Speech:Exps 0110

From Openitware
Jump to: navigation, search

Modified Senone Value (4000): Test on Train 0110


Description

Author: Eric Beikman

Date: 6/20/2013

Purpose: To run a test on train for Experiment 0107.

Details: For this Experiment, we utilized the 30 minute last_5hr/test corpus to test train 0107. In Experiment 0107, we increased the amount of senones to 4000 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. As such, this also allows us to compare the generated score with Experiment 0104, which had also had a modified senone value. Results

The following score was created during this experiment:

       ,-----------------------------------------------------------------.
      |                            hyp.trans                            |
      |-----------------------------------------------------------------|
      | SPKR    | # Snt # Wrd | Corr    Sub    Del    Ins    Err  S.Err |
      |---------+-------------+-----------------------------------------|
      | Sum/Avg |  437   6474 | 89.2    7.6    3.2   14.4   25.2   92.2 |
      |=================================================================|
      |  Mean   | 36.4  539.5 | 89.0    7.8    3.3   15.6   26.6   92.3 |
      |  S.D.   |  8.3  143.2 |  2.9    2.2    1.2    6.3    7.5    6.3 |
      | Median  | 32.5  546.5 | 88.8    8.0    3.4   15.2   26.4   95.6 |
      `-----------------------------------------------------------------'

This experiment sets a new record for the word error rate for this corpus. By slightly increasing the senone value (by 1000), we were able to get about a small improvement in our word error rate; however, this improvement is smaller than previous jumps.

Experiments 0104, and 0108 through 0110 prove that adjusting the senone rates will improve the word error rate. But how much it will improve the word error rate depends on the size of the corpus.

SenoneGraph.png

According to the graph above, for a 5 hour corpus size, the improvement is substantial until a senone value of 2500, after which the word error rate improvement per 500 seones drops. It can be concluded that for a 5 hour corpus, a senone value of 2500 is optimal, anything more than that will result in more overhead in training and decoding without a reasonable improvement in word errors.