SONAR

Benchmark-ready packaging of the SONAR synthetic-audio detection evaluation set (arXiv 2410.04324), for speech anti-spoofing and synthetic / deepfake voice detection.

Overview

SONAR is a binary-classification benchmark: bonafide (genuine human speech) vs. spoof (AI-synthesized speech). The spoof side aggregates clips from eight modern speech-synthesis systems, deliberately spanning architectures and providers so that a detector cannot win by memorising one generator's artifacts. The bonafide side is real clips sampled from the LibriTTS clean-test set. Original code and data: https://github.com/Jessegator/SONAR.

Side	Source	n
bonafide	LibriTTS clean-test	2274
spoof	OpenAI TTS	600
spoof	xTTS	600
spoof	FlashSpeech	118
spoof	VoiceBox	104
spoof	AudioGen	100
spoof	VALL-E	95
spoof	NaturalSpeech 3	32
spoof	PromptTTS 2	25

Seed-TTS excluded (2026-06-07). The upstream seedtts_testset/ directory holds the SeedTTS reference prompts — real Common Voice audio (common_voice_en_<prompt>-<target>.wav), not synthesized speech (the actual SeedTTS-generated outputs are not redistributed on disk). Labeling those 600 real clips as spoof was incorrect, so they are dropped from this package. This matches the SONAR authors' own released eval set (3948 clips).

License & redistribution

This packaging is released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license — see LICENSE.txt. SONAR is an aggregate of audio from multiple synthesis systems and from LibriTTS, each carrying its own upstream terms; the original authors note that the constituent sources "may be subject to different distribution licenses and usage restrictions." This repackaging is provided for non-commercial research use only; the underlying sources retain their own terms, and downstream users should consult the respective upstream providers' usage policies. Audio is re-encoded to 16 kHz mono FLAC for a uniform schema; labels are unmodified.

Schema

Column	Type	Description
path	string	source-relative path (e.g. OpenAI/alloy_69.wav), unique
audio	Audio(16000)	16 kHz mono FLAC (re-encoded from heterogeneous source wav/mp3)
label	ClassLabel	"bonafide" (0) / "spoof" (1)
notes	string	JSON: utterance_id, system, language

notes example:

{"utterance_id": "SONAR_OpenAI_alloy_69", "system": "OpenAI", "language": "en"}

Quick Start

from datasets import load_dataset

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

Stats

Stat	Value
Total trials	3948
Bonafide	2274
Spoof	1674

Source provenance

• Paper: https://arxiv.org/abs/2410.04324
• Code / data: https://github.com/Jessegator/SONAR
• Labels derived from the source directory layout (real_samples/ = bonafide; the eight synthesis-system directories = spoof; seedtts_testset/ excluded — real reference prompts).

Evaluation

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

Citation

@article{li2024sonar,
  title   = {Where are we in audio deepfake detection? A systematic analysis over
             generative and detection models},
  author  = {Li, Xiang and Chen, Pin-Yu and Wei, Wenqi},
  journal = {arXiv preprint arXiv:2410.04324},
  year    = {2024}
}

Maintainer

Contact: k.n.borodin@mtuci.ru