EuroSpeech Dataset

Dataset Description

EuroSpeech is a large-scale multilingual speech corpus containing high-quality aligned parliamentary speech across 22 European languages. The dataset was constructed by processing parliamentary proceedings using a robust alignment pipeline that handles diverse audio formats and non-verbatim transcripts. More information can be found in the paper.

This dataset is 16 kHz, the 24 kHz version of EuroSpeech can be found at disco-eth/EuroSpeech-24kHz

Dataset Summary

• Languages: 22 European languages (see detailed breakdown below)
• Total aligned hours: ~78,100 hours of initially aligned speech-text data
• Quality-filtered subsets:
  • CER < 30%: approximately 61,000 hours
  • CER < 20%: approximately 50,500 hours (this is the primary subset provided directly through the Hugging Face Datasets interface for all languages)
  • CER < 10%: approximately 32,200 hours
• Domain: Parliamentary proceedings (formal speaking style)
• Audio segment length: Typically 3-20 seconds
• Format: Audio segments with paired transcriptions

Languages

EuroSpeech provides substantial data for previously under-resourced languages:

• 19 languages exceed 1,000 hours of data (CER < 20%)
• 22 languages exceed 500 hours of data (CER < 20%)

Language	Code	Total Aligned (h)	CER < 30% (h)	CER < 20% (h)	CER < 10% (h)
Croatia	hr	7484.9	5899.7	5615.8	4592.0
Denmark	da	7014.2	6435.0	5559.8	3443.7
Norway	no	5326.2	4578.8	3866.7	2252.2
Portugal	pt	5096.3	4036.7	3293.5	2105.9
Italy	it	4812.8	3539.6	2813.7	1767.3
Lithuania	lt	5537.9	3971.0	2681.2	956.6
United Kingdom	en	5212.2	3790.7	2609.3	1175.0
Slovakia	sk	2863.4	2722.4	2553.6	2070.8
Greece	el	3096.7	2717.6	2395.4	1620.9
Sweden	sv	3819.4	2862.6	2312.8	1360.1
France	fr	5476.8	2972.1	2249.8	1347.6
Bulgaria	bg	3419.6	2570.4	2200.1	1472.8
Germany	de	2472.2	2354.2	2184.4	1698.4
Serbia	sr	2263.1	1985.1	1855.7	1374.1
Finland	fi	2130.6	1991.4	1848.2	1442.2
Latvia	lv	2047.4	1627.9	1218.8	499.9
Ukraine	uk	1287.8	1238.3	1191.1	1029.8
Slovenia	sl	1338.2	1241.7	1156.4	900.5
Estonia	et	1701.1	1430.9	1014.9	382.5
Bosnia & Herz.	bs	860.2	781.9	691.3	447.8
Iceland	is	1586.1	974.1	647.4	171.4
Malta	mt	3281.6	1284.3	613.0	143.9
Total		78128.6	61006.4	50572.9	32255.5

Dataset Structure

Data Instances

Each instance in the dataset consists of:

• Audio segment (3-20 seconds)
• Corresponding transcript text
• Metadata including language, source session, alignment quality metrics

Data Splits

The dataset provides predefined train, development, and test splits for each language. To ensure data integrity and prevent leakage between sets, these splits are constructed by assigning entire parliamentary sessions (i.e., all segments derived from a single original long audio recording) exclusively to one of the train, development, or test sets. The exact proportions follow common practices (e.g., 80/10/10).

Dataset Creation

Source Data

The data was collected from parliamentary proceedings across 22 European nations. Parliamentary sessions offer high-quality speech in a formal register, typically featuring clear speech with good audio quality and professional transcripts.

Data Collection and Processing

The dataset was constructed using a multi-stage pipeline:

1. Data Sourcing and Metadata Collection: Manual and scripted gathering of media/transcript links from parliamentary websites.

2. Download Pipeline: Automated retrieval of audio, video, and transcript files using specialized handlers for diverse source formats.

