Hugging Face's logo

Datasets:
aangelakis
/
STOMA

Tasks:
Text-to-Speech
Automatic Speech Recognition
Audio Classification
Modalities:
Audio
Text
Formats:
parquet
Languages:
Greek
Size:
10K - 100K
Tags:
speech_corpus
greek_language
text-to-speech
neural_TTS
speech_synthesis
under_resourced_languages
Libraries:
Datasets
Dask
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
STOMA / README.md
aangelakis's picture
aangelakis
Update README.md
ea413ab verified
preview code
|
raw
history blame contribute delete
14.6 kB
metadata 
pretty_name: 'STOMA: A Multi-Speaker Greek Speech Corpus'
annotations_creators:
  - expert-generated
language:
  - el
license: cc-by-4.0
task_categories:
  - text-to-speech
  - automatic-speech-recognition
  - audio-classification
tags:
  - speech_corpus
  - greek_language
  - text-to-speech
  - neural_TTS
  - speech_synthesis
  - under_resourced_languages
  - audio
size_categories:
  - 10K<n<100K

STOMA: A Multi-Speaker Greek Speech Corpus

STOMA is a new multi-speaker Greek speech corpus designed to advance research in text-to-speech (TTS) synthesis and related speech technologies for Greek, an under-resourced language. The corpus comprises approximately 23 hours of studio-recorded read speech from six native speakers (three male and three female), captured under controlled studio conditions using a dual-booth setup to ensure acoustic consistency and high signal quality. The spoken material was selected from the Greek Harvard Corpus and the Text Bank of the Center for the Greek Language, specifically from texts corresponding to the B2, C1, and C2 proficiency levels of the Certification of Attainment in Greek, ensuring linguistically rich and pedagogically well-balanced content. All recordings were standardized to 44.1 kHz, 16-bit mono PCM format and processed through a hybrid quality-control pipeline combining automated normalization and manual verification. To assess dataset quality, we trained state-of-the-art neural TTS systems based on the FastSpeech2 acoustic model and the HiFi-GAN vocoder, achieving natural and intelligible synthesized speech. The resulting corpus provides a publicly accessible, high-quality resource that supports both linguistic research and the development of modern speech synthesis systems in Greek.

Dataset Structure

The dataset is provided in sharded Parquet format for optimized loading. The original folder structure ("Speakers/" and "Text/") and naming conventions are preserved as metadata features ("speaker_id", "section", "session", etc.) within the dataset.

Data Fields

When you load the dataset, the metadata originally encoded in the raw file names is automatically parsed into the following distinct columns for easier filtering and analysis:

  • speaker_id (string): The speaker identifier.
    • M: Main Male Speaker
    • F: Main Female Speaker
    • M1...MX: Other Male Speakers
    • F1...FX: Other Female Speakers
  • section (string): The section of the corpus the text originated from (e.g., Harvard, B2, C1, C2).
  • session (int64): The recording session number.
  • sentence_id (int64): The unique sentence identifier within that specific session.
  • text (string): The Greek transcript of the audio.
  • audio (audio): The loaded audio feature, which includes the raw audio array and the sampling rate.
  • file_name (string): The original filename of the source .wav audio file (e.g., Greek-F1-B2-05-20.wav).

Usage

You can load this dataset directly in Python using the Hugging Face datasets library. 

from datasets import load_dataset

# Download and Load the dataset
dataset = load_dataset("aangelakis/STOMA", split="train")

# Inspect the first example
sample = dataset[0]

# Print the Greek text
print("Text:", sample["text"])

# Print metadata
print("Speaker:", sample["speaker_id"])
print("Corpus Section:", sample["section"])
print("Session:", sample["session"])
print("Sentence ID:", sample["sentence_id"])
print("Original File Name:", sample["file_name"])

# The 'audio' column is automatically decoded into a dictionary with 'array' and 'sampling_rate'
audio_data = sample["audio"]["array"]
sampling_rate = sample["audio"]["sampling_rate"]

License

STOMA is licensed under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

Citation

If you use this work, please cite:

TBA

References

More Information

Check out our GitHub repository for the training and inference procedures: STOMA: A Multi-Speaker Greek Speech Corpus

Table 1: Text statistics for the full text collection (Greek Harvard, B2, C1, and C2), recorded in its entirety by the two primary speakers.

Statistic Gr. Harvard B2 C1 C2 Total
Total Sentences 720 1,087 1,571 1,296 4,674
Total Words 5,539 15,102 25,263 22,722 68,626
Total Characters 32,301 95,119 164,700 152,904 445,024
Distinct Words 3,343 5,606 8,965 8,275 20,253
Mean Words per Sentence 7.69 13.89 16.08 17.53 14.68
Min Words per Sentence 5 1 1 1 1
Max Words per Sentence 9 32 41 39 41

Table 2: Text statistics for the reduced text subset selected from the full collection, used for recordings by the secondary speakers.

Statistic Gr. Harvard B2 C1 C2 Subset Total
Total Sentences 720 190 194 197 1,301
Total Words 5,539 2,913 3,057 3,265 14,774
Total Characters 32,301 18,286 20,190 22,110 92,887
Distinct Words 3,343 1,519 1,604 1,676 6,899
Mean Words per Sentence 7.69 15.33 15.76 16.57 11.36
Min Words per Sentence 5 4 3 4 3
Max Words per Sentence 9 28 28 30 30

Table 3: Speech statistics of the STOMA corpus per speaker and overall.

Speaker ID Clips Total Duration Mean Duration (s) Min–Max (s)
F 4,674 7:46:21 5.99 0.41–16.60
M 4,674 8:53:37 6.85 0.48–22.31
F1 1,301 1:46:58 4.93 1.65–15.08
F2 1,301 1:27:38 4.04 1.46–11.13
M1 1,301 1:47:06 4.94 1.64–14.99
M2 1,301 1:26:43 4.00 1.10–10.39
Total 14,552 23:08:26 5.72 0.41–22.31

Table 4: Speaker-specific acoustic characteristics of the STOMA corpus. SR: average speech rate (syllables/s, including pauses); AR: average articulation rate (syllables/s, excluding pauses); ASD: average syllable duration (s), computed across all utterances per speaker.

Speaker ID Mean F0 (Hz) Std F0 (Hz) SR AR ASD (s)
M 118.73 20.80 4.143 4.286 0.236
F 195.53 40.16 4.462 4.662 0.218
M1 128.30 23.34 4.165 4.403 0.230
F1 200.93 23.57 3.887 3.974 0.257
M2 109.33 15.49 4.407 4.527 0.225
F2 187.29 26.88 4.238 4.409 0.232

Table 5: Demographic information of the STOMA corpus speakers.

Speaker ID Age Sex Region Raised
M 29 Male Athens
F 37 Female Crete
M1 30 Male Alexandroupoli
F1 21 Female Larisa
M2 22 Male Larisa
F2 26 Female Athens

Table 6: Audio format specifications of the STOMA corpus recordings.

Property Value
File format WAV (RIFF)
Encoding Pulse-Code Modulation (PCM)
Compression None (uncompressed)
Channels 1 (mono)
Sampling rate 44.1 kHz
Bit depth 16-bit signed integer