Whisper Base Slo Artur - Full FT

This model is a fine-tuned Slovenian version of openai/whisper-base on the Artur 1.0 Full Dataset.
It was fine-tuned on Vega supercomputer.
Best model at step 32000 achieves the following results on the test set:

• Loss: 0.1458
• Wer: 11.38

Model description

Example usage:

from transformers import WhisperForConditionalGeneration, WhisperProcessor
import whisper

processor = WhisperProcessor.from_pretrained("blko/whisper-base-sl-artur-full-ft")

# load our base model at step 32000
model = WhisperForConditionalGeneration.from_pretrained(
    "blko/whisper-base-sl-artur-full-ft",
    revision="772cbcea0383a8f4359d3bd8457aa63ca881c47b"
    )

audio_path = "./test_recording.wav"    # 16kHz sample rate required
mel_from_file = whisper.audio.log_mel_spectrogram(audio_path).unsqueeze(dim=0)
output = model.generate(mel_from_file)
print(processor.batch_decode(output))

Training and evaluation data

Pre-processed dataset used in this work can be found on HuggingFace at link Artur 1.0

Training procedure

For details including training and testing code refer to GitHub link at UHO: Slovenian speech recognition in an application for people with hearing impairment

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2.5e-05
• train_batch_size: 32
• eval_batch_size: 16
• 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: 51000
• mixed_precision_training: Native AMP

Framework versions

• Transformers 4.44.2
• Pytorch 2.1.0+cu121
• Datasets 2.21.0
• Tokenizers 0.19.1