PACC-P

PACC-P is 5,992 bona fide speech clips under 50 presentation conditions: additive noise, babble and music at five SNRs, simulated rooms and echo, band limiting, pitch and tempo changes, pitch correction, and two compound cafe scenes passed through a codec. Every condition contains the same 5,992 clips as PACC-T, and every clip has a per-clip record logged at generation time.

It contains no synthetic or spoofed speech. It is a reference for what the acoustic path between talker and microphone does to real speech: a bona fide baseline, a robustness benchmark and a condition-shift test set.

Companion dataset: PACC-T (Telecoms), the same 5,992 sources through 34 codecs, 12 tandem chains and 5 resample controls. doi:10.5281/zenodo.23026393

Before you measure

SNR is set against the clean source's speech level

Every SNR is relative to the ITU-T P.56 active speech level of the clean source clip. Noise level is the RMS over 20 ms frames within 40 dB of the loudest frame. Each mixed-in layer is scaled independently against the speech level. Speech is never rescaled on its own: if a finished mix would clip, the whole mix is scaled so its peak is 0.999 of full scale, and that gain is logged per clip (safety_gain_db). Achieved SNR is measured after mixing and logged.

Duration changes break sample alignment

tempo_* and pitch_resample_* change clip length. These conditions are parallel by clip, not by sample: align in time before any frame-by-frame comparison. All other conditions keep the source length.

Band limiting and formants

bandpass_300_3400 and lowpass_2400 remove content above 3.4 and 2.4 kHz. LPC formant trackers configured for wideband speech (Praat's default ceiling is 5500 Hz) fit spurious poles in the empty band, pulling F2 and F3 low with no error flag; on lowpass_2400, F3 is physically absent. Set the analysis ceiling from content bandwidth, or treat F2 and above as unreliable in these conditions.

Cafe conditions are also codec conditions

cafe_voip (Opus 16 kbps, 48 kHz) and cafe_cellular (AMR-WB 23.85 kbps, 16 kHz) mix three layers and then encode. To separate the scene from the codec, compare with cafe_noise_snr15 and music_instrumental_snr20 here, and with the matching codec and resample control in PACC-T.

Native bandwidth differs by pool

AMI sources are native 16 kHz (content to 8 kHz); VCTK sources are 48 kHz. Outputs keep the source rate except the cafe conditions. Compare within a pool, or account for this when pooling.

Sources

The same 5,992 source clips as PACC-T, shipped once in pacc_base.tar with per-source metadata.

VCTK mic pairing: 1,485 utterances appear in both mic pools; 15 in each pool are unpaired. Joining mic1 to mic2 on utterance ID gives 1,485 pairs, not 1,500. All sources are mono 16-bit PCM. AMI file IDs encode meeting, channel, segment index and start/end time in seconds.

Per-source metadata ships as sources/metadata.csv in pacc_base.tar: speaker, gender and where the label came from, VCTK age and utterance, AMI meeting, channel, participant and segment times.

Gender balance

Sources were selected for gender balance. VCTK is balanced exactly. AMI is not: its clips were selected using gender inferred from the audio (pitch and apparent vocal-tract length), and those labels proved wrong for 17% of clips, mostly men labelled as women (441 clips, against 68 the other way). Checked against AMI's own participant metadata, the AMI pool is 62% male. The gender column gives the metadata label; the inferred label is kept, for transparency only, as gender_inferred_not_recommended. For TS meetings, which have no entry in AMI's participants.xml, sex is taken from the M/F prefix of the participant ID, which agrees with participants.xml for all 189 participants listed there.

8 AMI clips with no usable speech are excluded from all of PACC (excluded_sources.csv). Earlier internal builds also held RAVDESS and CREMA-D clips; they were removed so that PACC is CC BY only.

