meldynamics/liepa-asr

Liepa ASR Dataset

Lithuanian Automatic Speech Recognition (ASR) dataset from the LIEPA project (Lietuvių šnekos garsynas LIEPA) developed at Vilnius University.
It provides a phonetically representative corpus for ASR and TTS research, capturing diverse speakers and recording styles.

---

---

Dataset Summary

The Liepa ASR dataset contains speech recordings and their transcriptions, designed for both speech recognition and speech synthesis research.

• Total speakers: 376 (248 female, 128 male), mostly aged 15–22.
• Total hours: 100h.
• Splits:
  • Train: ~90% of speakers
  • Test: ~10% of speakers
• Splits are balanced by age groups and gender so that evaluation is representative.

---

Segments

• 100 Hours Segment:
  Contains speech recordings with annotated words, phrases, full sentences, and parts thereof.

---

Record Structure

Each record has the following fields:

Field	Description
transcription	Text transcription of the audio sample (UTF-8).
filename	Original filename, e.g. S010Mi_001_01.wav.
filepath	Full relative path to the file in the archive.
audio	Audio data as a datasets.Audio object (22 kHz, 16-bit, mono).
type	Utterance type: "Sentence" or "Word".
voice_id	Numeric speaker ID (e.g. "010").
gender	"Female" or "Male".
age	Explicit age or age range (e.g. 16, 21–25, >61).
group_id	Sentence/word group identifier (e.g. 001, 039).
recording	Recording number within the group (string, may be empty if unspecified).

---

Example

{
  "transcription": "Labas rytas",
  "filename": "S010Mi_001_01.wav",
  "filepath": "MII_LIEPA_REC_V1/D002/S001/S010Mi_001_01.wav",
  "audio": {
    "path": "MII_LIEPA_REC_V1/D002/S001/S010Mi_001_01.wav",
    "array": [...],
    "sampling_rate": 22000
  },
  "type": "Sentence",
  "voice_id": "010",
  "gender": "Female",
  "age": "18",
  "group_id": "001",
  "recording": "01"
}

Licensing

This dataset is distributed under the CC BY license. Please include appropriate attribution to Vilniaus universitetas and the LIEPA project when using or redistributing the data.

---

Contact and Further Information

For more details, dataset access, or inquiries, please refer to the LIEPA project page.

---

Acknowledgements

This dataset was developed as part of the LIEPA project ("Lietuvių šneka valdomos paslaugos") at Vilniaus universitetas between 2013 and 2015. We gratefully acknowledge the contributions of the speakers, researchers, and partners involved in its creation.

This dataset card is provided to support the understanding and use of the Lithuanian ASR dataset for speech recognition research.