vadette/macro_prosody_sample_set

Alexandria Voice Corpus — Multilingual Macro-Prosody Telemetry

Version 1.1 — Replacement release

This pack supersedes the earlier Korean & Hindi two-language release. That release was built on a pipeline with several unresolved quality-gate bugs (documented below). This version corrects all known issues and expands to seven typologically diverse languages.

No audio is included. This is a structured acoustic feature dataset for linguistic research, speech technology, and cross-linguistic prosody analysis.

---

What changed from the previous release

The earlier korean_cv24 and hindi_cv24 files in the two-language pack were generated before three gate-level bugs were resolved:

Bug	Effect on data
BUG-029 — Missing C50 floor on T2 gate	High-reverb recordings admitted to STUDIO tier. Clips recorded in tiled bathrooms or reverberant rooms were incorrectly marked T2, inflating that tier and contaminating features like F0 smoothness and spectral tilt.
BUG-030 — Missing speech-ratio floor on T2 gate	Sparse/silence-heavy clips admitted to STUDIO tier. Clips where >70% of the audio was ambient room tone were stamped T2. This inflated articulation_rate and skewed nPVI values.
High-SNR branch C50 fix	The fast path for very loud recordings (SNR ≥ 35 dB) also lacked the C50 floor, so a loud reverberant recording could bypass the room quality check entirely.

All three are fixed in this release. The full T2 gate now requires simultaneously: SNR ≥ 25 dB and C50 ≥ 20 dB and speech_ratio ≥ 0.30. Tier labels have been retroactively corrected across all ledgers.

If you downloaded the previous pack, please replace both files.

---

Dataset Details

What is macro-prosody telemetry?

Macro-prosody refers to the suprasegmental properties of speech — pitch contour, rhythm, intensity, voice quality — measured at the utterance level. Each row in this dataset is one spoken clip with 30+ acoustic features extracted from it.

This is distinct from transcription, phoneme alignment, or word-level data. It is designed for population-level acoustic analysis, cross-linguistic typology research, and training prosody-aware speech models.

---

Language Coverage

Language	Code	Family	Corpus	Speech type	Clips	T1%	MetaGender%	Notes
Korean	ko	Koreanic (isolate)	CV24	Scripted	6,805	68%	61%	Replacement for v1.0
Hindi	hi	Indo-Aryan	CV24	Scripted	18,435	67%	55%	Replacement for v1.0
Hebrew	he	Semitic	CV24	Scripted	5,455	73%	93%	Strong demographic metadata; heavily male-skewed (see notes)
Manx	gv	Celtic (Goidelic)	CV24	Scripted	6,579	98%	0%	Near-extinct revival language; near-entirely PRISTINE quality
Tzeltal	tzh	Mayan	CV24	Scripted	5,585	57%	68%	All meta-gender records are female (see notes)
Maguindanao	mdh	Austronesian	SPS2	Spontaneous	5,536	33%	0%	No TSV demographics; gender is pitch-inferred only
Lasi	lss	Indo-Iranian (Sindhi)	SPS2	Spontaneous	10,435	59%	0%	Large spontaneous corpus; no TSV demographics

Total: 58,830 clips (395 suppressed by k-anonymity, 0.7% suppression rate)

Typological notes

This pack was selected to span typologically distinct language families and speech types:

• Korean is a language isolate with phrase-final focus marking and complex mora timing — a useful contrast to the stress-timed Indo-Aryan languages.
• Hindi is the largest corpus here and provides strong statistical power for Indo-Aryan prosody baselines.
• Hebrew is a VSO Semitic language with root-and-pattern morphology; the high metadata coverage makes it useful for demographic-stratified analyses.
• Manx is a Celtic revival language with a tiny native speaker community. The 98% PRISTINE rate reflects the controlled recording conditions of motivated community contributors.
• Tzeltal is a Mayan language with ergative-absolutive alignment and a distinctive tonal register system. It is rarely represented in acoustic datasets.
• Maguindanao (SPS2) is spontaneous speech from a Philippine Austronesian language. The T2-heavy distribution reflects the naturalistic recording conditions of the SPS2 corpus.
• Lasi (SPS2) is a Sindhi variety spoken in Balochistan. Shorter median clip duration (3.4s vs 5–6s for CV24 languages) reflects the spontaneous speech format.

---

Dataset Structure

Each Parquet file contains one row per utterance. Files are Snappy-compressed.

