akrasi/pl-asr-benchmark

Polish ASR Benchmark

A benchmark dataset for evaluating Polish Automatic Speech Recognition (ASR) systems on medical and healthcare-related speech.

The dataset combines real-world speech recordings sourced from publicly available Creative Commons videos with synthetic speech generated from Polish healthcare articles. It is intended for evaluating transcription quality on domain-specific vocabulary, including medical terminology, pharmaceuticals, anatomy, and public health topics.

Overview

Polish ASR systems often perform well on general speech but struggle with specialized terminology. This benchmark focuses on healthcare and medical language, providing evaluation material that includes disease names, medications, anatomy, clinical concepts, and public health information.

The dataset contains two categories of speech:

1. Real speech recordings obtained from publicly available YouTube videos released under CC BY 4.0.
2. Synthetic speech (TTS) generated from Polish healthcare articles.

The benchmark is intended for evaluation purposes.

Real Speech Sources

The following recordings were used as source material:

ID	Title	Channel	Duration	Speakers	Transcript Type
K1	Dr Joanna Tomiczek-Szwiec: Raka	Czas na Opole	10:12	2	Automatic
K3	KNT: Kodeina, leki z apteki	Drug Pedia	4:23	1	Manual
K5	Pacjent z migrenami	NeuroProjekt	42:45	2	Automatic

Characteristics

• Healthcare and medical subject matter.
• Domain-specific terminology.
• Single-speaker and multi-speaker recordings.
• Moderate to high audio quality.
• Combination of manually prepared and automatically generated transcripts.

Synthetic Speech Sources

Synthetic speech samples were generated from the following healthcare-related articles:

Topic	Source
Jak sobie radzić z alergią na pyłki	pacjent.gov.pl
Szczepienie ratuje życie	pacjent.gov.pl
Co jeść przy nadmiernych wzdęciach	pacjent.gov.pl
Kodeina	doz.pl
Odwodnienie dorosłego	doz.pl
Ból pleców – domowe sposoby i leki z apteki	doz.pl
Stosowanie insuliny	mp.pl
Dopamina	mp.pl
Schizofrenia	mp.pl

These texts were selected to provide vocabulary coverage across multiple healthcare domains, including pharmacology, psychiatry, neurology, endocrinology, allergy treatment, and public health.

Dataset Characteristics

The benchmark includes:

• Polish-language speech only.
• Real human speech recordings.
• Synthetic speech generated from healthcare texts.
• Reference transcriptions.
• Medical and healthcare vocabulary.
• Drug names and pharmaceutical terminology.
• Disease names and clinical concepts.
• Long-form speech content.

Intended Use

This dataset is designed for:

• Benchmarking Polish ASR systems.
• Comparing ASR models on medical-domain speech.
• Evaluating recognition of specialized terminology.
• Measuring Word Error Rate (WER).
• Measuring Character Error Rate (CER).
• Tracking model improvements over time.

Limitations

• The dataset focuses on healthcare and medical topics and is not representative of all spoken Polish.
• Speaker diversity is limited by the available source recordings.
• Results obtained on this benchmark may not generalize to non-medical domains.

License

The dataset is distributed under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

Source recordings were selected from content published under CC BY 4.0. Users are responsible for complying with attribution requirements associated with the original materials.

Acknowledgements

The benchmark was created using publicly available educational and healthcare content from:

• Czas na Opole
• Drug Pedia
• NeuroProjekt
• pacjent.gov.pl
• doz.pl
• mp.pl

Their work made it possible to create an evaluation benchmark for Polish-language medical ASR systems.