SAS-CF: Scene Audio Spoofing — Codec Fake

SAS-CF is an environmental sound codec-fake dataset for anti-spoofing and deepfake detection research. It applies the same 15 neural audio codec models from CodecFake (Wu et al., Interspeech 2024) to re-synthesise real-world acoustic scenes from the TAU Urban Acoustic Scenes 2019 dataset.

Each recording is passed through a codec encoder → decoder pipeline, producing a re-synthesised version that preserves the acoustic content but carries codec-specific compression artifacts — the same approach used to build the original speech-based CodecFake dataset.

---

Dataset Structure

SAS-CF/
└── env_fake/          ← 226,635 fake recordings (15 codecs × 15,109 files)
    ├── A/     SpeechTokenizer
    ├── B1/    AcademiCodec hifi_16k_320d
    ├── B2/    AcademiCodec hifi_16k_320d_large_uni
    ├── B3/    AcademiCodec hifi_24k_320d
    ├── C/     AudioDec 24k
    ├── D1/    DAC 16k
    ├── D2/    DAC 24k
    ├── D3/    DAC 44k
    ├── E/     EnCodec 24k
    ├── F1/    FunCodec en_libritts_16k_gr1nq32ds320
    ├── F2/    FunCodec en_libritts_16k_gr8nq32ds320
    ├── F3/    FunCodec en_libritts_16k_nq32ds320
    ├── F4/    FunCodec en_libritts_16k_nq32ds640
    ├── F5/    FunCodec zh_en_16k_nq32ds320
    └── F6/    FunCodec zh_en_16k_nq32ds640

Columns:

Column	Type	Description
audio	Audio	Re-synthesised recording (48 kHz stereo, 10 s)
scene	string	Acoustic scene class (airport, bus, metro, …)
codec	string	Codec ID: A, B1, …, F6
label	int8	0 = fake (all samples here are codec-re-synthesised)

Note on label: All samples in env_fake/ are fake (label = 0). The label column exists for compatibility with anti-spoofing frameworks and for future dataset versions that will include real recordings (label = 1).

---

Loading the Dataset

from datasets import load_dataset

# Load one codec
ds = load_dataset("ggirishg/SAS-CF", "E")

# Load all codecs
ds = load_dataset("ggirishg/SAS-CF")

# Stream without downloading everything
ds = load_dataset("ggirishg/SAS-CF", "A", streaming=True)

---

Source Data

TAU Urban Acoustic Scenes 2019 Openset

• 15,110 × 10-second stereo WAV recordings at 48 kHz
• 10 scene classes: airport, bus, metro, metro_station, park, public_square, shopping_mall, street_pedestrian, street_traffic, tram
• 710 additional "unknown" open-set recordings from DCASE 2017

---

Codec Models (CodecFake Table 1)

ID	Framework	Model	SR (kHz)	Codebooks
A	SpeechTokenizer	SpeechTokenizer	16	8
B1	AcademiCodec	hifi_16k_320d	16	4
B2	AcademiCodec	hifi_16k_320d_large_uni	16	4
B3	AcademiCodec	hifi_24k_320d	24	4
C	AudioDec	24k_320d	24	8
D1	DAC	16k	16	12
D2	DAC	24k	24	32
D3	DAC	44k	44.1	9
E	EnCodec	24k	24	8
F1	FunCodec	en_libritts_16k_gr1nq32ds320	16	32
F2	FunCodec	en_libritts_16k_gr8nq32ds320	16	32
F3	FunCodec	en_libritts_16k_nq32ds320	16	32
F4	FunCodec	en_libritts_16k_nq32ds640	16	32
F5	FunCodec	zh_en_16k_nq32ds320	16	32
F6	FunCodec	zh_en_16k_nq32ds640	16	32

---

Statistics

• Total files: 226,635 (15 codecs × 15,109 files each)
• Source files: 15,109 usable (1 corrupt file skipped consistently across all codecs)
• Duration: ~627 hours total (15,109 × 10 s × 15 codecs)
• No predefined split — all files are provided as a single unsplit collection; users apply their own scene-based partitioning via the scene column

---

Intended Use

• Training and evaluating audio deepfake detection models on environmental sounds
• Cross-domain evaluation: models trained on speech fakes (CodecFake) tested on scene fakes (SAS-CF)
• Studying whether codec compression artifacts are modality-agnostic
• Suggested baseline detector: AASIST-L (clovaai/aasist)

---

Generation

Dataset generated using the Neural-Codecs pipeline on Kelvin2 HPC (NVIDIA A100 GPUs).

---

Citation

If you use SAS-CF, please also cite the original CodecFake paper:

@inproceedings{wu2024codecfake,
  title     = {CodecFake: Enhancing Anti-Spoofing Models Against Deepfake Audios from Codec-Based Speech Synthesis Systems},
  author    = {Wu, Haibin and Tseng, Yuan and Lee, Hung-yi},
  booktitle = {Interspeech},
  year      = {2024}
}

---

License

CC BY 4.0 — Environmental source audio from TAU Urban Acoustic Scenes 2019 used under its original license.