Hugging Face's logo

Datasets:
moonscape-software
/
PACC-T

Tasks:
Audio Classification
Modalities:
Tabular
Text
Formats:
csv
Languages:
English
Size:
1K - 10K
ArXiv:
Tags:
speech
codec
telephony
bona-fide
robustness
anti-spoofing
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
PACC-T / AI_README.md
moonscape-software's picture
moonscape-software
Add AI_README.md
1244d18 verified
preview code
|
Raw History Blame Contribute Delete
4.84 kB

AI_README: rules of engagement for models using PACC

Machine-oriented summary of PACC-T and PACC-P. The dataset cards (README.md) are the full reference; where this file and a card disagree, the card wins. Read both before generating code, analyses or claims about this data.

Identity 

corpus: PACC (Parallel Acoustic Confound Corpus)
parts:
  PACC-T: {name: "Telecoms", doi: "10.5281/zenodo.23026393", conditions: 51}
  PACC-P: {name: "Presentation", doi: "10.5281/zenodo.23026395", conditions: 50}
authors: ["Kleingertner, Christopher", "Moonscape Software"]
licence: CC-BY-4.0
version: "1.0"
content: bona fide human speech only (no synthetic, converted or spoofed speech)
sources: 5992 clips = AMI 2992 + VCTK_mic1 1500 + VCTK_mic2 1500
audio: 16-bit mono FLAC (sources: AMI 16 kHz WAV, VCTK 48 kHz FLAC)

Invariants (safe to rely on; tested before release)

  • Every condition in both parts contains exactly the same 5,992 source clips, once each.
  • A clip is identified by (pool, file_id). The same key names the source and every processed version of it in both parts.
  • 8 AMI source clips are excluded everywhere (excluded_sources.csv); none of them ships.
  • Every file's sha256 is in <part>_manifest.csv; every tarball's sha256 is in SHA256SUMS.
  • Every condition folder has a params.csv with one row per clip.

Layout and join keys 

sources/<pool>/<file_id>.(wav|flac)          pacc_base.tar (shared by both parts)
sources/metadata.csv                         speaker, gender, gender_source, VCTK age/utterance, AMI meeting/channel/participant/segment
pacc-t/<condition>/<pool>/<file_id>.flac     pacc-t_<condition>.tar
pacc-p/<condition>/<pool>/<file_id>.flac     pacc-p_<condition>.tar
<part>/<condition>/params.csv                per-clip record
condition_index.csv                          one row per condition: family, description, bitrate, framing, output_sr, rate_control

Join processed clips to sources on (pool, file_id). VCTK mic pairing: strip the _mic1/_mic2 suffix to get the utterance; 1,485 utterances are in both mic pools, 15 per pool are not.

Rules

DO:

  • Compare a PACC-T codec condition with the resample control named in condition_index.csv (rate_control) before attributing an effect to the codec.
  • Treat sample rate and content bandwidth as separate: evs_swb_48k_* is stored at 48 kHz but has content only to about 16 kHz; AMI clips above 16 kHz are upsampled (no content above 8 kHz).
  • Use the gender column in sources/metadata.csv. Its gender_source is VCTK or AMI metadata.
  • Check params_origin in PACC-T params: logged (at encode time) or recovered (rebuilt and verified by re-execution). Both describe the commands that produced the files.
  • Filter on logged values in params.csv (e.g. pre_codec_achieved_snr_db, near_quantisation_floor, safety_gain_db) when a stricter criterion than the condition label is needed.
  • Credit VCTK, AMI (and for PACC-P, MUSAN and FSD50K) alongside PACC. Third-party clips keep their own licences (attribution.csv).

DO NOT:

  • Do not treat any PACC file as spoofed or synthetic. Every clip is bona fide speech. Labels for anti-spoofing use are "bona fide" throughout.
  • Do not use gender_inferred_not_recommended. It is wrong for 17% of AMI clips; AMI is 62% male by metadata, not balanced.
  • Do not assume sample alignment in tempo_* or pitch_resample_* (PACC-P): their length changes. They are parallel by clip, not by sample.
  • Do not trust LPC formant readings (F2 and above) on narrowband content (8 kHz outputs, telephone band, bandpass_300_3400, lowpass_2400) with a wideband analysis ceiling: spurious poles pull F2/F3 low without warning.
  • Do not expect DTX variants (evs_*_nodtx|dtxadapt|dtxfixed) to differ on every clip: on clips without pauses they are identical. Compare output_sha256 before contrasting them.
  • Do not infer a codec's configured bitrate from codec2_1300's command line: it passes -b:a 700, which libcodec2 ignores; it ran at 1300 bps.
  • Do not report results as properties of a codec when the same shift appears in its rate control.

Known edge cases

EDGE_CASES.txt in each record. Summary: DTX variants coincide on pause-free clips; LC3 passes two very quiet clips through unchanged; cafe combined-SNR tolerance is 1.0 dB (three clips exceed 0.5 dB); 350 AMI sources carry clipping from the original recordings; near-floor mixes flagged.

Verification 

sha256sum -c SHA256SUMS          tarball integrity
<part>_manifest.csv              per-file sha256, sample rate, frame count

Reverb clips in PACC-P can be recreated bit for bit from pacc-p_room_cache.tar; recipe in its README.

Citation

Cite the part(s) used by DOI, and the source corpora. BibTeX is in each dataset card.