wav2vec2-large-xls-r-300m-japanese-colab

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

• Loss: 0.8060
• Wer: 0.1393

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: 100
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
No log	2.44	100	3.7830	1.0
No log	4.88	200	2.9520	0.9999
No log	7.32	300	1.1938	0.2940
3.7558	9.76	400	0.7278	0.1977
3.7558	12.2	500	0.6456	0.1668
3.7558	14.63	600	0.6702	0.1530
3.7558	17.07	700	0.7131	0.1568
0.2503	19.51	800	0.7277	0.1488
0.2503	21.95	900	0.7558	0.1630
0.2503	24.39	1000	0.7611	0.1437
0.2503	26.83	1100	0.7501	0.1426
0.1316	29.27	1200	0.7635	0.1445
0.1316	31.71	1300	0.8348	0.1578
0.1316	34.15	1400	0.7285	0.1545
0.1316	36.59	1500	0.7949	0.1491
0.0974	39.02	1600	0.7706	0.1524
0.0974	41.46	1700	0.8180	0.1432
0.0974	43.9	1800	0.7718	0.1281
0.0974	46.34	1900	0.7915	0.1315
0.0731	48.78	2000	0.7905	0.1337
0.0731	51.22	2100	0.8401	0.1340
0.0731	53.66	2200	0.7810	0.1410
0.0731	56.1	2300	0.8034	0.1418
0.0569	58.54	2400	0.8219	0.1472
0.0569	60.98	2500	0.7661	0.1432
0.0569	63.41	2600	0.7989	0.1442
0.0569	65.85	2700	0.8212	0.1440
0.0456	68.29	2800	0.8029	0.1395
0.0456	70.73	2900	0.8113	0.1425
0.0456	73.17	3000	0.8298	0.1434
0.0456	75.61	3100	0.8131	0.1403
0.0343	78.05	3200	0.8313	0.1415
0.0343	80.49	3300	0.8395	0.1434
0.0343	82.93	3400	0.8048	0.1386
0.0343	85.37	3500	0.8126	0.1393
0.026	87.8	3600	0.7933	0.1378
0.026	90.24	3700	0.8317	0.1389
0.026	92.68	3800	0.8005	0.1378
0.026	95.12	3900	0.8059	0.1385
0.0204	97.56	4000	0.8071	0.1389
0.0204	100.0	4100	0.8060	0.1393

Framework versions

• Transformers 4.17.0
• Pytorch 1.10.0+cu113
• Datasets 1.18.3
• Tokenizers 0.12.1