Speech:Exps 0312

From Openitware
Jump to: navigation, search

Description

Author: Hannah Yudkin

Date: 6-27-2018

Purpose: Root experiment for fixing LDA

Details: The purpose of these experiments was to determine an optimal starting point for an LDA experiment. This was done by choosing a different starting configuration by utilizing a different seed each time and seeing which provided the lowest WER. The seed was chosen randomly and was printed on screen during this process. Experiment 000 was used as a template and within that the mllt.py file was edited to allow a different starting configuration based on the specific seed used. We then ran 20 experiments on 5 hour switchboard corpus to train and then test-on-train on a subset.

Results: After testing 20 different random seeds, Experiment 006 with random number 2038113770 for a 5 hour corpus with 1 1/4 hour (first 1000 utterances) test set was the best result. Below is the scoring log for Exp 006.

    ,-----------------------------------------------------------------.
    |                            hyp.trans                            |
    |-----------------------------------------------------------------|
    | SPKR    | # Snt # Wrd | Corr    Sub    Del    Ins    Err  S.Err |
    |=================================================================|
    | Sum/Avg | 1000  14889 | 75.8   11.4   12.8    2.1   26.3   76.8 |
    |=================================================================|
    |  Mean   |  1.5   22.6 | 79.1   10.9   10.0    5.5   26.4   76.9 |
    |  S.D.   |  0.7   19.5 | 17.2   11.2   11.0   16.3   21.5   37.5 |
    | Median  |  1.0   16.0 | 80.0    8.7    7.8    0.0   25.0  100.0 |
    `-----------------------------------------------------------------'