README.md

metadata

language:
  - or
license: apache-2.0
tags:
  - automatic-speech-recognition
  - mozilla-foundation/common_voice_8_0
  - generated_from_trainer
  - or
  - robust-speech-event
  - model_for_talk
datasets:
  - common_voice
model-index:
  - name: wav2vec2-large-xls-r-300m-or-dx12
    results:
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Common Voice 8
          type: mozilla-foundation/common_voice_8_0
          args: or
        metrics:
          - name: Test WER
            type: wer
            value: 0.5947242206235012
          - name: Test CER
            type: cer
            value: 0.18272388876724327
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Robust Speech Event - Dev Data
          type: speech-recognition-community-v2/dev_data
          args: or
        metrics:
          - name: Test WER
            type: wer
            value: NA
          - name: Test CER
            type: cer
            value: NA

wav2vec2-large-xls-r-300m-or-dx12

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: 1.4638
• Wer: 0.5602

Evaluation Commands

1. To evaluate on mozilla-foundation/common_voice_8_0 with test split

python eval.py --model_id DrishtiSharma/wav2vec2-large-xls-r-300m-or-dx12 --dataset mozilla-foundation/common_voice_8_0 --config or --split test --log_outputs

2. To evaluate on speech-recognition-community-v2/dev_data

Oriya language isn't available in speech-recognition-community-v2/dev_data

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0004
• 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: 1000
• num_epochs: 200
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
13.5059	4.17	100	10.3789	1.0
4.5964	8.33	200	4.3294	1.0
3.4448	12.5	300	3.7903	1.0
3.3683	16.67	400	3.5289	1.0
2.042	20.83	500	1.1531	0.7857
0.5721	25.0	600	1.0267	0.7646
0.3274	29.17	700	1.0773	0.6938
0.2466	33.33	800	1.0323	0.6647
0.2047	37.5	900	1.1255	0.6733
0.1847	41.67	1000	1.1194	0.6515
0.1453	45.83	1100	1.1215	0.6601
0.1367	50.0	1200	1.1898	0.6627
0.1334	54.17	1300	1.3082	0.6687
0.1041	58.33	1400	1.2514	0.6177
0.1024	62.5	1500	1.2055	0.6528
0.0919	66.67	1600	1.4125	0.6369
0.074	70.83	1700	1.4006	0.6634
0.0681	75.0	1800	1.3943	0.6131
0.0709	79.17	1900	1.3545	0.6296
0.064	83.33	2000	1.2437	0.6237
0.0552	87.5	2100	1.3762	0.6190
0.056	91.67	2200	1.3763	0.6323
0.0514	95.83	2300	1.2897	0.6164
0.0409	100.0	2400	1.4257	0.6104
0.0379	104.17	2500	1.4219	0.5853
0.0367	108.33	2600	1.4361	0.6032
0.0412	112.5	2700	1.4713	0.6098
0.0353	116.67	2800	1.4132	0.6369
0.0336	120.83	2900	1.5210	0.6098
0.0302	125.0	3000	1.4686	0.5939
0.0398	129.17	3100	1.5456	0.6204
0.0291	133.33	3200	1.4111	0.5827
0.0247	137.5	3300	1.3866	0.6151
0.0196	141.67	3400	1.4513	0.5880
0.0218	145.83	3500	1.5100	0.5899
0.0196	150.0	3600	1.4936	0.5999
0.0164	154.17	3700	1.5012	0.5701
0.0168	158.33	3800	1.5601	0.5919
0.0151	162.5	3900	1.4891	0.5761
0.0137	166.67	4000	1.4839	0.5800
0.0143	170.83	4100	1.4826	0.5754
0.0114	175.0	4200	1.4950	0.5708
0.0092	179.17	4300	1.5008	0.5694
0.0104	183.33	4400	1.4774	0.5728
0.0096	187.5	4500	1.4948	0.5767
0.0105	191.67	4600	1.4557	0.5694
0.009	195.83	4700	1.4615	0.5628
0.0081	200.0	4800	1.4638	0.5602

Framework versions

• Transformers 4.16.2
• Pytorch 1.10.0+cu111
• Datasets 1.18.3
• Tokenizers 0.11.0