xtreme_s_xlsr_300m_fleurs_asr_western_european

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the GOOGLE/XTREME_S - FLEURS.ALL dataset. It achieves the following results on the evaluation set:

• Cer: 0.2484
• Cer Ast Es: 0.1598
• Cer Bs Ba: 0.1749
• Cer Ca Es: 0.1655
• Cer Cy Gb: 0.2280
• Cer Da Dk: 0.3616
• Cer De De: 0.1287
• Cer El Gr: 0.6020
• Cer En Us: 0.1938
• Cer Es 419: 0.1288
• Cer Fi Fi: 0.2050
• Cer Fr Fr: 0.1811
• Cer Ga Ie: 0.4474
• Cer Gl Es: 0.1324
• Cer Hr Hr: 0.1555
• Cer Hu Hu: 0.3911
• Cer Is Is: 0.4646
• Cer It It: 0.1283
• Cer Kea Cv: 0.1818
• Cer Lb Lu: 0.2594
• Cer Mt Mt: 0.3628
• Cer Nb No: 0.2254
• Cer Nl Nl: 0.1790
• Cer Oci Fr: 0.2159
• Cer Pt Br: 0.2275
• Cer Sv Se: 0.3092
• Loss: 1.3089
• Loss Ast Es: 0.7715
• Loss Bs Ba: 0.7378
• Loss Ca Es: 0.7868
• Loss Cy Gb: 1.1441
• Loss Da Dk: 1.9130
• Loss De De: 0.5391
• Loss El Gr: 3.4904
• Loss En Us: 0.9632
• Loss Es 419: 0.6186
• Loss Fi Fi: 0.8953
• Loss Fr Fr: 0.9076
• Loss Ga Ie: 3.0217
• Loss Gl Es: 0.5788
• Loss Hr Hr: 0.6462
• Loss Hu Hu: 1.9029
• Loss Is Is: 2.6551
• Loss It It: 0.6052
• Loss Kea Cv: 0.9107
• Loss Lb Lu: 1.3705
• Loss Mt Mt: 2.3651
• Loss Nb No: 1.1518
• Loss Nl Nl: 0.8490
• Loss Oci Fr: 1.1421
• Loss Pt Br: 1.1641
• Loss Sv Se: 1.5910
• Wer: 0.6451
• Wer Ast Es: 0.4654
• Wer Bs Ba: 0.5443
• Wer Ca Es: 0.4979
• Wer Cy Gb: 0.5962
• Wer Da Dk: 0.8455
• Wer De De: 0.4221
• Wer El Gr: 0.9805
• Wer En Us: 0.4556
• Wer Es 419: 0.3928
• Wer Fi Fi: 0.8116
• Wer Fr Fr: 0.4690
• Wer Ga Ie: 0.8519
• Wer Gl Es: 0.4245
• Wer Hr Hr: 0.4895
• Wer Hu Hu: 0.9099
• Wer Is Is: 0.9960
• Wer It It: 0.4415
• Wer Kea Cv: 0.5202
• Wer Lb Lu: 0.7225
• Wer Mt Mt: 1.0096
• Wer Nb No: 0.6541
• Wer Nl Nl: 0.5257
• Wer Oci Fr: 0.5770
• Wer Pt Br: 0.6685
• Wer Sv Se: 0.8546
• Predict Samples: 20043

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0003
• train_batch_size: 8
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• total_train_batch_size: 64
• total_eval_batch_size: 8
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 1000
• num_epochs: 20.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
3.1411	0.49	500	3.1673	1.0	1.0
0.6397	0.97	1000	0.9039	0.7171	0.2862
0.4033	1.46	1500	0.8914	0.6862	0.2763
0.3473	1.94	2000	0.8017	0.6505	0.2536
0.3143	2.43	2500	0.8568	0.6566	0.2627
0.3004	2.91	3000	0.8898	0.6640	0.2686
0.282	3.4	3500	0.8489	0.6637	0.2571
0.2489	3.88	4000	0.8955	0.6744	0.2691
0.1706	4.37	4500	0.9190	0.6788	0.2688
0.3336	4.85	5000	0.8915	0.6594	0.2572
0.1426	5.34	5500	0.9501	0.6784	0.2686
0.2301	5.83	6000	1.0217	0.6719	0.2735
0.1325	6.31	6500	0.9578	0.6691	0.2655
0.1145	6.8	7000	0.9129	0.6680	0.2593
0.1202	7.28	7500	0.9646	0.6749	0.2619
0.143	7.77	8000	0.9200	0.6554	0.2554
0.1012	8.25	8500	0.9553	0.6787	0.2628
0.1018	8.74	9000	0.9455	0.6445	0.2511
0.1148	9.22	9500	1.0206	0.6725	0.2629
0.0794	9.71	10000	0.9305	0.6547	0.2526
0.2891	10.19	10500	1.0424	0.6709	0.2570
0.1665	10.68	11000	0.9760	0.6596	0.2507
0.1956	11.17	11500	0.9549	0.6340	0.2440
0.0828	11.65	12000	0.9598	0.6403	0.2460
0.059	12.14	12500	0.9972	0.6574	0.2531
0.0505	12.62	13000	0.9836	0.6534	0.2525
0.0336	13.11	13500	1.0619	0.6564	0.2519
0.0435	13.59	14000	1.0844	0.6480	0.2543
0.0216	14.08	14500	1.1084	0.6512	0.2521
0.0265	14.56	15000	1.1152	0.6607	0.2563
0.0975	15.05	15500	1.1060	0.6456	0.2471
0.1396	15.53	16000	1.1100	0.6337	0.2418
0.0701	16.02	16500	1.1731	0.6309	0.2415
0.1171	16.5	17000	1.1302	0.6315	0.2396
0.0778	16.99	17500	1.1485	0.6379	0.2447
0.0642	17.48	18000	1.2009	0.6400	0.2464
0.0322	17.96	18500	1.2028	0.6357	0.2425
0.031	18.45	19000	1.2381	0.6285	0.2416
0.0579	18.93	19500	1.2299	0.6265	0.2409
0.0628	19.42	20000	1.2582	0.6277	0.2395
0.074	19.9	20500	1.2572	0.6278	0.2394

Framework versions

• Transformers 4.18.0.dev0
• Pytorch 1.10.1+cu111
• Datasets 1.18.4.dev0
• Tokenizers 0.11.6