README.md

metadata

license: apache-2.0
tags:
  - generated_from_trainer
  - hf-asr-leaderboard
  - hf-asr-leaderboard
datasets:
  - librispeech_asr
model-index:
  - name: hubert-base-libri-clean-ft100h-v3
    results:
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: LibriSpeech
          type: librispeech_asr
          config: clean
          split: test
          args:
            language: en
        metrics:
          - name: Test WER
            type: wer
            value: '8.1938'
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: LibriSpeech
          type: librispeech_asr
          config: other
          split: test
          args:
            language: en
        metrics:
          - name: Test WER
            type: wer
            value: '16.9783'
language:
  - en

hubert-base-libri-clean-ft100h-v3

This model is a fine-tuned version of facebook/hubert-base-ls960 on the librispeech_asr dataset. It achieves the following results on the evaluation set:

• Loss: 0.1120
• Wer: 0.1332

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: 5e-05
• train_batch_size: 8
• eval_batch_size: 16
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 16
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 600
• num_epochs: 8
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
5.201	0.14	250	3.9799	1.0
2.8893	0.28	500	3.4838	1.0
2.8603	0.42	750	3.3505	1.0
2.7216	0.56	1000	2.1194	0.9989
1.3372	0.7	1250	0.8124	0.6574
0.8238	0.84	1500	0.5712	0.5257
0.6449	0.98	1750	0.4442	0.4428
0.5241	1.12	2000	0.3442	0.3672
0.4458	1.26	2250	0.2850	0.3186
0.3959	1.4	2500	0.2507	0.2882
0.3641	1.54	2750	0.2257	0.2637
0.3307	1.68	3000	0.2044	0.2434
0.2996	1.82	3250	0.1969	0.2313
0.2794	1.96	3500	0.1823	0.2193
0.2596	2.1	3750	0.1717	0.2096
0.2563	2.24	4000	0.1653	0.2000
0.2532	2.38	4250	0.1615	0.1971
0.2376	2.52	4500	0.1559	0.1916
0.2341	2.66	4750	0.1494	0.1855
0.2102	2.8	5000	0.1464	0.1781
0.2222	2.94	5250	0.1399	0.1732
0.2081	3.08	5500	0.1450	0.1707
0.1963	3.22	5750	0.1337	0.1655
0.2107	3.36	6000	0.1344	0.1633
0.1866	3.5	6250	0.1339	0.1611
0.186	3.64	6500	0.1311	0.1563
0.1703	3.78	6750	0.1307	0.1537
0.1819	3.92	7000	0.1277	0.1555
0.176	4.06	7250	0.1280	0.1515
0.1837	4.2	7500	0.1249	0.1504
0.1678	4.34	7750	0.1236	0.1480
0.1624	4.48	8000	0.1194	0.1456
0.1631	4.62	8250	0.1215	0.1462
0.1736	4.76	8500	0.1192	0.1451
0.1752	4.9	8750	0.1206	0.1432
0.1578	5.04	9000	0.1151	0.1415
0.1537	5.18	9250	0.1185	0.1402
0.1771	5.33	9500	0.1165	0.1414
0.1481	5.47	9750	0.1152	0.1413
0.1509	5.61	10000	0.1152	0.1382
0.146	5.75	10250	0.1133	0.1385
0.1464	5.89	10500	0.1139	0.1371
0.1442	6.03	10750	0.1162	0.1365
0.128	6.17	11000	0.1147	0.1371
0.1381	6.31	11250	0.1148	0.1378
0.1343	6.45	11500	0.1113	0.1363
0.1325	6.59	11750	0.1134	0.1355
0.1442	6.73	12000	0.1142	0.1358
0.1286	6.87	12250	0.1133	0.1352
0.1349	7.01	12500	0.1129	0.1344
0.1338	7.15	12750	0.1131	0.1328
0.1403	7.29	13000	0.1124	0.1338
0.1314	7.43	13250	0.1141	0.1335
0.1283	7.57	13500	0.1124	0.1332
0.1347	7.71	13750	0.1107	0.1332
0.1195	7.85	14000	0.1119	0.1332
0.1326	7.99	14250	0.1120	0.1332

Framework versions

• Transformers 4.17.0
• Pytorch 1.11.0+cu113
• Datasets 1.18.3
• Tokenizers 0.12.1