Replies: 2 comments
-
Hi @lhanzl, |
Beta Was this translation helpful? Give feedback.
-
Hi @mravanelli, I have an additional question about this topic. lhanzl gives the example with output values ranging from "-0.2018069326877594" to "0.5425097942352295". The SpeechBrain documentation says that the output of the "classify_file" method is the "log posterior probabilities of each class ([batch, N_class])" and that the "score" "is the value of the log-posterior for the best class ([batch,])". However, as far as I know, log values that correspond to posterior probabilities should only have values smaller or equal to 0. Am I missing something here, or is there just a mistake in the documentation? What sort of thing are really the out_prob and score outputs of classify_files? |
Beta Was this translation helpful? Give feedback.
-
Hi, I am using ecapa-tdnn for classification tasks and found that when making predictions, the difference in the scores of the output nodes is very small. details as follows:
The above output comes from the officially released language recognition model.
The above output is a gender classification model trained by myself, using CommonLanguage recipe.
The score difference of the output nodes is very small, especially in the two-class model. But my gender classification model has a good accuracy rate.
Thank you very much.
Beta Was this translation helpful? Give feedback.
All reactions