PLDA_LDA module Error: 'StatObject_SB' object has no attribute 'mean_stat_per_model'

JakubPex commented 9 months ago

Describe the bug

When trying to run PLDA scoring of StatObject_SB instance containing non-unique models the AttributeError: 'StatObject_SB' object has no attribute 'mean_stat_per_model' occurs.

The mean_stat_per_model() method is not implemented.

Expected behaviour

StatObject_SB validity check shall be created in PLDA_LDA.fast_PLDA_scoring() and informative message returned in case of not supported model set.
mean_stat_per_model() method shall be created.

To Reproduce

No response

Environment Details

No response

Relevant Log Output

No response

Additional Context

No response

HimaJyothi17 commented 6 months ago

Any update on this issue. Or is there way to overcome this error?

JakubPex commented 6 months ago

You can pre-compute average embedding yourself, (average pooling) and use just one embedding per label. This is how I did it. Embedding average-pooling is a common way to do it when PLDA is used. Remember to length-normalize the averaged embedding before scoring.

speechbrain / speechbrain