nb-asr-north-saami-parakeet-v6-tokenremap-v5

Fine-tuned NeMo Parakeet TDT model for North Saami ASR.

Run Info

• Run name: parakeet_v6_full_v5_lr1e-3_20260327_013645
• Model artifact: /home/javierr/git/nb-asr-parakeet/outputs/parakeet_v6_full_v5_lr1e-3_20260327_013645/artifacts/model.nemo
• Weights & Biases run: https://wandb.ai/versae/nbailab/runs/2026-03-27_01-36-48

Validation Metrics During Training

step	val_loss	val_wer
500	1.265893	0.561911
1000	0.975794	0.476518
1500	0.813457	0.405448
2001	0.801883	0.401064
2501	0.815664	0.383845
3001	0.807974	0.355980
3502	0.859701	0.358172
4002	0.848246	0.345022
4502	0.908038	0.340952
5003	0.975302	0.354728
5503	0.961213	0.345022
6003	1.013317	0.340326
6504	1.028613	0.350344
7004	1.080084	0.338760
7504	1.163356	0.351597
8004	1.132094	0.340326
8505	1.175803	0.341891
9005	1.175904	0.342830
9505	1.192001	0.336882
10006	1.227019	0.329994
10506	1.244950	0.329681
11006	1.235673	0.327176
11507	1.244888	0.323106
12007	1.310125	0.322793
12507	1.301051	0.317783
13008	1.347271	0.323106
13508	1.329711	0.320601
14008	1.350736	0.328428
14509	1.354969	0.320288
15009	1.356913	0.320288
15509	1.368101	0.316844
16009	1.378582	0.318723

• Best during training: step 15509, val_loss 1.368101, val_wer 0.316844
• Final during training: step 16009, val_loss 1.378582, val_wer 0.318723

Final Evaluation Metrics

split	loss	wer
validation	1.370026	0.314026
test	2.539732	0.486678

Notes

• Validation/test metrics above are computed post-training from the exported .nemo model.
• Training validation table is parsed from logged Validation metrics logged to wandb entries.