asosawelford/es-TTS-subjective-naturalness

"A Dataset for Automatic Assessment of TTS Quality in Spanish" - Interspeech 2025

This dataset provides a collection of Spanish text-to-speech (TTS) audio samples with human naturalness ratings, aimed at advancing research on automatic TTS quality assessment in Spanish.

Dataset Description

The dataset contains samples from 52 different speakers, including 12 TTS systems and 6 real human voices. The samples cover multiple Spanish dialects, speaker genders, and speech synthesis methods.

To collect subjective labels, 92 native Spanish speakers rated the samples following the ITU-T Rec. P.807 standard on a 5-point Mean Opinion Score (MOS) scale, where 5 is completely natural and 1 is completely unnatural.

Dataset Statistics

Statistic	Value
Number of ratings	4,326
Number of speakers	52 (12 TTS + 6 human)
Average duration	3.5 seconds
Dialects	Rioplatense, Castilian, Central American
Raters	92 Spanish native speakers

Data Collection

• TTS systems include neural, concatenative, parametric, and proprietary models covering various dialects.
• Real human samples serve as high-quality references.
• Two data augmentation methods (Vocal Tract Length Perturbation and Griffin-Lim phase alteration) were applied to increase diversity. They are identified by "_VTLP" and "_GL" suffixes.

Usage

You can load this dataset easily with the 🤗 Hugging Face datasets library:

from datasets import load_dataset

dataset = load_dataset("asosawelford/es-TTS-subjective-naturalness")

print(dataset["train"][0])

Citation

If you use this dataset, please cite the following paper:

A Dataset for Automatic Assessment of TTS Quality in Spanish Alejandro Sosa Welford, Leonardo Pepino https://arxiv.org/abs/2507.01805

---

tags: - speech - tts - spanish - mos - naturalness - speech-quality - deep-learning