search

ppwwyyxx / speaker-recognition

A Speaker Recognition System
Apache License 2.0
675 stars 276 forks source link

Unpredictable results. #51

Closed rhythmize closed 6 years ago

rhythmize commented 7 years ago

Over the same dataset I trained on different machines, the misclassified samples are different. Although few of the all data samples are misclassified over the given dataset, the results varies from machine to machine. Test samples properly identified on one machine are mislabeled on the other. Do you have any idea about this?

ppwwyyxx commented 7 years ago

The algorithms have randomness inherently.

rhythmize commented 7 years ago

So, there's no way the results can be certain ?

ppwwyyxx commented 7 years ago

right.

richardm47 commented 6 years ago

@code-player by mislabeled you mean the results are incorrect & the matching is not showing correct results ?

rhythmize commented 6 years ago

Yeah. Definitely. Over the same dataset (say A-Z) trained over two different machines. One machine misidentifies some users (let say D, M, P) and perfectly identifies the rest, the other machine misidentifies some other users, (like F, O, X) and identifies the rest properly. This was the issue I asked about.