wav2vec2-large-xls-r-romansh-colab

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the common_voice_13_0 dataset. It achieves the following results on the evaluation set:

• Loss: 0.3954
• Wer: 0.3372

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: 0.0001
• train_batch_size: 4
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 8
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 50
• num_epochs: 30
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
13.0876	0.38	50	4.9608	1.0
4.0129	0.76	100	3.1470	1.0
3.0629	1.14	150	3.0178	1.0
3.0276	1.52	200	3.0792	1.0
3.0054	1.9	250	2.9988	1.0
3.0259	2.29	300	3.0606	1.0
2.9827	2.67	350	3.0001	1.0
3.0394	3.05	400	3.0010	1.0
2.97	3.43	450	2.9920	1.0
3.0238	3.81	500	2.9967	1.0
2.9976	4.2	550	2.9906	1.0
3.0268	4.58	600	2.9893	1.0
2.9899	4.96	650	2.9904	1.0
3.0395	5.34	700	2.9889	1.0
2.9797	5.72	750	2.9912	1.0
3.002	6.11	800	2.9816	1.0
3.0043	6.49	850	2.9881	1.0
2.9599	6.87	900	2.9814	1.0
3.0148	7.25	950	2.9962	1.0
2.9611	7.63	1000	2.9739	1.0
3.0657	8.02	1050	2.9980	1.0
2.9598	8.4	1100	2.9725	1.0
2.9749	8.78	1150	2.9997	1.0
2.9801	9.16	1200	2.9577	1.0
2.9597	9.54	1250	2.9457	1.0
2.9335	9.92	1300	2.9349	1.0
2.9576	10.3	1350	2.9185	1.0
2.8992	10.68	1400	2.7701	1.0
2.7045	11.07	1450	2.2958	1.0
2.0933	11.45	1500	1.4031	0.9998
1.5523	11.83	1550	1.0655	0.9029
1.3192	12.21	1600	0.9047	0.7736
1.1209	12.59	1650	0.7879	0.6763
1.0312	12.97	1700	0.7086	0.6616
0.9216	13.36	1750	0.6601	0.6118
0.8778	13.74	1800	0.6042	0.5971
0.7868	14.12	1850	0.5748	0.5675
0.7491	14.5	1900	0.5503	0.5484
0.7181	14.88	1950	0.5365	0.5240
0.7099	15.27	2000	0.5032	0.4984
0.6294	15.65	2050	0.4871	0.4900
0.6283	16.03	2100	0.4667	0.4853
0.5798	16.41	2150	0.4702	0.4690
0.5826	16.79	2200	0.4708	0.4555
0.5622	17.17	2250	0.4682	0.4537
0.5244	17.56	2300	0.4468	0.4353
0.5177	17.94	2350	0.4523	0.4432
0.469	18.32	2400	0.4368	0.4152
0.4963	18.7	2450	0.4260	0.4094
0.4644	19.08	2500	0.4250	0.3905
0.4588	19.46	2550	0.4265	0.3952
0.4273	19.84	2600	0.4357	0.3961
0.4519	20.23	2650	0.4196	0.3819
0.4161	20.61	2700	0.4192	0.3880
0.4205	20.99	2750	0.4137	0.3761
0.3993	21.37	2800	0.4216	0.3843
0.3937	21.75	2850	0.4189	0.3798
0.3767	22.14	2900	0.4130	0.3719
0.3879	22.52	2950	0.4004	0.3619
0.385	22.9	3000	0.4112	0.3605
0.3859	23.28	3050	0.4042	0.3591
0.3743	23.66	3100	0.4197	0.3703
0.3385	24.05	3150	0.3952	0.3510
0.3405	24.43	3200	0.3935	0.3537
0.363	24.81	3250	0.3908	0.3463
0.3257	25.19	3300	0.3972	0.3421
0.3487	25.57	3350	0.3991	0.3433
0.3478	25.95	3400	0.4119	0.3563
0.3435	26.33	3450	0.3948	0.3435
0.3346	26.71	3500	0.4106	0.3449
0.3093	27.1	3550	0.4008	0.3405
0.3304	27.48	3600	0.4025	0.3416
0.3335	27.86	3650	0.3950	0.3386
0.3179	28.24	3700	0.3924	0.3374
0.3141	28.62	3750	0.3928	0.3370
0.3335	29.01	3800	0.3965	0.3379
0.3198	29.39	3850	0.3949	0.3370
0.3201	29.77	3900	0.3954	0.3372

Framework versions

• Transformers 4.26.0
• Pytorch 2.0.1+cu118
• Datasets 2.14.4
• Tokenizers 0.13.3