You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
This repository has been archived by the owner on Sep 1, 2024. It is now read-only.
Hi,
I am trying to reproduce the results of base_noise_pt_noise_ft_30h.pt by fine-tuning the pre-trained checkpoint base_noise_pt_lrs3_vox_iter5.pt.
By directly decoding base_noise_pt_noise_ft_30h.pt on the clean set, I get WER 4%. But by fine-tuning base_noise_pt_lrs3_vox_iter5.pt with the config base_noise_pt_noise_ft_30h.yaml and decoding on the clean set, I get WER around 4.6%. For decoding, I used the parameters generation.beam=20 generation.lenpen=1. For fine-tuning, I used update_freq: [8] since I only used on GPU, and I used musan/tsv/all as the noise wav. The rest of the parameters have not been modified.
I also noticed that the performance gap between the provided fine-tuned model and the one that I fine-tuned by myself is even more significant on the noisy testing set.
Do you know what is going on, or did I miss something?
Thanks
The text was updated successfully, but these errors were encountered:
Sign up for freeto subscribe to this conversation on GitHub.
Already have an account?
Sign in.
Hi,
I am trying to reproduce the results of
base_noise_pt_noise_ft_30h.pt
by fine-tuning the pre-trained checkpointbase_noise_pt_lrs3_vox_iter5.pt
.By directly decoding
base_noise_pt_noise_ft_30h.pt
on the clean set, I get WER 4%. But by fine-tuning base_noise_pt_lrs3_vox_iter5.pt with the configbase_noise_pt_noise_ft_30h.yaml
and decoding on the clean set, I get WER around 4.6%. For decoding, I used the parametersgeneration.beam=20 generation.lenpen=1
. For fine-tuning, I usedupdate_freq: [8]
since I only used on GPU, and I usedmusan/tsv/all
as the noise wav. The rest of the parameters have not been modified.I also noticed that the performance gap between the provided fine-tuned model and the one that I fine-tuned by myself is even more significant on the noisy testing set.
Do you know what is going on, or did I miss something?
Thanks
The text was updated successfully, but these errors were encountered: