Speech:Exps 0091

From Openitware
Jump to: navigation, search

Last_5hr Test on Train (0089) using the existing tools for LM creation.


Description

Author:Team B&C

Date: 4/23/2013

Purpose: This experiment was designed to test the newly created acoustic model from 0089 and to test using a Language model derived the existing tools.

Details: The previous versions of the genTrans script merely removed markers which indicated transcribers' notes, leaving whatever was in those markers intact. This resulted in inaccurate models as the trainer was attempting to align and train phones listed in the transcript but did not exist in the audio file.

As the existing processes for creating the transcript used for the Language model creation suffered from the same issues described above, the language models it created were likely inaccurate as well as the grammar structures it was analyzing was not representative of normal English.

The transcript used in this experiment was derived from the traditional processes; namely, the ParseTranscript.perl script.

In otherwords, unlike the previous experiment, the Language model creation, decode, and scoring steps were not deviated from.

Results The experiment was successful in that the all steps were completed. The following score was produced:



                       SYSTEM SUMMARY PERCENTAGES by SPEAKER

      ,-----------------------------------------------------------------.
      |                               0091                              |
      |-----------------------------------------------------------------|
      | SPKR    | # Snt # Wrd | Corr    Sub    Del    Ins    Err  S.Err |
      |=================================================================|
      | Sum/Avg |  437   6474 | 81.5   14.1    4.4   13.9   32.5   96.1 |
      |=================================================================|
      |  Mean   | 36.4  539.5 | 81.6   14.1    4.3   15.0   33.4   96.3 |
      |  S.D.   |  8.3  143.2 |  4.6    3.5    1.8    5.5    7.3    4.3 |
      | Median  | 32.5  546.5 | 82.3   14.2    4.3   14.3   35.5   97.7 |
      `-----------------------------------------------------------------'

This experiment is a continuation of Experiment 0090 where we performed a similar decode and score with a Language model created using a genTrans5-derived transcript.

To compare between the various scores created during this set of experiments and the previous best-scoring experiment (0075). Please reference the following chart:



      ,-------------------------------------------------------------------------------------.
      |-------------------------------------------------------------------------------------|
      |        SPKR                 | # Snt # Wrd | Corr    Sub    Del    Ins    Err  S.Err |
      |=====================================================================================|
      | Sum/Avg For Experiment 0075 |  437   6569 | 79.4   15.6    5.0   12.5   33.1   96.1 |
      |=====================================================================================|
      | Sum/Avg For Experiment 0090 |  437   6474 | 82.4   13.2    4.4   14.3   31.9   96.1 |
      |=====================================================================================|
      | Sum/Avg For Experiment 0091 |  437   6474 | 81.5   14.1    4.4   13.9   32.5   96.1 |
      `-------------------------------------------------------------------------------------'

The results as compared to the first test decode and score of the acoustic model created using the Last_5hr train corpus are slightly better; however, this isn't as good as what was found in experiment 0090. Since Experiment 0090 and Experiment 0091 both used the same decoding transcript, dictionary, and acoustic model (0089), we can conclude that the Language model created in Experiment 0090 was superior than the one created in Experiment 0091.