LJS-MOS-120: Human MOS Ratings for 120 Samples of the LJ Speech Dataset

LJS-MOS-120 provides Mean Opinion Score (MOS) ratings for 120 text-to-speech (TTS) samples based on the LJ Speech dataset. Each sample was rated by human annotators for intelligibility and naturalness across four experimental TTS conditions. The dataset follows the Tidy data format, with one row per rating per dimension.

Ratings were collected via Amazon Mechanical Turk (MTurk) in 2022. Each entry includes the worker ID (anonymized), gender (male, female), age group (18–29, 30–49, 50+), and headphone usage.

TTS Conditions

• gt: Ground truth (original human recording)
• wg: Ground truth Mel-spectrogram resynthesized with WaveGlow
• impl: Mel-spectrogram generated using an implicit duration model (Zenodo) and synthesized with WaveGlow
• expl: Mel-spectrogram generated using an explicit duration model (Zenodo) and synthesized with WaveGlow

Dataset Format

Each row contains:

• audio: path to the WAV file
• file: filename (e.g., LJ002-0092)
• condition: one of gt, wg, impl, expl
• score_dimension: intelligibility or naturalness
• score: integer rating (1–5)
• worker_id: anonymized integer ID
• gender, age_group, used_headphone: rater metadata
• assignment_id, hit_id, state: MTurk metadata

License

• LJ Speech: Public Domain (CC0)
• This dataset: MIT License