Column reference

Column	Type	Description
clip_id	string	Anonymized sequential ID (e.g. korean_cv24_004521)
lang_code	string	BCP-47 language code
lang_name	string	Language name
corpus_id	string	Source corpus (cv24 or sps2)
speech_type	string	scripted or spontaneous
tier	int	Quality tier: 1 (PRISTINE) or 2 (STUDIO). T3/T4 suppressed at export.
tier_label	string	PRISTINE or STUDIO
duration_ms	int	Clip duration bucketed to nearest 100ms
gender	string	male / female / other / unknown
gender_source	string	meta (self-reported) / inferred (pitch-based) / unknown
age	string	Age bracket from CV metadata (where available; blank for SPS2)
syllable_count_approx	int	Approximate syllable count (vowel-count proxy; transcript removed)
pitch_mean	float32	Mean F0 (Hz)
pitch_std	float32	F0 standard deviation (Hz)
pitch_range	float32	F0 range max–min (Hz)
pitch_velocity_max	float32	Maximum rate of F0 change (Hz/s)
intensity_mean	float32	Mean RMS intensity (dB)
intensity_max	float32	Peak intensity (dB)
intensity_range	float32	Dynamic range (dB)
intensity_velocity_max	float32	Maximum rate of intensity change
hnr_mean	float32	Harmonics-to-noise ratio (dB)
cpps	float32	Cepstral peak prominence smoothed (breathiness)
jitter_local	float32	Cycle-to-cycle pitch perturbation
shimmer_local	float32	Cycle-to-cycle amplitude perturbation
spectral_centroid_mean	float32	Mean spectral centroid (Hz)
spectral_tilt	float32	Spectral slope (voice effort indicator)
mfcc_delta_mean	float32	Mean MFCC delta (rate of spectral change)
zcr_mean	float32	Zero-crossing rate
teo_mean	float32	Teager energy operator mean
teo_std	float32	Teager energy operator std dev
f1_mean	float32	First formant mean (Hz)
f2_mean	float32	Second formant mean (Hz)
f3_mean	float32	Third formant mean (Hz)
formant_dispersion	float32	F3–F1 dispersion (Hz)
npvi	float32	Normalized pairwise variability index (rhythm)
articulation_rate	float32	Syllables per second (speech intervals only)
snr_median	float32	Median SNR (Brouhaha)
c50_median	float32	Median C50 clarity (Brouhaha)
speech_ratio	float32	Proportion of clip containing voiced speech
emotion_score	float32	Brouhaha emotional arousal proxy
dialect_tag	string	Accent/dialect slug from CV metadata (where available; blank for SPS2)
sample_type	string	core (cream-selected) or null

Quality tiers

Clips were graded using Brouhaha acoustic scoring. The T2 gate requires all three conditions simultaneously (corrected in v1.1):

Tier	Label	SNR	C50	Speech ratio	Description
T1	PRISTINE	≥ 35 dB	≥ 35 dB	≥ 0.30	Studio/near-studio quality
T2	STUDIO	≥ 25 dB	≥ 20 dB	≥ 0.30	Clean field recording, low reverb
T3	WILD	≥ 10 dB	any	≥ 0.10	Usable but noisy — not exported
T4	TRASH	< 10 dB	any	< 0.10	Rejected — not exported

Only T1 and T2 clips appear in this dataset. T3 and T4 are excluded at export.

Files

korean_cv24.parquet         6,805 rows   ~600 KB
hindi_cv24.parquet         18,435 rows   ~1.1 MB
hebrew_cv24.parquet         5,455 rows   ~475 KB
manx_cv24.parquet           6,579 rows   ~440 KB
tzeltal_cv24.parquet        5,585 rows   ~500 KB
mdn_cv24.parquet            5,536 rows   ~500 KB
lasi_cv24.parquet          10,435 rows   ~860 KB

---

Dataset Creation

Curation Rationale

This release addresses two gaps: the quality-gate bugs in the initial Korean/Hindi release, and the absence of any low-resource, non-Western language representation in the initial pack. Tzeltal, Manx, Lasi, and Maguindanao are rarely seen in structured acoustic datasets. The SPS2 spontaneous speech languages (Maguindanao, Lasi) provide a direct contrast to the CV24 scripted speech languages within this same release.

Source Data

Processing Pipeline

1. MP3 source audio converted to 16 kHz mono WAV (ffmpeg, −20 dBFS normalization)
2. Quality grading via Brouhaha (SNR, C50, VAD) — T1/T2 retained only
3. Tier retroactively corrected post BUG-029/030 fix via regrade_tiers.py
4. Acoustic feature extraction via Parselmouth/Praat at 16 kHz
5. Cream selection (demographically balanced 25-minute representative subset per language) recorded in sample_type field
6. Anonymization and precision degradation applied at export

Source Data Producers

CV24 recordings were made by volunteer contributors to the Mozilla Common Voice project under CC0. SPS2 recordings were collected by the Mozilla Spontaneous Speech project under CC0. Contributors self-reported demographic metadata where willing; many rows will have blank age/accent fields.

Anonymization — moonscape_k5_v1

• Original Mozilla filenames replaced with sequential anonymized clip IDs (mapping kept internal, never distributed)
• Transcripts removed entirely (approximate syllable count provided as proxy)
• All continuous acoustic variables truncated to 2 decimal places, stored as float32
• Duration bucketed to nearest 100ms
• K-anonymity at k=5 on {gender, age_bucket, duration_bucket} — rows in groups smaller than k=5 suppressed (395 rows suppressed across 7 languages, 0.7%)

---

Bias, Risks, and Limitations

Gender coverage varies significantly by language. Hebrew has 93% self-reported gender metadata but is heavily male-skewed (4,982 male vs 95 female meta-labeled records). Tzeltal has 68% metadata coverage but all meta-labeled records are female — this reflects the contributor demographics of the CV24 Tzeltal community at time of collection, not the language population. Manx, Maguindanao, and Lasi have 0% self-reported gender; all gender labels are pitch-inferred and should be treated accordingly. Always inspect gender_source before demographic analyses.

Scripted vs spontaneous speech are not directly comparable. CV24 (Korean, Hindi, Hebrew, Manx, Tzeltal) is read speech from volunteer recordings of prompted sentences. SPS2 (Maguindanao, Lasi) is spontaneous conversational speech. Articulation rate, pause rate, nPVI, and pitch dynamics will differ systematically between the two corpus types — this is a real typological signal but also a corpus-type confound. The speech_type and corpus_id columns allow you to stratify analyses accordingly.

Recording conditions are uncontrolled for CV24. Common Voice contributors record at home on personal devices. Acoustic conditions vary widely; the quality gate reduces but does not eliminate this variance. Manx is an exception — the 98% PRISTINE rate suggests a tightly controlled recording campaign by the community.

Lasi and Maguindanao lack TSV demographic metadata. The SPS2 corpus was released without validated speaker demographic files. Age, accent, and dialect fields will be blank for these languages. Gender is pitch-inferred only.

Formant data (F1/F2/F3) is present but may be unreliable for tonal languages. The Parselmouth formant tracker can produce artifacts in tonal contexts. Cross-validate against known formant benchmarks before using F1/F2/F3 for Korean, Tzeltal, or Maguindanao.

Prohibited use: Do not attempt to use this dataset for speaker identification, speaker re-linking to source audio, or any form of individual re-identification. This is prohibited regardless of technical feasibility and violates the terms of use.

Recommendations

Use gender_source == 'meta' to filter to self-reported labels for any demographic analysis. Use corpus_id to separate scripted from spontaneous comparisons. For rhythm typology work, npvi and articulation_rate are the most reliable features in this release. Treat intensity_mean and intensity_max with caution — Mozilla applies normalization during encoding which compresses the true dynamic range.

---

Citation

@dataset{alexandria_multilingual_prosody_v1_1_2026,
  title     = {Alexandria Voice Corpus --- Multilingual Macro-Prosody Telemetry v1.1},
  author    = {MoonScape Software},
  year      = {2026},
  license   = {CC0-1.0},
  note      = {Derived from Mozilla Common Voice CV24 and Spontaneous Speech SPS2 (CC0).
               Acoustic features extracted via Parselmouth/Praat. v1.1 corrects
               quality-gate bugs BUG-029, BUG-030, and high-SNR C50 bypass present
               in the earlier Korean/Hindi release.}
}

@misc{mozilla_common_voice,
  title     = {Common Voice: A Massively-Multilingual Speech Corpus},
  author    = {Ardila, Rosana and others},
  year      = {2020},
  url       = {https://commonvoice.mozilla.org}
}

---

Glossary

Term	Definition
F0 / pitch_mean	Fundamental frequency — perceived pitch, Hz
HNR	Harmonics-to-noise ratio — higher = cleaner, more periodic voice
CPPS	Cepstral peak prominence smoothed — lower = breathier voice
nPVI	Normalized pairwise variability index — durational variability between adjacent syllables; higher in stress-timed languages
C50	Clarity metric — higher = less reverb/echo
SNR	Signal-to-noise ratio — higher = cleaner recording
Brouhaha	Quality scoring model: github.com/marianne-m/brouhaha-vad
T1 / PRISTINE	SNR ≥ 35, C50 ≥ 35, speech_ratio ≥ 0.30
T2 / STUDIO	SNR ≥ 25, C50 ≥ 20, speech_ratio ≥ 0.30 (all three required simultaneously)
moonscape_k5_v1	Anonymization: k=5 suppression + sequential IDs + 2dp truncation + 100ms duration bucketing
CV24	Mozilla Common Voice 24.0 — scripted read speech
SPS2	Mozilla Spontaneous Speech 2.0 — unscripted conversational speech

---

Contact

moonscapesoftware@gmail.com