Edit model card

Whisper Small Bashkir

This model is a fine-tuned version of openai/whisper-small on the mozilla-foundation/common_voice_11_0 ba dataset. It achieves the following results on the evaluation set:

  • Loss: 0.2589
  • Wer: 15.0723

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: 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: 500
  • training_steps: 30000
  • mixed_precision_training: Native AMP

Training results

Training Loss Epoch Step Validation Loss Wer
0.1637 1.01 2000 0.2555 26.4682
0.1375 2.01 4000 0.2223 21.5394
0.0851 3.02 6000 0.2086 19.6725
0.0573 4.02 8000 0.2178 18.4280
0.036 5.03 10000 0.2312 17.8248
0.0238 6.04 12000 0.2621 17.4096
0.0733 7.04 14000 0.2120 16.5656
0.0111 8.05 16000 0.2682 16.2291
0.0155 9.05 18000 0.2677 15.9242
0.0041 10.06 20000 0.3178 15.9534
0.0023 12.01 22000 0.3218 16.0536
0.0621 13.01 24000 0.2313 15.6169
0.0022 14.02 26000 0.2887 15.1083
0.0199 15.02 28000 0.2553 15.1848
0.0083 16.03 30000 0.2589 15.0723

Framework versions

  • Transformers 4.26.0.dev0
  • Pytorch 1.13.0+cu117
  • Datasets 2.7.1.dev0
  • Tokenizers 0.13.2
Downloads last month
10
Inference Examples
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.

Dataset used to train AigizK/bashkir-whisper-small

Space using AigizK/bashkir-whisper-small 1

Evaluation results