Hungarians/whisper-small-cv16-hu-v1.5

Whisper Small Hungarian (training in progress)

This model is a fine-tuned version of openai/whisper-small on the Common Voice 16 dataset of Mozilla Foundation. It achieves the following results on the evaluation set:

Tempolary at step 11000:

• Wer: 8.4969

Unfortunatly the colab disconected, this is the end... :( maybe later continue

My own hungarian language specific compare test result (on CV11):

Modell	WER	CER	NORMALISED WER	NORMALISED CER
openai/whisper-tiny	112.1	51.33	108.79	49.64
openai/whisper-base	95.87	42.84	95.68	41.38
openai/whisper-small	53.65	15.89	49.8	14.63
Hungarians/whisper-tiny-cv16-hu	30.57	8.52	27.71	7.86
Hungarians/whisper-tiny-cv16-hu-v2	16.99	4.98	15.27	4.49
Hungarians/whisper-base-cv16-hu	15.55	4.07	13.68	3.67
Hungarians/whisper-base-cv16-hu-v2	12.63	3.55	11.39	3.26
Hungarians/whisper-small-cv16-hu	17.86	4.1	15.27	3.58
sarpba/whisper-small-cv16-v1.5-hu	9.94	2.41	8.50	2.14

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: 1.25e-05
• train_batch_size: 8
• eval_batch_size: 4
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: constant_with_warmup
• lr_scheduler_warmup_steps: 400
• planed training_steps: 15000
• executed steps: 11000 only (colab dc)
• mixed_precision_training: Native AMP

Training results

Steps	Training Loss	Validation Loss	Wer Ortho	Wer

Framework versions

• Transformers 4.36.2
• Pytorch 2.1.0+cu121
• Datasets 2.16.0
• Tokenizers 0.15.0