To access the Moonscape Human Speech Atlas, please provide your institutional details and agree to the biometric privacy terms. NOTE: Requests utilizing non-institutional emails (@gmail, @hotmail, @yahoo, @outlook, etc.) will be automatically rejected for the Restricted Tier.

Moonscape Human Speech Atlas (HSA)

Moonscape Software — Human Speech Atlas

A curated corpus of pre-extracted acoustic feature matrices for multilingual prosody research. The HSA covers 90+ languages across 12 linguistic family trees, derived from Mozilla Common Voice (CV24.0) and Spontaneous Speech (SPS2.0) — both CC0.

The HSA is a companion project to the Synthetic Speech Atlas (SSA), which provides parallel feature-space representations of synthetic and bonafide speech for deepfake detection. Together they form a complete acoustic telemetry platform.

Audio files are not included. Each file is a Snappy-compressed Parquet dataset of 38 hand-crafted classical signal-processing features extracted via Parselmouth/Praat and Brouhaha, with full biometric protection applied at export.

Export version: HSA_v1_2026 Anonymization standard: moonscape_k5_fp16_v1 Watermark: HMAC-SHA256 seeded FP16-resolution noise per column

Repository Structure

This is a single gated repository. All 12 linguistic family trees are available as named subsets (configurations) within it. Each subset maps to a subdirectory of Parquet files under data/.

Human_Speech_Atlas/
├── README.md
├── LICENSE.md
└── data/
    ├── Indo-European/        ← ~148K cream clips, 50 tables
    ├── Niger-Congo/          ← ~144K cream clips, 34 tables
    ├── Austronesian/         ← ~46K cream clips, 17 tables
    ├── Mesoamerican/         ← ~28K cream clips, 13 tables
    ├── Americas-Other/       ← ~16K cream clips, 4 tables
    ├── Afro-Asiatic/         ← ~14K cream clips, 5 tables
    ├── Nilo-Saharan/         ← ~12K cream clips, 4 tables
    ├── Trans-New-Guinea/     ← ~11K cream clips, 3 tables
    ├── Eurasian-Minor/       ← ~8K cream clips, 8 tables
    ├── Turkic/               ← ~8K cream clips, 6 tables
    ├── Asian-Minor/          ← ~8K cream clips, 10 tables
    └── Isolates/             ← ~6K cream clips, 2 tables

Loading a Subset

from datasets import load_dataset

# Load a single family tree
df = load_dataset("moonscape-software/Human_Speech_Atlas", "Indo-European")

# Load multiple trees
indo = load_dataset("moonscape-software/Human_Speech_Atlas", "Indo-European")
niger = load_dataset("moonscape-software/Human_Speech_Atlas", "Niger-Congo")

# Load all trees (large — ~531K rows total)
all_trees = {
    tree: load_dataset("moonscape-software/Human_Speech_Atlas", tree)
    for tree in [
        "Indo-European", "Niger-Congo", "Austronesian", "Mesoamerican",
        "Americas-Other", "Afro-Asiatic", "Nilo-Saharan", "Trans-New-Guinea",
        "Eurasian-Minor", "Turkic", "Asian-Minor", "Isolates"
    ]
}

