whisper-small-uz-en-ru-lang-id

This model is a fine-tuned version of openai/whisper-small on the "mozilla-foundation/common_voice_16_1"(uz/en/ru) dataset. It achieves the following results on the validation set during training:

• Loss: 0.2065
• Accuracy: 0.9747
• F1: 0.9746

Accuracy on the test (evaluation) dataset: 92.4%.

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

# datasets for each language from the set {uz: Uzbek, en: English, ru: Russian}
common_voice_train_uz = load_dataset("mozilla-foundation/common_voice_16_1", "uz", split='train', trust_remote_code=True, token=env('HUGGING_TOKEN'), streaming=True)
common_voice_train_ru = load_dataset("mozilla-foundation/common_voice_16_1", "ru", split='train', trust_remote_code=True, token=env('HUGGING_TOKEN'), streaming=True)
common_voice_train_en = load_dataset("mozilla-foundation/common_voice_16_1", "en", split='train', trust_remote_code=True, token=env('HUGGING_TOKEN'), streaming=True)
common_voice_valid_uz = load_dataset("mozilla-foundation/common_voice_16_1", "uz", split='validation', trust_remote_code=True, token=env('HUGGING_TOKEN'), streaming=True)
common_voice_valid_ru = load_dataset("mozilla-foundation/common_voice_16_1", "ru", split='validation', trust_remote_code=True, token=env('HUGGING_TOKEN'), streaming=True)
common_voice_valid_en = load_dataset("mozilla-foundation/common_voice_16_1", "en", split='validation', trust_remote_code=True, token=env('HUGGING_TOKEN'), streaming=True)

# code to shuffle and to take limited size of data. Rows per set: Train-24000, Validation-3000.
... 
# concatenate 3 datasets
common_voice['train'] = concatenate_datasets([common_voice_train_uz, common_voice_train_ru, common_voice_train_en])

Training procedure

Used Trainer from transformers. Training and evaluation process are described in the Jupyter notebook, storing in the following github repository:

https://github.com/fitlemon/whisper-small-uz-en-ru-lang-id

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 3e-05
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 8
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.1
• training_steps: 9000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy	F1
0.0252	1	3000	0.3089	0.953	0.9525
0.0357	2	6000	0.1732	0.964	0.9637
0.0	3	9000	0.2065	0.9747	0.9746

Framework versions

• Transformers 4.38.2
• Pytorch 2.2.1+cu121
• Datasets 2.17.1
• Tokenizers 0.15.2