Whisper Small Uzbek

This model is a fine-tuned version of openai/whisper-small trained and evaluated on the mozilla-foundation/common_voice_11_0 uz and google/fleurs uz_uz datasets.

It achieves the following results on the common_voice_11_0 evaluation set:

• Loss: 0.3872
• Wer: 23.6509

It achieves the following results on the FLEURS evaluation set:

• Wer: 47.15

Model description

This model was created as part of the Whisper fine-tune sprint event.

Based on eval, this model achieves a WER of 23.6509 against the Common Voice 11 dataset and 47.15 against the FLEURS dataset.

This is a significant improvement over the smallest reported WER of 90.2 for the Uzbek language recorded on the Whisper article:

Intended uses & limitations

More information needed

Training and evaluation data

Training was performed using the train and evaluation splits from Mozilla's Common Voice 11 and Google's FLEURS datasets.

Testing was performed using the test splits from the same datasets.

Training procedure

Training and CV11 testing was performed using a modified version of Hugging Face's run_speech_recognition_seq2seq_streaming.py script by Michael Kamfonas.

FLEURS testing was performed using the standard run_eval_whisper_streaming.py script.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 64
• eval_batch_size: 32
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 400
• training_steps: 5000
• mixed_precision_training: Native AMP

Training results

Framework versions

• Transformers 4.26.0.dev0
• Pytorch 1.13.0+cu117
• Datasets 2.7.1.dev0
• Tokenizers 0.13.2