Speech:Exps 0104

From Openitware
Jump to: navigation, search

Modified Senome Value (2000): Test on Train 0102


Description

Author: Eric Beikman

Date: 6/16/2013

Purpose: To run a test on train for Experiment 0102.

Details: For this Experiment, we utilized the 30 minute last_5hr/test corpus to test train 0102. In Experiment 0102, we increased the amount of senones to 2000 to determine the effect on the resulting model.


To enable us to compare to our previous "Best" scoring experiment for this corpus (Experiment 0090), this experiment utilizes the language model, dictionaries, and phone list from that particular experiment. Results The initial attempts to start the decoder were met with failure. After some troubleshooting, it was determined that the Sphinx Trainer incorporates the amount of senones in the model into the filenames of the resulting models. The script which runs the Sphinx Trainer (run_decode.pl) has the file path set to utilize models with the default value (1000). After adjusting the filepath according to models created in experiment 0102, the decoder was able to start.

The following score was created during this experiment:

       ,-----------------------------------------------------------------.
      |                            hyp.trans                            |
      |-----------------------------------------------------------------|
      | SPKR    | # Snt # Wrd | Corr    Sub    Del    Ins    Err  S.Err |
      |---------+-------------+-----------------------------------------|
      | Sum/Avg |  437   6474 | 85.2   10.9    3.9   14.6   29.3   95.2 |
      |=================================================================|
      |  Mean   | 36.4  539.5 | 85.3   10.9    3.8   15.6   30.2   95.2 |
      |  S.D.   |  8.3  143.2 |  3.7    2.8    1.6    5.6    6.8    5.1 |
      | Median  | 32.5  546.5 | 84.7   11.7    3.8   16.3   31.1   96.3 |
      `-----------------------------------------------------------------'

This experiment sets a new record for the word error rate for this corpus.

The results of this experiment may explain why we achieved a higher word error rate utilizing the 5-hour long corpuses as compared to what we achieved using the 1 hour corpuses. The default senone value of 1000 is recommended for corpus data between 1 to 3 hours. While CMU recommends that a 5 hour long corpus utilize between 1000-2500 senomes. When we increase the training data, we need to adjust the senone value to corespond with the[chart provided by CMU].