Family Tree Index

Subset (config_name)	Languages	Cream Clips	Notable Languages
`Indo-European`	50 tables	~148K	Hindi, Cornish, Gujari, Dhatki, Khowar, Kalasha, Manx, Lasi, Gawri
`Niger-Congo`	34 tables	~144K	Hausa, Nawdm, Massa, Chokwe, Fang, Igbo, Cameroon Grassfields
`Austronesian`	17 tables	~46K	Seediq, Batak Toba, Gotontalo, Cuyonon, Melanau
`Mesoamerican`	13 tables	~28K	Tzeltal, 3x Mazatec, Mixtec, Totonac, Huichol
`Americas-Other`	4 tables	~16K	Central Alaskan Yupik, Seri, Quechua
`Afro-Asiatic`	5 tables	~14K	Hausa, Hebrew, Tashlhiyt Berber
`Nilo-Saharan`	4 tables	~12K	Dinka Ruweng, Kenyan Luo
`Trans-New-Guinea`	3 tables	~11K	Mauwake, Ukuriguma (Papuan deep-time)
`Eurasian-Minor`	8 tables	~8K	Adyghe, Kabardian, Lak, Moksha, Votic
`Turkic`	6 tables	~8K	Kazakh, Tuvan, Bashkir
`Asian-Minor`	10 tables	~8K	Min Dong, Puxian Min, Keazi, Bodo
`Isolates`	2 tables	~6K	Korean, Georgian

Clip counts are post-k-anonymisation (k=5). Rows where quasi-identifier groups contain fewer than 5 members are suppressed.

Why This Dataset Exists

Cross-linguistic prosody research is underserved by existing open corpora. The HSA enables researchers to:

Study prosodic typology across genetically unrelated language families in a single unified schema and a single gated repository
Train cross-lingual speech models without audio infrastructure
Benchmark acoustic features across tonal, stress-timed, and syllable-timed languages
Use the Isolates subset (Korean, Georgian) as a control group for studying convergent vs inherited prosodic features
Access rare and endangered language data — Votic (~5 remaining speakers), Tashlhiyt Berber (consonant-only syllabics), Trans-New Guinea Papuan languages

Privacy & Legal Framework

Why This Corpus Is Gated

These files contain acoustic features derived from recordings of real human speakers. Although audio is not included, acoustic feature vectors carry residual biometric information. The following controls are applied:

Access restricted to verifiable institutional actors
Non-institutional emails rejected at intake
Commercial use requires a separate EULA with Moonscape Software
All users agree to non-re-identification and watermark integrity obligations

Biometric Protection Measures — `moonscape_k5_fp16_v1`

UID link cut — file_id, client_id, source_file, and transcript stripped entirely. No path back to source speaker or sentence identifiers.
K-anonymity (k≥5) — Rows suppressed where the quasi-identifier group {gender, age_bucket, duration_bucket} contains fewer than 5 members.
Duration bucketing — duration_ms rounded to nearest 100ms.
Precision reduction — All continuous acoustic variables rounded to 2dp.
FP16 reinflation watermark — Each 2dp-rounded value is reinflated with deterministic FP16-resolution seeded noise:
```
seed  = HMAC-SHA256(HSA_EXPORT_SECRET, col_name + "|HSA_v1_2026")
noise ~ Uniform(-0.004, +0.004)  [seeded deterministically per column]
value = float16(round(raw, 2) + noise)  [stored as float32]
```
This destroys the backward-engineering path to raw biometric values while preserving statistical validity. Every row carries a unique verifiable provenance signature.

Source Data License

All source audio is Mozilla Common Voice CV24.0 (CC0-1.0) and Mozilla Spontaneous Speech SPS2.0 (CC0-1.0). No audio is redistributed. Feature extraction pipeline and methodology are copyright Moonscape Software.

Dataset Schema

All HSA parquet files share an identical 46-column canonical schema regardless of language, family tree, or source corpus. Column order is fixed.

Identity & Provenance

Column	Type	Description
`clip_id`	string	Anonymous sequential ID. Format: `{lang_corpus_NNNNNN}`
`lang_code`	string	ISO 639-3 language code (e.g. `ko`, `ha`, `btv`)
`lang_name`	string	Human-readable language name
`corpus`	string	Source corpus: `cv24` or `sps2`
`speech_type`	string	`scripted` (CV24) or `spontaneous` (SPS2)
`source_dataset`	string	Full source name (e.g. `Mozilla Common Voice CV24.0`)
`sentence_domain`	string	Text domain: `wikipedia` \| `news` \| etc.

