xlsr-he-adap-de

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the common_voice_17_0 dataset. It achieves the following results on the evaluation set:

• Loss: 1.1481
• Wer: 0.5333
• Cer: 0.1968

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: 16
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 50
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
3.697	0.8368	100	3.7808	1.0	1.0
3.1842	1.6736	200	3.5119	1.0	1.0
3.3472	2.5105	300	3.4405	1.0	1.0
1.6876	3.3473	400	2.0791	0.9664	0.5128
1.1828	4.1841	500	1.5367	0.8851	0.3903
0.9151	5.0209	600	1.2217	0.8210	0.3806
0.7117	5.8577	700	1.0726	0.7824	0.3500
0.809	6.6946	800	1.1018	0.8094	0.3451
0.9377	7.5314	900	0.9955	0.7438	0.3255
0.5836	8.3682	1000	0.9658	0.7605	0.3209
0.5226	9.2050	1100	0.9701	0.7316	0.3125
0.4732	10.0418	1200	0.9576	0.7636	0.2993
0.5439	10.8787	1300	0.9689	0.7743	0.2976
0.3479	11.7155	1400	1.0207	0.7026	0.2813
0.4111	12.5523	1500	1.0051	0.6873	0.2725
0.2865	13.3891	1600	0.9566	0.7087	0.2716
0.3942	14.2259	1700	1.0009	0.6929	0.2730
0.3058	15.0628	1800	0.9195	0.6695	0.2583
0.2141	15.8996	1900	0.9707	0.6523	0.2532
0.4893	16.7364	2000	1.0019	0.6772	0.2548
0.2922	17.5732	2100	1.0317	0.6721	0.2645
0.3056	18.4100	2200	1.0440	0.6385	0.2595
0.3616	19.2469	2300	1.1057	0.6406	0.2516
0.271	20.0837	2400	1.1302	0.6411	0.2532
0.2183	20.9205	2500	1.2060	0.6050	0.2513
0.3128	21.7573	2600	1.1261	0.6436	0.2522
0.1602	22.5941	2700	1.1014	0.6141	0.2394
0.2255	23.4310	2800	1.2601	0.6009	0.2480
0.3142	24.2678	2900	1.0729	0.6151	0.2410
0.1815	25.1046	3000	1.0396	0.6111	0.2314
0.2507	25.9414	3100	1.1343	0.5760	0.2236
0.151	26.7782	3200	1.1477	0.6263	0.2382
0.1531	27.6151	3300	1.0935	0.5984	0.2281
0.1943	28.4519	3400	1.0250	0.5689	0.2150
0.2592	29.2887	3500	1.0309	0.5780	0.2115
0.2394	30.1255	3600	1.0363	0.5735	0.2176
0.2146	30.9623	3700	1.0521	0.5582	0.2098
0.1629	31.7992	3800	1.0586	0.5816	0.2116
0.099	32.6360	3900	1.0348	0.5643	0.2100
0.1748	33.4728	4000	1.0983	0.5841	0.2147
0.1143	34.3096	4100	1.0979	0.5567	0.2059
0.1364	35.1464	4200	1.1404	0.5663	0.2094
0.1552	35.9833	4300	1.0805	0.5628	0.2085
0.1121	36.8201	4400	1.1262	0.5628	0.2061
0.1051	37.6569	4500	1.1390	0.5425	0.2059
0.1384	38.4937	4600	1.1252	0.5394	0.2016
0.1268	39.3305	4700	1.1607	0.5552	0.2068
0.1233	40.1674	4800	1.1776	0.5618	0.2072
0.2489	41.0042	4900	1.1335	0.5399	0.1977
0.1468	41.8410	5000	1.1419	0.5404	0.1964
0.1148	42.6778	5100	1.1404	0.5455	0.2008
0.1415	43.5146	5200	1.1149	0.5425	0.2005
0.1358	44.3515	5300	1.1354	0.5430	0.2013
0.1231	45.1883	5400	1.1457	0.5374	0.1999
0.0898	46.0251	5500	1.1218	0.5343	0.1989
0.1271	46.8619	5600	1.1404	0.5353	0.1977
0.1467	47.6987	5700	1.1765	0.5318	0.1961
0.1757	48.5356	5800	1.1517	0.5292	0.1973
0.1471	49.3724	5900	1.1481	0.5333	0.1968

Framework versions

• Transformers 4.42.0.dev0
• Pytorch 2.3.1+cu121
• Datasets 2.19.2
• Tokenizers 0.19.1