README.md

metadata

language:
  - tr
license: apache-2.0
tags:
  - automatic-speech-recognition
  - common_voice
  - generated_from_trainer
datasets:
  - common_voice
model-index:
  - name: ''
    results: []

This model is a fine-tuned version of facebook/wav2vec2-xls-r-1b on the COMMON_VOICE - TR dataset. It achieves the following results on the evaluation set:

• Loss: 0.7314
• Wer: 0.4692
• Cer: 0.1321

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: 0.0005
• train_batch_size: 32
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 100.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
1.0779	4.59	500	0.8260	0.7395	0.2354
0.7573	9.17	1000	0.7544	0.6960	0.2100
0.8225	13.76	1500	0.6867	0.6672	0.2021
0.621	18.35	2000	0.6824	0.6209	0.1874
0.6362	22.94	2500	0.6712	0.6286	0.1904
0.624	27.52	3000	0.6940	0.6116	0.1820
0.4781	32.11	3500	0.6966	0.5989	0.1735
0.5685	36.7	4000	0.6742	0.5971	0.1769
0.4384	41.28	4500	0.6904	0.5999	0.1767
0.5509	45.87	5000	0.6734	0.5641	0.1692
0.3665	50.46	5500	0.7018	0.5662	0.1680
0.3914	55.05	6000	0.7121	0.5552	0.1631
0.2467	59.63	6500	0.6657	0.5374	0.1563
0.2576	64.22	7000	0.6920	0.5316	0.1554
0.2711	68.81	7500	0.6900	0.5176	0.1495
0.2626	73.39	8000	0.6843	0.5043	0.1454
0.1377	77.98	8500	0.7383	0.5101	0.1470
0.2005	82.57	9000	0.7228	0.5045	0.1430
0.1355	87.16	9500	0.7231	0.4869	0.1375
0.0431	91.74	10000	0.7397	0.4749	0.1350
0.0586	96.33	10500	0.7360	0.4754	0.1339

Framework versions

• Transformers 4.17.0.dev0
• Pytorch 1.10.2+cu102
• Datasets 1.18.2.dev0
• Tokenizers 0.11.0