I am trying to adapt a semi-continuous acoustic model. I use map_adapt for adaptation. I have the adaptation corpus in several batches. So I am performing the adaptation process in iterations as follows:
I perform adaptation on acoustic model using first batch of audio data. The newly created acoustic model has good accuracy.
Now I use the newly created model as the base model for adaption of second batch of audio. The newly created acoustic model is again used for adaptation of third batch of data and so on...
After a while, I noticed that while I was getting good accuracy for first batch of audio data at the beginning, the accuracy has considerably reduced with the newest acoustic model developed after some iterations.
Is it because of doing the adaptation in iterations or something else? If yes, how can I solve that problem? The adaptation corpus I have is in several batches and it would be infeasible for me to combine it in one single batch for adaptation.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
This is expected given you understand what adaptation is doing. It learns to recognize only the last batch and forgets the first batch.
I am not sure what do you mean about "infeasible", modern systems could be easily trained on thousands hours of data, you do not need more. Maybe you are trying to solve some different problem. In that case to get help you need to provide more details on what problem are you going to solve and what data do you have availlable.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I am trying to adapt a semi-continuous acoustic model. I use map_adapt for adaptation. I have the adaptation corpus in several batches. So I am performing the adaptation process in iterations as follows:
I perform adaptation on acoustic model using first batch of audio data. The newly created acoustic model has good accuracy.
Now I use the newly created model as the base model for adaption of second batch of audio. The newly created acoustic model is again used for adaptation of third batch of data and so on...
After a while, I noticed that while I was getting good accuracy for first batch of audio data at the beginning, the accuracy has considerably reduced with the newest acoustic model developed after some iterations.
Is it because of doing the adaptation in iterations or something else? If yes, how can I solve that problem? The adaptation corpus I have is in several batches and it would be infeasible for me to combine it in one single batch for adaptation.
This is expected given you understand what adaptation is doing. It learns to recognize only the last batch and forgets the first batch.
I am not sure what do you mean about "infeasible", modern systems could be easily trained on thousands hours of data, you do not need more. Maybe you are trying to solve some different problem. In that case to get help you need to provide more details on what problem are you going to solve and what data do you have availlable.