Demographics

Column	Type	Description
`gender`	string	`male` \| `female` \| `other` \| `unknown`
`age`	string	Age bracket (e.g. `20-29`) or `unknown`
`accent`	string	Self-reported accent/dialect label
`dialect_tag`	string	Normalised dialect code
`sample_type`	string	`cream_t1` \| `cream_t2` \| `fill_t3` \| `fill_t4`

Temporal

Column	Type	Description
`duration_ms`	Int64	Clip duration bucketed to nearest 100ms
`duration_s`	float32	`duration_ms / 1000`

Quality Gate (Brouhaha)

Column	Type	Description
`tier`	int	1=PRISTINE 2=STUDIO 3=AMBIENT 4=TRASH
`tier_label`	string	`PRISTINE` \| `STUDIO` \| `AMBIENT` \| `TRASH`
`snr_median`	float32	Median signal-to-noise ratio (dB)
`snr_mean`	float32	Mean SNR (dB)
`c50_median`	float32	Median room clarity C50 (dB)
`speech_ratio`	float32	Active speech fraction (0-1)

Acoustic Features (float32, FP16 watermarked)

Column	Units	Description
`pitch_mean`	Hz	Mean F0 (VAD-bounded, voiced frames only)
`pitch_std`	Hz	F0 standard deviation
`pitch_range`	Hz	95th-5th percentile F0
`pitch_velocity_max`	Hz/frame	Max F0 rate-of-change
`jitter_local`	%	Cycle-to-cycle period variation (MP3 fidelity caveat)
`shimmer_local`	%	Cycle-to-cycle amplitude variation (MP3 fidelity caveat)
`hnr_mean`	dB	Harmonics-to-noise ratio
`cpps`	—	Cepstral peak prominence smoothed
`intensity_mean`	dB	Mean intensity (normalised — see limitations)
`intensity_max`	dB	Peak intensity (normalised — see limitations)
`intensity_range`	dB	Dynamic range
`intensity_velocity_max`	dB/frame	Max intensity rate-of-change
`spectral_centroid_mean`	Hz	Mean spectral centroid
`spectral_tilt`	dB/kHz	Log-power spectrum slope
`mfcc_delta_mean`	—	Mean first-order MFCC delta
`zcr_mean`	—	Zero crossing rate
`teo_mean`	—	Mean Teager Energy Operator
`teo_std`	—	TEO standard deviation
`f1_mean`	Hz	Mean first formant
`f2_mean`	Hz	Mean second formant
`f3_mean`	Hz	Mean third formant
`formant_dispersion`	Hz	`(F3-F1)/2` — vocal tract length proxy
`npvi`	—	Normalised Pairwise Variability Index (0.0 pending MFA)
`articulation_rate`	syl/s	Syllable rate (0.0 pending MFA)
`emotion_score`	0-1	Composite vocal intensity score
`syllable_count_approx`	int	Vowel-count syllable proxy

Known Limitations

intensity_mean / intensity_max — Mozilla normalises source audio to -20 dBFS. These columns are dead vectors. Cross-speaker intensity comparison is invalid.
jitter_local / shimmer_local — MP3 codec degrades sub-ms glottal measurements. HNR and CPPS are more robust alternatives for this corpus.
npvi / articulation_rate — Return 0.0 pending Phase 2 MFA phoneme alignment.
Tonal languages — In Niger-Congo, Tai-Kadai, and Sino-Tibetan languages, pitch_mean/std/range measure lexical tone, not prosodic stress.

