SAS-CF: Speech-Acoustic Scene CodecFake Dataset

SAS-CF is a large-scale benchmark dataset for anti-spoofing in real-world acoustic environments. It combines codec-faked speech and environmental audio to create realistic mixed conditions for training and evaluating spoofing detection models.

Dataset Summary

Condition	Description	Files	Hours	Label
M_rr	Real speech + Real env	15,109	42	1 (genuine)
M_ff	Fake speech + Fake env (same codec)	226,635	630	0 (spoof)
M_rf	Real speech + Fake env	226,635	630	1 (genuine)
M_fr	Fake speech + Real env	226,635	630	0 (spoof)
m_f	Codec-applied full mix C(M_rr)	226,635	630	0 (spoof)
env_fake	Fake environmental audio (source)	226,635	630	0 (spoof)
Total		~1,147,784	~3,192

Conditions Explained

Source Components

• s^r: Real speech — VCTK corpus (~110 speakers, 44,242 utterances)
• s^f: Fake speech — s^r passed through each of 15 neural codecs
• e^r: Real environment — TAU Urban Acoustic Scenes 2019 (open-set), 15,109 files, 10 scene classes
• e^f: Fake environment — e^r passed through each of 15 neural codecs

Seen Mix Conditions

Symbol	Construction	Acoustic Content
M_rr	mix(s^r, e^r)	Genuine speech in real environment
M_ff	mix(s^f_X, e^f_X)	Codec-X fake speech in codec-X fake environment
M_rf	mix(s^r, e^f_X)	Genuine speech in codec-X fake environment
M_fr	mix(s^f_X, e^r)	Codec-X fake speech in real environment

Unseen Joint-Mixture Condition

Symbol	Construction	Description
m^f	codec(M_rr)	Entire real mix passed through codec — both speech and env are degraded simultaneously

Key distinction: In seen conditions (M_ff), fake components are generated before mixing. In m^f, the real mix is faked after mixing — the codec acts on the combined signal, creating unseen artifact patterns.

Mixing Protocol

All mixes follow a consistent procedure:

1. Resample both streams to 16 kHz mono
2. Normalise speech to −20 dBFS, environment to −30 dBFS
3. Add and peak-normalise the mixture
4. Trim/pad to 10-second clips

Neural Codecs (15 total)

ID	Codec	Sample Rate
A	SpeechTokenizer	16 kHz
B1	DAC 16 kHz	16 kHz
B2	DAC 24 kHz	24 kHz
B3	DAC 44 kHz	44 kHz
C	AudioDec (libritts_v1)	24 kHz
D1	AcademiCodec HiFi 16k 320d	16 kHz
D2	AcademiCodec HiFi 24k 320d	24 kHz
D3	AcademiCodec HiFi 16k large	16 kHz
E	EnCodec 24 kHz	24 kHz
F1	FunCodec en gr1 16k	16 kHz
F2	FunCodec en gr8 16k	16 kHz
F3	FunCodec en nq32ds320 16k	16 kHz
F4	FunCodec en nq32ds640 16k	16 kHz
F5	FunCodec zh-en nq32ds320 16k	16 kHz
F6	FunCodec zh-en nq32ds640 16k	16 kHz

Acoustic Scenes (10 classes)

airport, bus, metro, metro_station, park, public_square, shopping_mall, street_pedestrian, street_traffic, tram

Schema

Each row contains:

{
  "audio":     {"bytes": bytes, "path": str},  # HF Audio-compatible
  "scene":     str,   # acoustic scene label
  "codec":     str,   # codec letter (A, B1, ..., F6) or "real" for M_rr
  "label":     int8,  # 1 = genuine, 0 = spoof
  "condition": str,   # M_rr / M_ff / M_rf / M_fr / m_f / env_fake
}

Usage

from datasets import load_dataset

# Load a specific condition
ds = load_dataset("ggirishg/SAS-CF", "M_rr")
ds = load_dataset("ggirishg/SAS-CF", "M_ff")
ds = load_dataset("ggirishg/SAS-CF", "m_f")

# Access audio
sample = ds["all"][0]
audio_array = sample["audio"]["array"]
label = sample["label"]   # 1=genuine, 0=spoof
scene = sample["scene"]   # e.g. "airport"
codec = sample["codec"]   # e.g. "E"

Source Datasets

• Speech: VCTK Corpus
• Environment: TAU Urban Acoustic Scenes 2019 (open-set)
• Codec implementations: SpeechTokenizer, DAC, AudioDec, AcademiCodec, EnCodec, FunCodec

Citation

If you use SAS-CF in your research, please cite:

@dataset{madaan2026sascf,
  title   = {SAS-CF: Speech-Acoustic Scene CodecFake Dataset},
  author  = {Madaan, Girish},
  year    = {2026},
  url     = {https://huggingface.co/datasets/ggirishg/SAS-CF}
}

License

CC BY 4.0