3. Alignment Pipeline:

   • Segmentation of long recordings into 3-20 second utterances using voice activity detection (VAD)
   • Transcription of segments using an ASR model to produce pseudo-labels
   • Alignment of segments to transcripts using a novel two-stage dynamic algorithm
   • Selection of best-aligned transcript formats and quality filtering

4. Filtering: CER-based filtering to create quality tiers (CER < 30%, < 20%, < 10%)

Alignment Algorithm

The core of the alignment process is a novel two-stage dynamic algorithm specifically engineered for extreme robustness when matching ASR pseudo-labels to noisy, non-verbatim parliamentary transcripts:

1. Coarse stage: Uses a sliding window to rapidly scan the transcript, efficiently bypassing large irrelevant sections to identify a set of top-k candidate text spans via Character Error Rate (CER).

2. Fine-tuning stage: Performs a local search around promising candidates, optimizing start position and window size for the best CER.

A fallback mechanism restarts the search if no initial match meets a predefined quality threshold.

Dataset Use

Intended Uses

The EuroSpeech dataset is intended for:

• Training and evaluating automatic speech recognition (ASR) systems
• Training and evaluating text-to-speech (TTS) systems
• Multilingual speech research
• Low-resource language speech technology development
• Cross-lingual transfer learning in speech models

Citation Information

If you use this dataset, please cite:

@inproceedings{pfisterereurospeech,
  title={EuroSpeech: A Multilingual Speech Corpus},
  author={Pfisterer, Samuel and Gr{\"o}tschla, Florian and Lanzend{\"o}rfer, Luca A and Yan, Florian and Wattenhofer, Roger},
  booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems Datasets and Benchmarks Track}
}

Considerations

Data Quality

The dataset provides multiple quality tiers based on Character Error Rate (CER):

• CER < 30%: More data, but potentially lower quality alignments
• CER < 20%: Balanced quality-quantity trade-off (recommended for most applications)
• CER < 10%: Highest quality alignments, but reduced quantity

Licensing Information

The licensing terms vary by country as each parliament has its own policies. The table below provides relevant sources for each parliament in our dataset.

Please note that we do not guarantee the accuracy of this information and take no responsibility for any use that conflicts with applicable licenses or laws. Users are responsible for ensuring compliance with relevant terms.

Copyright and Licensing Information for each Parliament

Country	Source
Croatia	Legal Notice
Denmark	Legal Notice
Norway	NLOD License
Portugal	Portuguese Copyright Code Article 75
Italy	Italian Parliament Website references CC By 4.0 License
Lithuania	Republic of Lithuania Law on Copyright and Related Rights Article 22
United Kingdom	Terms and Conditions for audio, Open Government Licence for transcripts
Slovakia	Slovak Copyright Act Chapter One Section 5e)
Greece	Greek Copyright Law Article 2(5) and Article 25(1)(b)
Sweden	Law (2022:818)
France	License Ouverte
Bulgaria	Copyright Policy references CC BY 2.5 BG
Germany	Terms of Use
Serbia	Serbian Law on Copyright and Related Rights Article 6(2)
Finland	Copyright Act Article 9, 22, and 25
Latvia	Latvian Copyright Law Section 21
Ukraine	Law of Ukraine on Copyright and Related Rights Article 8(1)(3)
Slovenia	Copyright and Related Rights Act Article 46-51
Estonia	Copyright Act, Estonian Youtube references CC BY SA
Bosnia & Herzegovina	Copyright Law Article 44 and 47
Iceland	Copyright Act Article 22
Malta	Re-Use of Public Sector Information Act Chapter 546

Limitations

• The dataset primarily represents formal parliamentary speech and may not generalize well to casual, spontaneous, or noisy speech environments.
• The dataset reflects the demographics and speaking styles of European parliamentarians, which may not be representative of the general population.
• Some languages have significantly more data than others, which could lead to performance disparities in multilingual models.

Additional Information

Dataset Curators

• Samuel Pfisterer (@SamuelPfisterer1)
• Florian Grötschla (@FloGr)
• Luca Lanzendörfer (@lucala)
• Florian Yan (@floyan)
• Roger Wattenhofer

Maintenance Status

[Information about maintenance and update plans]

Links

• EuroSpeech on Hugging Face Datasets
• EuroSpeech GitHub Repository