michaelszhu/wav2vec2-base-finetuned-radio-ASR-2

BANG WAV2VEC v1 (EN)

This model is a fine-tuned version of openai/wav2vec2 on the Radio-Modified Common Voice 11.0 dataset. It achieves the following results on the evaluation set:

• Loss: 316.8398
• Wer: 95.7239

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: 1e-05
• train_batch_size: 8
• eval_batch_size: 8
• 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: 500
• training_steps: 15000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
483.2584	2.0083	1000	434.6134	100.0
405.653	4.0167	2000	344.3590	96.6701
370.4857	6.025	3000	342.5145	99.8213
369.1107	9.0042	4000	327.9243	99.9830
346.2929	11.0125	5000	318.8220	99.3436
343.8343	13.0208	6000	312.6764	98.5592
341.4097	15.0292	7000	314.1781	99.1016
353.2187	18.0083	8000	337.7428	99.9430
362.3456	20.0167	9000	312.4731	97.1724
340.6648	22.025	10000	309.5160	96.7086
346.2773	25.0042	11000	307.8313	96.0953
324.1138	27.0125	12000	314.9616	95.1754
331.969	29.0208	13000	311.7163	95.5329
325.9136	31.0292	14000	315.2805	95.6577
335.0141	34.0083	15000	316.8398	95.7239

Framework versions

• Transformers 4.41.2
• Pytorch 2.3.1+cu121
• Datasets 2.19.2
• Tokenizers 0.19.1