README.md

metadata

language:
  - sv-SE
license: cc0-1.0
tags:
  - automatic-speech-recognition
  - mozilla-foundation/common_voice_8_0
  - generated_from_trainer
  - sv
  - robust-speech-event
  - model_for_talk
datasets:
  - mozilla-foundation/common_voice_8_0
  - marinone94/nst_sv
model-index:
  - name: XLS-R-300M - Swedish
    results:
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: mozilla-foundation/common_voice_8_0
          type: mozilla-foundation/common_voice_8_0
          args: sv-SE
        metrics:
          - name: Test WER
            type: wer
            value: 12.68
          - name: Test CER
            type: cer
            value: 3.79
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: speech-recognition-community-v2/dev_data
          type: speech-recognition-community-v2/dev_data
          args: sv
        metrics:
          - name: Test WER
            type: wer
            value: 27.55
          - name: Test CER
            type: cer
            value: 9.79

This model is a fine-tuned version of KBLab/wav2vec2-large-voxrex on 2 epochs of the MARINONE94/NST_SV - SV dataset (80% random split with seed 42 as the dataset for now has only the "train" split), and then on 50 epochs of the the MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - SV-SE dataset ("train+validation" split). See run.sh to have a complete overview of all the training steps. NOTE: the first training for now didn't work as expected, so it might be useless or even degrade performance. Further investigation and development is needed.

It achieves the following results on the MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - SV-SE "test" set, without any language model:

• Loss: 0.1497
• Wer: 0.1261

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: 0.00025
• train_batch_size: 32
• eval_batch_size: 32
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 128
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 2000
• num_epochs: 50.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
3.3533	1.1	100	3.2807	1.0
3.1709	2.2	200	3.1325	1.0
3.0573	3.3	300	3.0615	1.0
3.0314	4.39	400	3.0990	1.0
3.0129	5.49	500	3.0400	1.0
2.9964	6.59	600	2.9990	1.0
2.9602	7.69	700	2.9620	1.0
2.8756	8.79	800	2.7302	1.0
2.2931	9.89	900	1.5058	0.9776
1.8427	10.98	1000	0.9155	0.7832
1.4286	12.09	1100	0.4075	0.3796
1.2229	13.19	1200	0.2893	0.2652
1.1106	14.28	1300	0.2469	0.2254
1.0663	15.38	1400	0.2219	0.1973
1.0667	16.48	1500	0.2129	0.1894
1.0193	17.58	1600	0.1991	0.1789
0.9816	18.68	1700	0.1940	0.1801
0.9814	19.78	1800	0.1860	0.1667
0.9787	20.87	1900	0.1888	0.1642
0.9699	21.97	2000	0.1875	0.1704
0.9616	23.08	2100	0.1802	0.1617
0.9378	24.17	2200	0.1793	0.1577
0.888	25.27	2300	0.1764	0.1545
0.8942	26.37	2400	0.1674	0.1492
0.8701	27.47	2500	0.1739	0.1512
0.8555	28.57	2600	0.1690	0.1446
0.8513	29.67	2700	0.1649	0.1477
0.8659	30.77	2800	0.1637	0.1422
0.8419	31.86	2900	0.1614	0.1397
0.8491	32.96	3000	0.1595	0.1401
0.8395	34.07	3100	0.1607	0.1376
0.83	35.16	3200	0.1538	0.1379
0.7835	36.26	3300	0.1602	0.1408
0.7703	37.36	3400	0.1601	0.1369
0.7474	38.46	3500	0.1514	0.1342
0.7719	39.56	3600	0.1593	0.1353
0.7638	40.66	3700	0.1536	0.1338
0.771	41.75	3800	0.1531	0.1317
0.7594	42.85	3900	0.1498	0.1288
0.7383	43.95	4000	0.1527	0.1300
0.7565	45.05	4100	0.1482	0.1289
0.7697	46.15	4200	0.1495	0.1272
0.7194	47.25	4300	0.1493	0.1269
0.7479	48.35	4400	0.1490	0.1276
0.7132	49.45	4500	0.1501	0.1265

Framework versions

• Transformers 4.17.0.dev0
• Pytorch 1.10.2+cu102
• Datasets 1.18.3
• Tokenizers 0.11.0