Dataset Card for LIEPA-2

Dataset Summary

The LIEPA-2 dataset is a large-scale annotated speech corpus for the Lithuanian language, developed under the project "Development of Services Controlled by Lithuanian Speech" (LIEPA-2). It is a phonetically representative, structured collection of data (audio recordings and annotations) designed for scientific research in speech technologies and the development of electronic services.

• Total Duration: 1000 hours
• Access: Due to the large size of the dataset, it is not available for direct download. Interested parties are requested to contact the creators at gediminas.navickas@mif.vu.lt.

The dataset was created by Vilnius University (Institute of Data Science and Digital Technologies, Faculty of Philology) and partners.

Dataset Statistics

• Total Duration: 1000 hours
• Total Speakers: 2,621
• Speech Types:
  • Read speech: 91%
  • Spontaneous speech: 9%

Recording Environments

• Studio recordings: 62%
• Dictaphone recordings: 28%
• Audio books: 3%
• TV broadcasts: 3%
• Radio broadcasts: 2%
• Telephone recordings: 2%

Speaker Demographics

• Gender Distribution:
  • Female: 56%
  • Male: 44%
• Age Distribution:
  • Under 12 years: 8%
  • 13–17 years: 3%
  • 18–25 years: 18%
  • 26–60 years: 61%
  • Over 60 years: 10%

Dataset Structure

Data Format

This dataset is SLICED for ASR training. Each row represents a single audio segment (utterance or noise event).

• Audio: The sliced audio segment (WAV, 16kHz).
• Text: The transcription of the segment.
  • For Noise Segments (is_noise=True), the text contains the noise label (e.g., +BREATH+, +SMACK+, +GARBAGE+).
• Metadata Fields:
  • is_noise: Boolean. True if the segment represents a noise event, False for speech. Note: The dataset explicitely includes noise segments as separate rows for training stability/negative sampling.
  • speaker_id: Unique speaker identifier.
  • gender: Female / Male.
  • age_group: e.g., 26-60, 0-12.
  • duration_ms: Length of segment in milliseconds.
  • tags: List of entry-level tags (e.g., ['abc'] for alphabet reading).
  • annotator: 2-letter code of the annotator (derived from Speaker ID).
  • entry_id: ID of the original recording (EAF file).
  • segment_id: ID of the segment within the recording.
  • start_ms / end_ms: Timestamp in the original file.

Original Documentation

For the full labeling protocol and noise class definitions, please refer to the readme.pdf included in the repository root.

Additional Information

Licensing Information

The LIEPA-2 dataset is distributed under the CC BY (Creative Commons Attribution) license.

Citation Information

When using this dataset for the development of other products, please mention in your product description that the dataset was created at Vilnius University under the project "Lietuvių šneka valdomų paslaugų plėtra – LIEPA 2".

Contact

For access to the original dataset, please contact: gediminas.navickas@mif.vu.lt

• Project Website: liepa2.rastija.lt
• Creators: Vilnius University (Institute of Mathematics and Informatics, Faculty of Philology).

References

• Garsynas LIEPA-2