meldynamics/liepa-tts

Liepa TTS Dataset

A high-quality Lithuanian text-to-speech (TTS) dataset created under the LIEPA project at Vilniaus universitetas (Vilnius University). This dataset serves as a benchmark for Lithuanian speech synthesis, featuring professional recordings from four distinct speakers.

---

Dataset Summary

The dataset contains approximately 3 hours of speech per speaker, totaling around 12 hours. It was originally developed for the study Lietuviško balso sintezatorių kokybės vertinimas (Evaluation of Lithuanian text-to-speech synthesizers).

Speakers

The dataset features four synthesized voices based on professional actors:

ID	Name	Gender	Age Group	Description
Aiste	Aistė Diržiūtė	Female	Young (22-24)	Clear, youthful tone.
Edvardas	Edvardas Kubilius	Male	Young (23-25)	Deep, resonant voice.
Regina	Regina Jokubauskaitė	Female	Older (64-66)	Mature, warm expressive voice.
Vladas	Vladas Bagdonas	Male	Older (64-66)	Authoritative, mature male voice.

---

Data Fields

Each entry in the dataset contains the following fields:

• audio: The audio recording file (wav).
• transcription: The raw orthographic text of the sentence.
• phonetic_transcription: The primary phonetic transcription recommended for use. This is identical to text_stressed_unicode_a (detailed below).
• fullname: Full name of the speaker.
• age: Age range of the speaker at the time of recording.
• gender: Gender of the speaker.
• filename: Original filename of the audio segment.
• id: Unique identifier for the sentence.

Advanced Stylistic & Phonetic Fields

The dataset includes detailed stress and accentuation variants useful for linguistic research and advanced TTS modeling.

• text_stressed_unicode_a (Recommended): Machine-predicted stress using standard Unicode combining diacritics. This is the most consistent representation for modern TTS systems.
• text_stressed_ascii_a: Machine-predicted stress using legacy ASCII symbols (^, ~, `).
• text_stressed_unicode_b: Human/Reference stress (from PDF) using Unicode diacritics.
• text_stressed_ascii_b: Human/Reference stress (from PDF) using legacy ASCII symbols.
• text_stressed_numbers: Machine-predicted stress represented in a number-based format used by the inner workings of the synthesis engine.

---

Transcription Variants Explained

This project contains two distinct sources of transcriptions for the corpus.

1. Reference Transcript (Variant B)

These correspond to fields ending in _b.

• Source: Extracted from the original atrinkti sak 2a.pdf provided by the dataset authors.
• Characteristics: Represents the intended text with human-validated stress marks.
• Note: Used as the ground truth for spelling and capitalization.

2. Machine-Predicted Transcript (Variant A)

These correspond to fields ending in _a.

• Source: Generated by the Liepa statistical accentuation model.
• Characteristics: Automated stress prediction. While highly accurate, it is a statistical model and may differ slightly from the human reference in ambiguous cases.
• Why use this? It provides a highly consistent and machine-readable format (text_stressed_unicode_a) that effectively handles raw text input scenarios.

Stress Symbol Mapping

We provide stress marks in both ASCII (legacy) and Unicode formats.

Stress Type	ASCII Symbol	Unicode Combining	Example (ASCII)
Grave (Short)	`	\u0300	a `
Acute (Long)	^	\u0301	a^
Tilde (Stressed)	~	\u0303	a~

---

Technical Details

• Sampling Rate: 22,050 Hz (22kHz)
• Format: WAV
• Total Sentences: ~5000 per speaker
• Phonetic Units: >161,000 total

License

This dataset is distributed under the CC BY 4.0 license. please attribute Vilniaus universitetas and the LIEPA project.

Contact

For further details, refer to the LIEPA project page.