Conditions

How the material was made

• Rooms: pyroomacoustics 0.10.1, image source plus ray tracing, air absorption, speaker to microphone 1.0 m at 1.5 m height, 4 layouts per room (one per clip, chosen by seed). Small 4.0 x 3.5 x 2.7 m (RT60 0.3 s), medium 7.0 x 5.0 x 3.0 m (0.6 s), large 12.0 x 9.0 x 4.0 m (1.0 s). The room impulse responses ship in pacc-p_room_cache.tar. Each reverb clip's params row names its room variant and the variant's rir_sha256; convolving the source with that response (recipe in the tarball's README) recreates the stored clip bit for bit.
• Babble: 6 English talkers (3 female, 3 male) from MUSAN's LibriVox speech, each at its own position in the medium room.
• Scenes: FSD50K clips, a continuous bed plus timed events at +6 dB relative to the bed. Cafe: tableware, pouring, footsteps, doors, light kitchen sounds, distant traffic, no voices, 0.4 events/s, in the medium room. Environment: home, office, street and in-car background, no voices, 0.3 events/s, no room.
• Music: MUSAN music (CC BY and public-domain tracks only), played through a loudspeaker in the medium room. The cafe's music layer is drawn exactly like music_instrumental_snr20.
• Echo: one copy delayed 150 or 300 ms at -10 or -20 dB.
• Filters: Butterworth bandpass 300-3400 Hz (design order 4) and lowpass 2400 Hz (order 8).
• Pitch: PSOLA keeps formants (Praat 6.1.38 via parselmouth 0.4.7, pitch 75-600 Hz); resampling moves formants with F0 (librosa 0.11.0, soxr_hq). Autotune snaps F0 to the nearest equal-tempered semitone (A4 = 440 Hz), retune 0 ms (hard) or 50 ms (natural).
• Tempo: ffmpeg 8.1 atempo, pitch unchanged.
• Randomness: every clip has its own seed, sha256 of 20260916|<condition>|<file_id>, so every draw (noise clip, room layout, event timing) is reproducible.
• Software: Python 3.11.9, numpy 1.24.4, scipy 1.15.3, soundfile 0.13.1, librosa 0.11.0, soxr 1.0.0, praat-parselmouth 0.4.7, pyroomacoustics 0.10.1, ffmpeg 8.1.

Files

All audio is 16-bit mono FLAC. SHA256SUMS and TARBALL_INDEX.csv list every tarball; pacc-p_manifest.csv lists every file with its sha256. Tarballs are deterministic.

Per-clip records

Each condition's params.csv has one row per clip, logged when the clip was generated: input and output sample rate and length, source speech level and peak, every target and achieved SNR, layer gains, safety gain, the near-floor flag, codec and codec length error for the cafe conditions.

Known edge cases

See EDGE_CASES.txt: the cafe conditions' combined-SNR tolerance (three clips deviate by more than 0.5 dB), rows near the 16-bit quantisation floor, inherited AMI clipping, and the excluded sources.

Motivation

The design of this corpus was motivated by Delgado et al. (ICASSP 2026), who argue that deepfake detection must account for how audio is presented through real communication channels, and by Lee et al. (ICASSP 2026), whose noise-aware multi-LoRA framework highlights real-world noise conditions. Neither group was involved in producing this dataset.

• H. Delgado, G. Ramondetti, E. Dalmasso, G. Karvitsky, D. Colibro, H. Talib. "On Deepfake Voice Detection - It's All in the Presentation." ICASSP 2026. arXiv:2509.26471.
• W. Lee, H. Dinh-Xuan, T.-P. Doan, S. Jung. "Dynamic Noise-Aware Multi LoRA Framework Towards Real-World Audio Deepfake Detection." ICASSP 2026.

Licence and credits

PACC-P is released under CC BY 4.0 by Christopher Kleingertner (Moonscape Software). It is derived from:

• CSTR VCTK Corpus 0.92. J. Yamagishi, C. Veaux, K. MacDonald. University of Edinburgh, CSTR, 2019. doi:10.7488/ds/2645. CC BY 4.0.
• AMI Meeting Corpus. University of Edinburgh et al. https://groups.inf.ed.ac.uk/ami/corpus/. CC BY 4.0. J. Carletta (2006), "Announcing the AMI Meeting Corpus", ELRA Newsletter 11(1).
• MUSAN. D. Snyder, G. Chen, D. Povey (2015), arXiv:1510.08484. https://www.openslr.org/17/. CC BY 4.0; only CC BY and public-domain files used.
• FSD50K. E. Fonseca et al. (2022), IEEE/ACM TASLP 30. doi:10.5281/zenodo.4060432. CC BY; only CC0 and CC BY clips used.

Third-party clips keep their own licences; per-clip credits are in attribution.csv (4,696 clips: CC0, CC BY 3.0/4.0 and public domain). Please credit these sources alongside PACC-P.

Citation

@dataset{moonscape_pacc_p_2026,
  author    = {Kleingertner, Christopher and {Moonscape Software}},
  title     = {{PACC-P: Parallel Acoustic Confound Corpus, Presentation}},
  year      = {2026},
  version   = {1.0},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.23026395}
}