Hugging Face's logo

Datasets:
alakxender
/
dhivehi-audios-82-spk

Tasks:
Automatic Speech Recognition
Text-to-Speech
Modalities:
Audio
Text
Formats:
parquet
Languages:
Divehi
Size:
1M - 10M
Libraries:
Datasets
Dask
License:
Dataset card Data Studio Files
xet
Community
1

You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

Dhivehi Synthetic Voice and Speech Augmentation Dataset

This dataset is a multi-speaker dataset containing 1.26 million synthetic audio samples (~2,627 hours total). Each sample pairs a Dhivehi sentence with an augmented waveform, created through controlled synthesis, voice-cloning, and heavy acoustic perturbations. The dataset was generated to enable ASR, TTS, and voice-representation research in low-resource Dhivehi, focusing on robustness across pronunciation, prosody, and timbre variance.

Process

  • Base text source: sentences from a Dhivehi news corpus.
  • TTS model: speech model fine-tuned for Dhivehi phonetics.
  • Voice cloning: reference recordings used to condition synthetic speakers.
  • Augmentations:
    • speed & tempo variation
    • dynamic range compression
    • pitch shifting (± semitones)
    • formant warping & spectral noise
    • reverb & background mix-in (random: 50-80 samples added to each subset)
    • pronunciation drift simulation
  • Generation time: ~36 hours of continuous synthesis.
  • Sampling rate: 16 kHz PCM WAV.

Each row in the metadata includes:

Field Description
audio Path to .wav file
sentence Dhivehi text string
speaker_id Original speaker tag (e.g. fh_00)
subset_id Merged canonical speaker (e.g. f_00)
gender male / female
presence heavy / medium / light (augmentation intensity)

Dataset

Split Total Samples Duration (hrs) Speakers Avg Len (s)
Female (heavy) 78 625 153.6 5 7.04
Female (medium) 235 862 459.6 15 7.02
Female (light) 157 299 306.8 12 7.02
Male (heavy) 77 639 169.3 5 7.85
Male (medium) 235 871 512.1 15 7.82
Male (light) 471 729 1 026.3 30 7.83

Total: ≈ 1 257 024 samples (≈ 2 627 hours, 82 unique speakers)

Speaker Composition

  • Male: ≈ 65 % of samples (45 speakers)
  • Female: ≈ 35 % of samples (37 speakers)
  • Per-speaker duration: ~30 – 34 hours (balanced)
  • Pronunciation depth: each speaker appears under multiple augmentation presets (heavy, medium, light), producing distinct acoustic conditions.

Text Statistics

  • Average sentence length: ~118 characters
  • Median length: ~112 characters
  • Range: 8 – 692 characters (max after token cleanup)
  • Texts cover news, politics, society, and general narration topics.

Intended Use

Designed primarily for:

  • Fine-tuning and evaluation of Dhivehi TTS systems
  • Automatic Speech Recognition (ASR) robustness tests
  • Speaker embedding and voice transfer experiments
  • Cross-speaker adaptation and data-augmentation research

Not recommended for direct human listening or production-grade speech models without validation.

Disclaimer

This dataset contains synthetic audio generated for research and evaluation purposes only.
It does not represent real human voices, nor does it reflect any individual’s identity or opinion.

Downloads last month
17