---
license: cc-by-4.0
task_categories:
- text-to-speech
language:
- lt
pretty_name: Liepa TTS
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: test
    path: data/test-*
dataset_info:
  features:
  - name: audio
    dtype: audio
  - name: transcription
    dtype: string
  - name: phonetic_transcription
    dtype: string
  - name: fullname
    dtype: string
  - name: age
    dtype: string
  - name: gender
    dtype: string
  - name: filename
    dtype: string
  - name: filepath
    dtype: string
  - name: text_stressed_numbers
    dtype: string
  - name: text_stressed_ascii_a
    dtype: string
  - name: text_stressed_unicode_a
    dtype: string
  - name: text_stressed_ascii_b
    dtype: string
  - name: text_stressed_unicode_b
    dtype: string
  - name: id
    dtype: int32
  - name: phoneme_transcription
    dtype: string
  splits:
  - name: train
    num_bytes: 1897690649
    num_examples: 18162
  - name: test
    num_bytes: 210852485
    num_examples: 2018
  download_size: 2069568677
  dataset_size: 2108543134
---

# LIEPA TTS Dataset

This dataset contains **20,180 recordings (13.21 hours)** from four Lithuanian
speakers in the Vilnius University **LIEPA** speech-synthesis project. It combines
the original audio, corrected orthographic transcripts, optional stress hints,
historical PDF-derived text, and the original per-recording phoneme annotations.

The original LIEPA project produced the recordings, voices and synthesis system.
This edition organizes those resources for use with Hugging Face `datasets` and
records subsequent transcript and stress corrections in auditable review files.

## Speakers and splits

| Speaker ID | Name | Gender | Age group | Recordings |
| --- | --- | --- | --- | ---: |
| Aiste | Aistė Diržiūtė | Female | 22–24 | 5,019 |
| Edvardas | Edvardas Kubilius | Male | 23–25 | 5,018 |
| Regina | Regina Jokubauskaitė | Female | 64–66 | 5,124 |
| Vladas | Vladas Bagdonas | Male | 64–66 | 5,019 |

The supplied splits contain **18,162 training** and **2,018 test** recordings.

## Data fields

| Column | Contents |
| --- | --- |
| `audio` | Embedded WAV recording. Original files are mono, 16-bit PCM at 22,050 Hz; see the decoding note below. |
| `transcription` | Current unstressed orthographic text, including accepted word corrections. |
| `phonetic_transcription` | Legacy name for current orthographic text with optional stress hints. Exactly equal to `text_stressed_unicode_a`; it is not a sequence of phonemes. |
| `phoneme_transcription` | Original per-recording phoneme string from this voice's `db_fon.txt`, in native LIEPA notation. It is not IPA or regenerated G2P. |
| `text_stressed_numbers` | Current Variant A text in the LIEPA numeric stress encoding. |
| `text_stressed_ascii_a` | Current Variant A text with ASCII stress marks. |
| `text_stressed_unicode_a` | Recommended current orthographic text with Unicode combining stress hints. |
| `text_stressed_ascii_b` | Historical PDF-derived Variant B text with ASCII stress marks. |
| `text_stressed_unicode_b` | Historical PDF-derived Variant B text with Unicode stress marks. |
| `fullname` | Speaker name. |
| `age` | Speaker age group. |
| `gender` | Speaker gender metadata. |
| `filename` | Original WAV filename; its numeric part is the zero-based recording index within that voice. |
| `filepath` | Original source path, including the voice identifier. |
| `id` | One-based recording number within the voice, equal to the WAV filename number plus one. It is not globally unique; use voice and filename together. |

Use `transcription` for unstressed text and `text_stressed_unicode_a` for text
with optional stress hints. The five current orthographic fields
(`transcription`, `phonetic_transcription`, `text_stressed_numbers`,
`text_stressed_ascii_a`, `text_stressed_unicode_a`) encode the same accepted words.
Their case and stress representations can differ.

Variant B preserves the historical PDF source. Its wording and stress can differ
from the corrected Variant A fields; it has not received the subsequent word
corrections. `phoneme_transcription` also remains a direct original source and
can disagree with corrected spelling.

## Text and stress provenance

Variant A began as engine-derived stressed spelling generated with the modified
LIEPA synthesis application published as
[`phonology_engine`](https://github.com/aleksas/phonology_engine). The original
voice resources informed that reconstruction. The current Variant A fields also
include subsequent listening-confirmed word corrections, supported stress
repairs, removals of misleading hints, and the explicit automatic subset
described below. They are not an untouched engine export or a complete set of
manually verified stress labels.

**Stress marks are optional pronunciation hints.** An unmarked word is valid,
including when a dictionary gives a stress mark or the source phonemes indicate
stress. Dictionary stress need not be fully realized in connected speech.
Missing hints are not errors or a reason to reject a recording. A hint whose
placement contradicts the recording can be misleading and may be removed.

