errors in prediction

[richardm47]

1. Trained the software on different voices.
2. Each voice sample used for training is length at least one minute long.
3. For each voice a separate .out model file is created.

Now I matched new sample files each of length 5 seconds and the predictions were very random. Sometimes they were accurate and other times they showed wrong results. Even testing voices that weren't in the database showed a close match. Can you tell me what I am doing wrong ? And whats the correct way to feed background noise and how much does it matter ?

Note: I edited the /testbench/gmmset.py file to get scores and the scores came around -5 ~ -10 for wrong matches.

[richardm47]

@ppwwyyxx ??

[ppwwyyxx]

For each voice a separate .out model file is created.

Why is this?

[richardm47]

@ppwwyyxx I need to update target speakers on a regular basis and selectively check for certain speakers only and not all the speakers. So I figured that it would be better to train one .out file for each speaker and call only the needed speaker's models while prediction. I am also getting output scores in the testbench/gmmset.py because the default system only shows the highest score and I've noticed that sometimes the other top scorers is the actual speaker. Their score is pretty high but not the highest.

What am I doing wrong ?

[richardm47]

@ppwwyyxx is that wrong ?

[ppwwyyxx]

This looks right. I don't know what part was going wrong.