lab260/mos260

Russian TTS MOS Evaluation Dataset

A curated Russian synthesized speech dataset with crowdsourced quality ratings.

Overview

This dataset contains Russian synthesized speech samples from multiple TTS systems and audio codec models, evaluated by human annotators via crowdsourcing. Quality ratings are provided in two dimensions: perceptual quality (MOS) and intelligibility (Int-MOS).

• Language: Russian only
• Sources: F5, FishSpeech, GPT-So-VITS, Tortoise, XTTS, EnCodec, WavTok, and internal encoder/decoder configurations
• Format: Single Parquet file with embedded audio bytes
• Access: Public

Usage

Load the dataset

from datasets import load_dataset

ds = load_dataset("YOUR_ORG/YOUR_DATASET_NAME")

Or just use script load_data_tts.py

Read audio from bytes

import io
import soundfile as sf

sample = ds["train"][0]

audio_array, sr = sf.read(io.BytesIO(sample["audio"]))

print(f"File:     {sample['filename']}")
print(f"MOS:      {sample['mos_mean']}")
print(f"Int-MOS:  {sample['int_mos_mean']}")

Data Structure

Each row corresponds to one audio file with aggregated scores from all annotators.

Column	Type	Description
inner_audio_id	int	Internal audio identifier from the labeling system
filename	string	Relative path to the audio file (e.g. F5/001234_RUSLAN.wav)
audio	binary	Raw audio bytes (wav)
mos_mean	float or null	Mean perceptual quality score (1–5); null if not rated
int_mos_mean	float or null	Mean intelligibility score (1–5); null if not rated
project	string	Source labeling project name

Evaluation Methodology

Ratings were collected via crowdsourcing across two independent evaluation tracks:

• MOS — perceptual quality assessment following ITU-T P.800 (1 = bad, 5 = excellent)
• Int-MOS — intelligibility assessment on the same 1–5 scale

Each sample may have one or both scores depending on which evaluation projects it was included in. A sample appearing in both tracks may carry two different inner_audio_id values — this is expected behavior from the labeling pipeline.

Model quality on unlabeled data was additionally estimated using Proxymos as a reference-free MOS predictor.

Contact

• Email: kborodin.research@gmail.com
• Telegram: @korallll_ai