Edit model card

Whisper Small Sr Yodas

This model is a fine-tuned version of openai/whisper-small on merged datasets Common Voice 16 + Fleurs + Juzne vesti (South news) + LBM + (Yodas)[https://huggingface.co/datasets/espnet/yodas] dataset and

Rupnik, Peter and Ljubešić, Nikola, 2022,
ASR training dataset for Serbian JuzneVesti-SR v1.0, Slovenian language resource repository CLARIN.SI, ISSN 2820-4042,
http://hdl.handle.net/11356/1679.

It achieves the following results on the evaluation set:

  • Loss: 0.3584
  • Wer Ortho: 0.2328
  • Wer: 0.1220

Model description

Added new dataset Yodas as test and experiment to improve results.

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: 16
  • 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: 50
  • num_epochs: 10
  • mixed_precision_training: Native AMP

Training results

Training Loss Epoch Step Validation Loss Wer Ortho Wer
0.6958 0.49 1000 0.2114 0.2528 0.1563
0.5941 0.98 2000 0.1857 0.2214 0.1269
0.3985 1.46 3000 0.1729 0.2106 0.1167
0.4187 1.95 4000 0.1745 0.2120 0.1147
0.3446 2.44 5000 0.1770 0.2074 0.1139
0.2992 2.93 6000 0.1710 0.2048 0.1061
0.2074 3.42 7000 0.1887 0.2090 0.1123
0.1958 3.91 8000 0.1871 0.2136 0.1131
0.1707 4.39 9000 0.2069 0.2230 0.1126
0.1403 4.88 10000 0.2092 0.2138 0.1110
0.0871 5.37 11000 0.2345 0.2216 0.1161
0.0856 5.86 12000 0.2384 0.2281 0.1161
0.0496 6.35 13000 0.2657 0.2327 0.1211
0.0542 6.84 14000 0.2760 0.2346 0.1198
0.0274 7.32 15000 0.3024 0.2304 0.1218
0.0281 7.81 16000 0.3134 0.2357 0.1216
0.0151 8.3 17000 0.3328 0.2276 0.1188
0.0165 8.79 18000 0.3417 0.2348 0.1220
0.0094 9.28 19000 0.3545 0.2318 0.1221
0.0125 9.77 20000 0.3584 0.2328 0.1220

Framework versions

  • Transformers 4.39.3
  • Pytorch 2.0.1+cu117
  • Datasets 2.18.0
  • Tokenizers 0.15.1
Downloads last month
20
Safetensors
Model size
242M params
Tensor type
F32
·
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.

Finetuned from

Datasets used to train Sagicc/whisper-small-sr-yodas-v2

Evaluation results