LJ-TTS: A Paired Real and Synthetic Speech Dataset for Single-Speaker TTS Analysis

LJ-TTS is a large-scale dataset containing real human speech and synthetic speech generated by 11 state-of-the-art text-to-speech (TTS) models.
The dataset is designed to support research in speech synthesis, deepfake detection, speech analysis, and comparative evaluation of generative models under a controlled single-speaker setting.

By providing utterance-level alignment between real and synthetic samples, LJ-TTS enables fine-grained comparisons across TTS architectures, isolating synthesis differences without the confounding effect of speaker variability.
The dataset supports systematic analyses across multiple dimensions, including source attribution, phoneme-level acoustic studies, and robustness evaluations of synthetic speech detectors.

---

🌟 Key Features

• Single-speaker design
  Ensures controlled comparisons without multi-speaker variation.

• Real + Synthetic speech pairs
  Each utterance in the REAL folder has corresponding synthesized versions from all TTS systems.

• 11 diverse TTS models
  Spanning both autoregressive and non-autoregressive architectures.

• 1:1 alignment
  Matching filenames and transcriptions across real and synthetic speech enable:

  • deepfake detection
  • spoofing analysis
  • model source tracing
  • perceptual evaluation
  • phoneme-level studies
  • benchmarking and reproducible comparisons

• High-quality data
  Built upon clean recordings of Linda Jonhson (LJSpeech).

---

📁 Dataset Structure

Extract LJ-TTS.zip. Each subfolder (real data folder, plus individual TTS folders) contains audio files with identical filenames, enabling direct pairing.

---

📚 Citation and License

If you use LJ-TTS in your work, please cite:

@misc{negroni2025ljtts,
  title        = {LJ-TTS: A Paired Real and Synthetic Speech Dataset for Single-Speaker TTS Analysis},
  author       = {Negroni, Viola and Salvi, Davide and Comanducci, Luca and Majid Wani, Taiba and Uecker, Madleen and Amerini, Irene and Tubaro, Stefano and Bestagini, Paolo},
  year         = {2025}
}


This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.