Speech:Exps 0257

<< Experiments

NOAA Experiment
Author: Marcel

Date: 9/30/14

Purpose: This set of experiments have the purpose to decode NOAA data as accurately as possible.


 * 001 - first NOAA decode experiment attempt using 0256/001 - result :65.4%
 * 002 - contains a failed attempt to replicate the Robust group tutorial setup
 * 003 - contains a language model build from NOAA corpus transcript  - outdated
 * 004 - decode NOAA using 0253/B12 training data  - result 61.3%
 * 005 - contains a language model build from NOAA corpus transcript  - updated
 * 006 - decode NOAA using 0253/B124hr_3170 training data  - result :92.7%
 * 007 - decode NOAA using 0253/B124hr_3170 training data different LM   - result :98.0%
 * 008 - decode NOAA using 0253/B124hr_3170 training data different features  - result :98.1%
 * 009 - decode NOAA using 0253/B12 training data with different LM  - result :94.7%
 * 010 - decode NOAA using 0253/A12 training data with different LM - result :94.6%
 * 011 - decode NOAA using 0253/B12 training data with different HMM - failed
 * 012 - decode NOAA using 0253/B12 training data with new LM(005) -result :45.4%
 * 013 - decode NOAA using A12 training data using new LM(005)
 * 014 - decode NOAA using 0253/B12 training data with NOAA dictionary - result :48.1%
 * 015 - decode NOAA using 0253/B12 training data with NOAA dictionary modified - result :48.2%
 * I noticed that for the first 70 wav files the error rate is way higher than in the rest of the files. This could be because I recorded the files at different times and maybe something wrong with the wav files. After I ran the sclite without them I got 37.9 %.
 * 016 - decode NOAA using 0253/B12 training data with new NOAA dictionary - result :48.0%
 * 017 - decode NOAA using 0253/B12 training data with new LM - result :47.2%
 * 018 - Statistical Language Model Using CMUCLMTK
 * 019 - decode NOAA using 0253/B12 training data with new LM (CMUCLMTK) - result :46.0%
 * 020 - decode NOAA using 0253/B12 training with new feats - result :47.0%
 * 021 - decode NOAA (first batch of bad wav files) using 0253/B12 training data   - result :77.0%
 * 022 - decode NOAA (remaining files) using 0253/B12 training data - result :37.4%