eduardem/lili-romanian-single-speaker-piper

---
language:
- ro
license: cc-by-4.0
task_categories:
- text-to-speech
tags:
- audio
- romanian
- tts
- piper
- single-speaker
pretty_name: Lili Romanian Single-Speaker Piper Dataset
size_categories:
- 10K<n<100K
---

# Lili Romanian Single-Speaker Piper Dataset

A curated Romanian single-speaker speech dataset prepared for Piper training.

| | |
|---|---|
| **Segments** | 10,738 |
| **Total duration** | 22.91 hours |
| **Speaker** | Lili |
| **Gender** | female |
| **Language** | Romanian (ro) |
| **Audio format** | WAV, 16-bit, mono, 22.05 kHz |
| **Segment duration** | 2.52 - 9.99 seconds |

## Summary

This dataset contains a single Romanian narrator exposed as `Lili`.
It is published as a Hugging Face Parquet-backed audio dataset, so the Hub viewer
can stream rows directly and play audio inline.

## Dataset Structure

| Column | Type | Description |
|--------|------|-------------|
| `audio` | Audio | WAV audio sample |
| `text` | string | Romanian transcript |
| `speaker` | string | Speaker name |
| `gender` | string | Speaker gender |
| `duration` | float | Duration in seconds |
| `source` | string | Broad source type (`audiobook`) |
| `source_id` | string | Original book/narrator bucket id from the upstream dataset |
| `dnsmos` | float | Quality score retained from preprocessing |
| `clip_key` | string | Stable clip identifier |

## Source Dataset

This dataset is a processed derivative of [`datadriven-company/TTS-Romanian`](https://huggingface.co/datasets/datadriven-company/TTS-Romanian).
The final speaker pool was consolidated from the following matched source ids:

- `cartia_478_Florian_Cristescu_Familia_Roademult`
- `cartia_486_Lloyd_Douglas_Camasa_lui_Cristos`
- `cartia_489_Anton_Pavlovici_Cehov_Calugarul_negru`
- `cartia_490_Lloyd_Douglas_Marele_Pescar`
- `cartia_505_Ionel_Teodoreanu_La_Medeleni_Volumul_1_Hotarul_nestatornic`
- `cartia_512_Ionel_Teodoreanu_La_Medeleni_Volumul_2_Drumuri`

## Processing

The published subset corresponds to the cleaned `<=10s` baseline used for Piper training.

- narrator pooling across matched audiobook ids
- global exact-text dedup before export
- edge-silence trimming
- clip-level QC filtering on noise, clipping, silence, speech rate, and transcript anomalies

Average duration: **7.68s**

QC summary from the final local cleaning pass:

- kept clips: **10,738**
- kept hours: **22.906**
- rejected clips from the `<=10s` pool: **788**
- rejected hours from the `<=10s` pool: **1.400**

## Usage

```python
from datasets import load_dataset

ds = load_dataset("eduardem/lili-romanian-single-speaker-piper")
sample = ds["train"][0]
print(sample["text"])
print(sample["audio"]["path"])
```

## Attribution

Please preserve attribution to [`datadriven-company/TTS-Romanian`](https://huggingface.co/datasets/datadriven-company/TTS-Romanian)
when redistributing this dataset or derivatives trained from it.