Datasets:

SpeechAntiSpoofingBenchmarks
/

PyAra

license: cc-by-nc-sa-4.0
language:
  - ru
pretty_name: PyAra — Russian Audio Deepfake Detection
task_categories:
  - audio-classification
size_categories:
  - 100K<n<1M
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test-*.parquet
tags:
  - anti-spoofing
  - audio-deepfake-detection
  - speech
  - benchmark
  - arena-ready
  - russian
arxiv: []

PyAra — Russian Audio Deepfake Detection

Benchmark-ready packaging of PyAra, a Russian-language audio anti-spoofing dataset (source: https://www.kaggle.com/datasets/alep079/pyara).

Overview

Binary classification: bonafide (genuine Russian human recordings) vs. spoof (synthetic / converted speech). The spoof side is produced by five generation algorithms (alg_1 … alg_5). Label is the top-level source directory (Real/ → bonafide, Fake/ → spoof), per the fake column of final_dataset.tsv.

License & redistribution

Released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0) license. Redistribution and adaptation are permitted for non-commercial use with attribution and ShareAlike; this packaging is itself CC BY-NC-SA 4.0. See LICENSE.txt. Audio is the original 16 kHz mono WAV, embedded bit-exactly (no re-encode — a full decode probe of all clips passed cleanly).

Schema

Column	Type	Description
`path`	`string`	source-relative path, e.g. `Fake/alg_1_0.wav`, unique
`audio`	`Audio(16000)`	16 kHz mono WAV
`label`	`ClassLabel`	`"bonafide"` (0) / `"spoof"` (1)
`notes`	`string`	JSON: `utterance_id`, `algorithm`, `gender`, `age`, `length`

notes example:

{"utterance_id": "alg_1_0", "algorithm": "alg_1", "gender": "male", "age": "twenties", "length": "5.55"}

utterance_id is the source filename stem (real_12713, alg_1_0) — unique across the whole set.

Quick Start

from datasets import load_dataset

ds = load_dataset("SpeechAntiSpoofingBenchmarks/PyAra", split="test")
print(ds[0])

Stats

Stat	Value
Total trials	201,778
Bonafide (real)	73,583
Spoof	128,195
Spoof algorithms	alg_1 (46,505), alg_2 (10,065), alg_3 (11,165), alg_4 (25,412), alg_5 (35,048)
Language	Russian
Sample rate	16 kHz mono

Source provenance

Dataset: PyAra, https://www.kaggle.com/datasets/alep079/pyara
No associated academic paper.

Evaluation

For evaluation instructions and submission format, see submissions/README.md.

Citation

PyAra — Russian Audio Deepfake Detection dataset.
https://www.kaggle.com/datasets/alep079/pyara

Maintainer

Contact: k.n.borodin@mtuci.ru