Datasets:
metadata
language:
- el
license: cc-by-4.0
task_categories:
- text-to-speech
- automatic-speech-recognition
pretty_name: TTS-Greek
size_categories:
- 10K<n<100K
tags:
- audio
- speech
- greek
- tts
- audiobooks
TTS-Greek
A large-scale, high-quality Greek speech dataset for text-to-speech and automatic speech recognition.
Data Sources
This dataset combines two sources:
| Source | Samples | Hours | License | Content |
|---|---|---|---|---|
| LibriVox | 34,727 | 96.8 | Public Domain | Modern Greek classic literature, philosophy, fiction |
| FLEURS-R (Google) | 4,124 | 12.6 | CC-BY 4.0 | Wikipedia-sourced sentences, AI-restored audio |
Dataset Statistics
| Metric | Value |
|---|---|
| Total samples | 38,851 |
| Total duration | 109 hours |
| Unique speakers | 35 |
| Average duration | 10.1 seconds |
| Average DNSMOS | 3.90 |
Gender Distribution
| Gender | Samples | Hours |
|---|---|---|
| Male | 15,300 | 44.1 |
| Female | 23,551 | 65.3 |
Features
| Field | Type | Description |
|---|---|---|
__key__ |
string | Unique sample identifier |
mp3 |
Audio | Audio (24kHz mono WAV) |
text |
string | Transcription |
language |
string | Language code (el) |
speaker_id |
string | Speaker/audiobook identifier |
gender |
string | male or female |
dnsmos |
float | Audio quality score (1-5) |
Processing Pipeline
Audio was processed through:
- Standardization (24kHz mono, -20dBFS normalization)
- Voice Activity Detection (Silero VAD)
- Transcription & forced alignment (WhisperX)
- Segmentation (5-15s segments at natural pauses)
- Noise reduction (DeepFilterNet)
- Quality filtering (DNSMOS)
Usage
from datasets import load_dataset
ds = load_dataset("datadriven-company/TTS-Greek", split="train")
# Access a sample
sample = ds[0]
print(sample["text"])
# Audio: sample["mp3"]
License
The combined dataset is licensed under CC-BY 4.0 (the more restrictive of the two source licenses). Attribution: LibriVox (Public Domain) + Google FLEURS-R (CC-BY 4.0).