my first two experiments I would like to keep this constant.

Results N/A