// About
What it is
At bards.ai, we fine-tuned OpenAI's Whisper model specifically for the Polish language as part of Hugging Face's global Whisper Fine-Tuning Sprint.
Our model took first place in the Polish category, standing out for its exceptional transcription accuracy and robustness across diverse Polish audio sources.
We focused on minimizing word error rate while preserving the natural flow of spoken Polish, ensuring the model performs well across interviews, podcasts, casual speech, and formal recordings. The result is a state-of-the-art Polish ASR model ready for real-world use in transcription, accessibility, and voice-driven applications.
// Performance
Evaluation metrics
Loss
0.3684
WER
7.2802%
// Training
Hyperparameters
- Learning rate
- 1e-05
- Train batch size
- 8
- Eval batch size
- 4
- Seed
- 42
- Gradient accumulation steps
- 8
- Total train batch size
- 64
- Optimizer
- Adam (betas=(0.9, 0.999), epsilon=1e-08)
- LR scheduler type
- linear
- LR scheduler warmup steps
- 500
- Training steps
- 2100
- Mixed precision training
- Native AMP
// Get the model
Available on Hugging Face
// More open source
You might want these too
Open Source
Jaskier 7b LLM
Top-ranked <=7B LLM (Feb 2024): Fast, smart, open-source. Fine-tuned for exceptional reasoning and instruction-following.
Explore the release
Open Source
Finance Sentiment
A family of language-specific models for detecting sentiment of financial news
Explore the release
Open Source
Fast Twitter Emotion Detection
Emotion classification model for Polish tweets. Detects joy, optimism, sadness, and anger with high accuracy and speed.
Explore the release
// Need a model like this?
We build production models. Then we open-source the useful ones.