Quality Tiers

Tier	Label	SNR	C50	Speech ratio
1	PRISTINE	>= 35 dB	>= 35 dB	>= 0.30
2	STUDIO	>= 25 dB	>= 20 dB	>= 0.30
3	AMBIENT	>= 10 dB	any	>= 0.10
4	TRASH	< 10 dB	any	< 0.10

Usage Examples

from datasets import load_dataset

# Load Indo-European family (default subset)
ds = load_dataset("moonscape-software/Human_Speech_Atlas", "Indo-European")
df = ds["train"].to_pandas()

# Cream clips only (T1+T2)
cream = df[df["sample_type"].str.startswith("cream", na=False)]

# Scripted (CV24) vs spontaneous (SPS2)
scripted    = df[df["corpus"] == "cv24"]
spontaneous = df[df["corpus"] == "sps2"]

# Cross-family pitch comparison (exclude tonal languages)
from datasets import load_dataset, concatenate_datasets
trees = ["Indo-European", "Austronesian", "Mesoamerican",
         "Eurasian-Minor", "Turkic", "Isolates"]
frames = [load_dataset("moonscape-software/Human_Speech_Atlas", t)["train"]
          for t in trees]
combined = concatenate_datasets(frames).to_pandas()
non_tonal = combined[~combined["lang_code"].isin(["th","cdo","cpx","brx"])]
print(non_tonal.groupby("lang_name")["pitch_mean"].mean().sort_values())

Extraction Pipeline

Pass 1 — Brouhaha (Lavechin et al., Interspeech 2022) — SNR, C50, VAD, tier.

Pass 2 — Classical Acoustic Features — Parselmouth/Praat + librosa, 38 features.

Pass 3 — MFA Phoneme Alignment (pending) — Montreal Forced Aligner on cream WAV sets. Will populate npvi and articulation_rate.

Licensing

Tier 1 — Moonscape Academic License (Non-Commercial)

Free for academic and non-commercial research on submission of institutional email. Full terms in LICENSE.md. Key obligations:

Cite Mozilla Common Voice and this dataset in any publications
Never attempt to re-identify human speakers from acoustic features
Never alter or remove forensic watermarks
Not redistribute as a standalone commercial product

Tier 2 / Tier 3 — Commercial License

Requires execution of a separate EULA with Moonscape Software prior to access.

Source	License
Mozilla Common Voice CV24.0	CC0-1.0
Mozilla Spontaneous Speech SPS2.0	CC0-1.0
Feature extraction pipeline	Copyright Moonscape Software

Citation

@dataset{kleingertner2026hsa,
  author    = {Kleingertner, Chris},
  title     = {Moonscape Human Speech Atlas (HSA)},
  year      = {2026},
  publisher = {Moonscape Software},
  note      = {Multilingual acoustic prosody feature corpus across 90+ languages
               in 12 linguistic family trees. Companion to the Synthetic Speech Atlas.},
  url       = {https://huggingface.co/datasets/moonscape-software/Human_Speech_Atlas}
}

@misc{mozilla2024commonvoice,
  title  = {Mozilla Common Voice},
  author = {Mozilla Foundation},
  year   = {2024},
  url    = {https://commonvoice.mozilla.org},
  note   = {CV24.0 release, CC0-1.0}
}

@inproceedings{lavechin2022brouhaha,
  title     = {Brouhaha: Multi-task Training for Noise Speech Detection and Assessment},
  author    = {Lavechin, Marvin and others},
  booktitle = {Interspeech},
  year      = {2022}
}

Access & Terms

This repository is gated. All users agree to:

Cite Mozilla Common Voice and this dataset in any publications
Never attempt to re-identify human speakers from acoustic features
Never alter, truncate, or remove the forensic watermarks (seeded FP16 noise)
Comply with upstream Mozilla Common Voice CC0 terms
Not redistribute this feature matrix as a standalone product without a separate commercial licence from Moonscape Software

Tier 1 (Academic / Non-Commercial): Approved on institutional email. Full terms in LICENSE.md.

Tier 2/3 (Commercial): Requires EULA execution with Moonscape Software.

Moonscape Software — Human Speech Atlas Export version: HSA_v1_2026 | Watermark: HMAC-SHA256 seeded FP16 noise

Downloads last month: -

You need to agree to share your contact information to access this dataset