I77/kyrgyz_whisper_tiny_finetuned

kyrgyz_asr

This model is a fine-tuned version of openai/whisper-tiny on the common_voice_17_0 dataset. It achieves the following results on the evaluation set:

• Loss: 0.3324
• Wer: 38.5075

Model description

This is a test fine-tuning of Whisper Tiny for the Kyrgyz language using a dataset from the Mozilla Foundation. The code is taken from this source.

Intended uses & limitations

More information needed

Training and evaluation data

mozilla-foundation/common_voice_17_0 (ky - kyrgyz)

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: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• training_steps: 4000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.59	0.4735	1000	0.5917	60.8051
0.4987	0.9470	2000	0.4195	47.8517
0.3932	1.4205	3000	0.3561	42.6685
0.3441	1.8939	4000	0.3324	38.5075

Framework versions

• Transformers 4.49.0
• Pytorch 2.5.1+cu124
• Datasets 3.3.2
• Tokenizers 0.21.0