This dataset quantifies the effect of 35 codec conditions on 80+ acoustic features extracted from 7,500 biological speech clips. It answers the question: "Which acoustic features survive telecommunications codec compression, and which are destroyed?"

The corpus is the empirical foundation for channel-aware gate calibration in deepfake audio detection. Every threshold in the SSA Gate 0 channel triage architecture is derived from measurements in this dataset.

262,463 rows | 35 codec conditions | 7,500 source clips | 4 speech corpora

Key Finding

Phase-smear decoupling (bico_f0_f1) survives CELP codec compression at 96.3%.

This is counterintuitive — CELP codecs (AMR-NB, GSM, G.711) completely destroy and parametrically reconstruct the waveform. Yet the first formant bicoherence is almost entirely preserved.

The physical reason: bico_f0_f1 measures nonlinear quadratic phase coupling established at glottal closure. CELP codec operations are entirely linear. Linear operations cannot create or destroy nonlinear phase relationships.

Contrast with modgd_var (spectral phase complexity — a linear property):

Codec Family	bico_f0_f1	modgd_var	bico retained	modgd retained
Source (biological)	0.465	5.332	100%	100%
EVS SWB 48kbps	0.456	5.362	98.1%	100.6%
Opus 32kbps	0.465	5.361	99.9%	100.5%
AMR-NB 4.75kbps	0.446	2.697	95.8%	50.6%
GSM 13kbps	0.446	2.804	95.8%	52.6%
G.711 A-Law	0.447	2.686	96.1%	50.4%
Codec2 700bps	0.475	2.770	102.1%	51.9%

Implication for deployment: bico_f0_f1-based deepfake detection works on telephony audio, VoIP intercepts, and any channel condition. The detection boundary is architectural — synthetic speech never had a biological glottis — not environmental.

Dataset Structure

Format: Long — one row per clip per codec condition. Use codec_condition column to filter to a specific codec.

import pandas as pd
df = pd.read_parquet('ssa_codec_degradation_study.parquet')

# Compare CELP vs modern codecs on primary detection signal
df.groupby('is_celp')['bico_f0_f1'].mean()

# Full degradation curve for modgd_var
df.groupby('codec_condition')['modgd_var'].mean().sort_values()

# All AMR-NB 4.75kbps clips
amr = df[df['codec_condition'] == 'amr_nb_475']

Source Corpora

Pool	Corpus	N clips	Licence	Recording conditions
VCTK_mic1	VCTK 0.92 (MKH800)	1,500	CC-BY-4.0	Anechoic studio, high-bandwidth
VCTK_mic2	VCTK 0.92 (AKG C535)	1,500	CC-BY-4.0	Anechoic studio, standard mic
AMI	AMI Meeting Corpus	3,000	CC-BY-4.0	Spontaneous conversational
CREMA-D	CREMA-D	1,500	ODC-BY	Emotional speech, controlled

RAVDESS excluded: CC-BY-NC-SA-4.0 licence would propagate NC to entire dataset. RAVDESS-derived results available in the NC variant of this release.

All clips stratified by gender, source corpus, and emotional valence to ensure biological diversity across the codec degradation curves.

Codec Conditions

Single Codec (27 conditions)

Condition	Family	Nom. Ceiling	Standard
source	Biological	—	Unencoded reference
evs_swb_48k	EVS-SWB	16kHz	3GPP TS 26.445
evs_swb_nodtx	EVS-SWB	16kHz	EVS without DTX
evs_24400	EVS	16kHz	3GPP R12
evs_24400_nodtx	EVS	16kHz	EVS 24.4k without DTX
evs_9600	EVS	8kHz	3GPP R12
evs_9600_nodtx	EVS	8kHz	EVS 9.6k without DTX
opus_32k	Opus-CELT	16kHz	IETF RFC6716
opus_16k	Opus-SILK	16kHz	IETF RFC6716
opus_6k	Opus-SILK	8kHz	IETF RFC6716
lc3	LC3	16kHz	Bluetooth LE Audio
aac_64k	Perceptual	16kHz	MPEG-4
aac_32k	Perceptual	16kHz	MPEG-4
mp3_128k	Perceptual	16kHz	MPEG-1 Layer III
mp3_32k	Perceptual	16kHz	MPEG-1 Layer III
g722	SB-ADPCM	7kHz	ITU-T G.722
amr_wb	CELP-WB	7kHz	3GPP AMR-WB
g726_32k	ADPCM	4kHz	ITU-T G.726
g726_24k	ADPCM	4kHz	ITU-T G.726
g726_16k	ADPCM	4kHz	ITU-T G.726
amr_nb_122	CELP	4kHz	3GPP AMR-NB 12.2kbps
amr_nb_475	CELP	4kHz	3GPP AMR-NB 4.75kbps
gsm	CELP	4kHz	ETSI GSM 13kbps
ilbc	CELP	4kHz	IETF RFC3951
speex_8k	CELP	4kHz	Xiph Speex 8kbps
g711_ulaw	PCM-Comp	4kHz	ITU-T G.711 μ-Law
g711_alaw	PCM-Comp	4kHz	ITU-T G.711 A-Law
codec2_700	CELP	4kHz	FreeDV Codec2 700bps

Tandem Chains (7 conditions)

Multi-hop codec degradation simulating real-world transmission paths:

Condition	Chain	Final ceiling
tandem_opus_32k_to_evs_24400	Opus 32k → EVS 24.4k	16kHz
tandem_evs_24400_to_amr_wb	EVS 24.4k → AMR-WB	7kHz
tandem_opus_32k_to_amr_wb	Opus 32k → AMR-WB	7kHz
tandem_amr_wb_to_g711_ulaw	AMR-WB → G.711	4kHz
tandem_evs_24400_to_amr_nb_475	EVS 24.4k → AMR-NB 4.75k	4kHz
tandem_evs_24400_to_g711_ulaw	EVS 24.4k → G.711	4kHz
tandem_opus_32k_to_g711_ulaw	Opus 32k → G.711	4kHz

Gate Calibration Reference

Empirically derived thresholds from this dataset:

Feature	Source mean	CELP mean	Suggested threshold	Gate use
modgd_var	5.332	2.960	4.0	Gate 0: below = CELP confirmed
bico_f0_f1	0.465	0.448	N/A	Channel-agnostic — no gate needed
codec_cutoff_hz	4393	2336	3000	Provenance: below = NB ceiling
f1_velocity	77.3	50.7	64.0	Degrades under CELP
spectral_floor_var	0.041	0.001	0.020	Dead room / digital vacuum gate
pause_cv	0.850	0.859	N/A	Preserved — structural, not spectral

Citation

If you use this dataset please cite:

@dataset{moonscape_ssa_codec_2026,
  title     = {SSA Codec Degradation Study — Acoustic Feature Exports},
  author    = {Moonscape Software},
  year      = {2026},
  publisher = {HuggingFace},
  note      = {Export version SSA\_CodecStudy\_v1\_2026}
}

and cite the source corpora:

VCTK: Yamagishi et al. (2019)
AMI: Carletta et al. (2005)
CREMA-D: Cao et al. (2014)

Licence

CC-BY-4.0 — source corpora are VCTK (CC-BY-4.0), AMI (CC-BY-4.0), CREMA-D (ODC-BY). Commercial use permitted with attribution. Academic and commercial licences available from Moonscape Software.

SSA Codec Degradation Study v1 | Moonscape Software | 2026 Export version: SSA_CodecStudy_v1_2026 For full data dictionary see SSA_CODEC_STUDY_DATA_DICTIONARY.md

Downloads last month: 13

Total file size:

58.7 MB