zalozbadev/whisper-large-v3-turbo-hsb-aug

Model Description

This model was fine-tuned on over 98 hours of transcribed upper sorbian speech, including colloquial speech. Part of the corpus was augmented to create additional training data of 89 hours.

Training Data

• Sources:
  • Załožba za serbski lud (https://zalozba.de/)
  • Korla Baier (https://huggingface.co/Korla/tts-modele)
• Volume: 5900 minutes original data, combined with additional 5388 minutes of augmented data, 5% Validation Set

Training Details

• Hyperparameters:
  • Batch size: 8
  • Gradient accumulation steps: 4
  • Learning rate: 5e-6, linear decay
• Warmup: 2000 steps
• Additional Techniques: BF16 training, initial 15 layers frozen

Performance

Metrics

• Model checkpoint: 8000
• Word Error Rate: 4.5

For a later checkpoint with better WER but worse robustness to noise, check this branch: https://huggingface.co/zalozbadev/whisper-large-v3-turbo-hsb-aug/tree/longest_trained

Usage

• Specify transcription language "czech" (model was finetuned this way)

Model Details

• Model Name: zalozbadev/whisper-large-v3-turbo-hsb-aug
• Publisher: Załožba za serbski lud
• Model Version: 1.0.0
• Model Date: 2026-16-01
• License: CC-BY-4.0
• Architecture: Whisper Large v3 Turbo
• Task: Automatic Speech Recognition