Replies: 3 comments
-
|
Hi and thanks for post. So I think the performance of the system depends on many factors such as:
With C=3072 and 4 GTX 2080Ti, I reached 0.80% EER (with s-norm) as reported. Therefore, I'd like to learn a bit more about your related configurations. |
Beta Was this translation helpful? Give feedback.
-
|
My configuration used C=1024 and an RTX 3090. While the achieved EER of 1.03% is better than what you reported, it still falls short of the figures mentioned in the papers (0.87% and 0.856%). I'm curious as to why this discrepancy exists. I'll need to investigate this further. To start, I plan on training the model for more epochs. |
Beta Was this translation helpful? Give feedback.
-
|
I have a similar issue. I was trying to use the default setup (C=1024) to reproduce the results in the paper. According to the paper, model (C=1024) should have 14.7M parameters, but the speechbrain setup comes ~20M parameteres. Any update on this @underdogliu? |
Beta Was this translation helpful? Give feedback.
Uh oh!
There was an error while loading. Please reload this page.
-
Hello,
In the papers 1 and 2, the ECAPA-TDNN (C=1024) model, when trained on VoxCeleb2 and tested on VoxCeleb1-O, reported EERs of 0.87% and 0.856%, respectively.
Both papers indicate that the model consists of 14.7M parameters.
In the speaker verification recipe from SpeechBrain, the channels are defined as [1024, 1024, 1024, 1024, 3072] in the YAML file.
This setup, however, results in 20.8M parameters.
By adjusting the channels to [1024, 1024, 1024, 1024, 1536], the parameter count becomes 14.7M, consistent with the papers.
Thus, it appears the latter configuration is the accurate one.
However, even with this configuration and using the same training and test datasets (trained 10 epochs), the EER reached is 1.03%.
Why does this performance deviate from the results mentioned in the papers?
I would greatly appreciate direct assistance or a referral to someone who could provide insight.
@mravanelli
Beta Was this translation helpful? Give feedback.
All reactions