rishchen/ukrainian-tts-audiobook-pani-nina-parquet-old

---
license: cc-by-nc-sa-4.0
task_categories:
- text-to-speech
language:
- uk
size_categories:
- 100K<n<1M
---

# Ukrainian TTS audiobook dataset (Parquet)

Segmented Ukrainian audiobook speech with aligned text, prepared for training and evaluating Text-to-Speech (TTS) models. The dataset is published as Hugging Face-compatible Parquet shards so the Hub **Dataset Preview** can render an `audio` column.
That was mabe by using `whisper` (https://github.com/openai/whisper) and `ffmpeg` (https://www.ffmpeg.org/), where with whisper we set start and end of voices + transcribe it and using ffmpeg slice into ~2-10 seconds pieces.

## Motivation / use case

- Train Ukrainian TTS / speech synthesis models on long-form narrated speech.
- Provide a simple tabular format (`audio` + `text` + metadata) that works well with `datasets`.

## Dataset format

Each example is one utterance:

- `id` (`int64`): sequential index (0..N-1)
- `path` (`string`): relative path of the original `.wav` (kept for traceability)
- `audio` (`Audio`): Hugging Face audio feature stored as a struct `{bytes, path}`
- `text` (`string`): Ukrainian transcript
- `duration` (`float32`): seconds
- `source` (`string`): original source recording name (from `metadata.jsonl`)

## Dataset stats

- Rows: `116,575`
- Total duration: `~114.9 hours`
- Sources: `39` unique recordings (see `source` field)
- Audio: mono, PCM16, 16 kHz WAV

## Install deps

```bash
python -m pip install huggingface_hub==0.30.2 datasets==3.5.0
```

## Load with `datasets`

From the Hub:

```python
from datasets import load_dataset

ds = load_dataset("rishchen/ukrainian-tts-audiobook-pani-nina-parquet", split="train")
```

From local Parquet shards:

```python
from datasets import load_dataset

ds = load_dataset("parquet", data_files={"train": "train/train-*.parquet"})["train"]
```

## Download from Hugging Face

This repo is set up to upload the contents of `train/` as the dataset repository root (so `wavs/...` paths match the Parquet).

```bash
export HF_TOKEN="hf_..."
python download.py
python unpack.py
```

## Notes for Hub Dataset Preview

- The Parquet shards include Hugging Face `datasets` feature metadata so `audio` is recognized as an `Audio` column.
- If you generate *path-only* Parquet (default), keep the referenced `.wav` files in the same repo so `audio.path` resolves.
- If you generate Parquet with `--embed-audio-bytes`, the dataset is self-contained and you can upload only Parquet.
- If you publish both `metadata.jsonl` and Parquet, the Hub may try to load either format; if you want Parquet-only, upload only the shards (or rename/move the raw JSONL out of the dataset root before pushing).