Variant B comes from the published `atrinkti_sak_2a.pdf`. It is retained as a
historical reference, rather than treated as the authority for recorded words
or stress. Original annotations, reconstructed stress and reviewer decisions
remain distinguishable in the linked audits.

| Stress type | ASCII | Unicode combining mark | Numeric encoding |
| --- | --- | --- | --- |
| Grave | ``a` `` | `a` + U+0300 | `A2` |
| Acute | `a^` | `a` + U+0301 | `A3` |
| Tilde | `a~` | `a` + U+0303 | `5A4` |

The numeric representation uses uppercase letters. Unicode stress marks are
combining characters; preserve them when processing the stressed text.

## Original phoneme annotations

`phoneme_transcription` is populated for **all 20,180 recordings**, using each
speaker's own `programos/<Voice>/db_fon.txt` from the original
[LIEPA engine release](https://github.com/aleksas/liepa_dataset/releases/download/data/SintezesVariklis_LIEPAprojektas.zip).
Captured source files and release files were matched against pinned SHA256
hashes. The zero-based WAV filename selects the corresponding zero-based line
within that voice; current text and table order do not determine alignment.

The strings preserve the original notation, excluding only line terminators:
phoneme case, palatalization apostrophes, and `+`, `-`, `_` boundary symbols.
For example:

```text
_+l A J-v a s+p l a w-k' I-m oo+j' Uo-s t oo s+p l oO-t' i s+_
```

These strings are original annotation evidence, not a guarantee of correct
spelling or an instruction to undo reviewed words. They can contain annotation
errors and pronunciation variants. No phonemes were regenerated from the
corrected transcripts for this column. See the
[source and alignment audit](reviews/original-phoneme-column-20261101.json).

## Transcript corrections and review status

The latest word-correction pass changed **76 recordings**: **47 explicitly
listener-confirmed corrections**, **12 automatic contextual word substitutions**
and **17 automatic spelling repairs**. The 29 automatic corrections are
identified as automatic in the audit; they are not claimed as listener-reviewed.
Every complete replacement sentence in this subset matched its voice's original
phoneme stream, with context used to resolve plausible spellings. A unique
dictionary suggestion alone was insufficient for acceptance. Explicit listener
decisions to retain the existing words were preserved.

All five current orthographic fields were updated together. Stress on unchanged
words was retained. Changed words kept projected old hints only when the exact
marked form was already attested in the corpus; otherwise the optional hint was
omitted. This pass did not predict new stress or change historical Variant B,
original phoneme strings, audio, IDs or splits. See the
[76-recording correction audit](reviews/obvious-word-fixes-20261101.json).

Earlier accepted corrections are also part of this edition. Their original
audit files preserve the decisions and their historical state:

- [Regina word corrections](reviews/regina-transcript-corrections-20261028.json) and [listening decisions](reviews/regina-listening-review-20261028.jsonl).
- [Four-voice word corrections](reviews/all-voice-transcript-corrections-20261031.json) and [listening decisions](reviews/all-voice-listening-review-20261031.jsonl).
- [Stress restoration](reviews/stress-restoration-20261031.json) and [subsequent listener decisions](reviews/listener-stress-review-20261031.json).
- [Optional-hint placement review](reviews/stress-placement-review-20261101.json) and [amended decisions](reviews/stress-placement-recheck-20261101.json).

The final placement decisions removed 20 misleading hints in the reviewed
sample. Older audit entries describing pending stress cases are historical;
absence of an optional hint is valid. The reviewed samples do not establish
accuracy for every remaining transcript or justify copying stress decisions
between voices. Further word/annotation conflicts remain unresolved.

## Audio format and loading

All original embedded WAV files are **22,050 Hz, mono, 16-bit PCM**. Their total
duration is **13.21 hours**, measured from all 20,180 original WAV headers.

The current Hugging Face `Audio` feature metadata specifies **22,000 Hz** for
decoding. A decoder honoring that setting resamples the original WAV to
22,000 Hz. To request decoding at the original rate explicitly:

```python
from datasets import Audio, load_dataset

dataset = load_dataset("meldynamics/liepa-tts", split="train")
dataset = dataset.cast_column("audio", Audio(sampling_rate=22050))
```

The source WAV bytes remain at 22,050 Hz. This description update does not alter
the existing decoding metadata, audio bytes, transcripts or split membership.

## Citation and attribution

Please cite the original LIEPA evaluation study:

**Lietuviško balso sintezatorių kokybės vertinimas / Evaluation of Lithuanian text-to-speech synthesizers**  
DOI: [10.5755/j01.sal.0.28.15130](https://doi.org/10.5755/j01.sal.0.28.15130).

This dataset edition is distributed under **CC BY 4.0**. Please attribute:

- **Vilnius University and the LIEPA project** for the original recordings, synthesis resources and source annotations.
- **MEL DYNAMICS — [`meldynamics/liepa-tts`](https://huggingface.co/datasets/meldynamics/liepa-tts)** for this structured edition, transcript repairs, stress representations and dataset packaging.

Related tooling: [`aleksas/liepa_dataset`](https://github.com/aleksas/liepa_dataset)
and [`aleksas/phonology_engine`](https://github.com/aleksas/phonology_engine).
For original project information, see the
[LIEPA project page](https://xn--ratija-ckb.lt/liepa/infrastrukturines-paslaugos/garsynas-liepa/).