Mikimi/phoible-voxangeles-formants-aggregated

PHOIBLE × VoxAngeles — Phoneme Inventories with Formants (Aggregated)

A cross-linguistic tabular dataset that pairs the phoneme inventories of the world's languages (from PHOIBLE) with acoustic formant measurements (F1–F5) derived from real audio recordings (from the VoxAngeles corpus).

Each row is one phoneme of one language inventory, annotated with the full set of PHOIBLE distinctive features and, where matching audio exists, with averaged formant frequencies.

Dataset summary

PHOIBLE is a large database of phonological inventories: it records, for each documented language, which phonemes that language uses, encoded as IPA symbols together with a matrix of distinctive features. PHOIBLE itself contains no audio — it is a purely symbolic resource.

VoxAngeles (a phonetically audited and segment-aligned subset of the UCLA Phonetics Lab Archive, covering roughly 95 languages under a CC-BY-NC 4.0 licence) does contain audio with phone-level segmentation. This dataset links the two: for every recorded phone in VoxAngeles, the first five formants are measured with Praat, and the measurements are matched back to PHOIBLE phonemes by the pair (language, IPA symbol).

This aggregated dataset reports, per phoneme inventory entry, the formant values averaged over all available realizations of that phoneme. A companion dataset (the realizations dataset) stores every individual recorded realization as its own row, with per-token formants and full provenance back to the source audio file.

Because audio coverage is limited to the languages present in VoxAngeles, the majority of PHOIBLE phonemes have no matching audio; for those rows the formant columns are NA. This is expected and is recorded explicitly via the has_audio flag.

Provenance and how rows are built

1. PHOIBLE CLDF tables (languages.csv, parameters.csv, values.csv, inventories.csv) are joined into one flat table: one row per phoneme per inventory, carrying language metadata and all distinctive features.
2. VoxAngeles ships pre-computed formant measurements in data/phonetic_measurements/voxangeles_formants_quartiles.tsv. These F1-F3 values were measured by the corpus authors with Praat (Burg method) at quartiles of vowel duration. This dataset reads those values directly; it does not recompute formants.
3. The measurements are normalized: one row per vowel realization, with F1-F3 averaged across the available duration quartiles. F4 and F5 are not provided by VoxAngeles and are therefore always NA.
4. Realizations are grouped by (language code, IPA symbol); formants are averaged within each group for the aggregated dataset.
5. A left join keeps all PHOIBLE phonemes and attaches the averaged formants where a match exists. Non-matching phonemes keep NA formants.
6. Three derived articulatory-effort proxy columns are computed. formant_distance_hz is the Euclidean distance of (F1, F2, F3) from a neutral schwa reference (500/1500/2500 Hz); formant_distance_bark is the same distance in the perceptual Bark scale; effort_proxy multiplies the per-token distance by the segment duration and is then averaged over the phoneme's realizations. These are coarse proxies, not measured acoustic energy. For phonemes without audio they are NA.
7. Speaker-normalized formants are added for cross-linguistic comparison. F1_lobanov–F3_lobanov are within-speaker z-scores (F − mean)/sd; F1_nearey–F3_nearey are Nearey log-mean values ln(F) − G, where G is the speaker's grand mean of ln(F1..F3). Statistics are pooled within iso639_3 (VoxAngeles provides ~one speaker per language). The Lobanov columns need at least two tokens per speaker, so lone tokens are NA there; Nearey stays defined but is unreliable for sparsely sampled speakers. In this aggregated dataset the normalized values are averaged per phoneme.

Data fields

Column	Type	Description
language_name	string	Language name (Glottolog).
iso639_3	string	ISO 639-3 language code.
glottocode	string	Glottolog languoid code.
macroarea	string	Macroarea (e.g. Eurasia, Africa, ...).
latitude	string	Language latitude.
longitude	string	Language longitude.
inventory_id	string	PHOIBLE inventory identifier.
phoneme_ipa	string	The phoneme, as an IPA symbol.
segment_class	string	consonant, vowel, or tone.
has_audio	bool	true if matching VoxAngeles audio was found.
n_segments	int	Number of recorded realizations the formants are averaged over (0 if none).
F1–F5	float	Mean formant frequencies in Hz, or NA.
formant_distance_hz	float	Euclidean distance of (F1,F2,F3) from the neutral schwa (500/1500/2500 Hz), in Hz. A proxy for articulatory deviation from the neutral vocal-tract setting — not true acoustic energy.
formant_distance_bark	float	The same distance computed in the perceptual Bark scale.
effort_proxy	float	formant_distance_hz x duration, averaged over the phoneme's realizations; a coarse proxy for cumulative articulatory effort.
F1_lobanov–F3_lobanov	float	Speaker-normalized formants — Lobanov z-score (F − mean)/sd pooled within speaker. Dimensionless, comparable across speakers/languages.
F1_nearey–F3_nearey	float	Speaker-normalized formants — Nearey log-mean ln(F) − G (G = speaker's grand mean of ln(F1..F3)). In log units.
feat_*	string	PHOIBLE distinctive features (+, -, 0); ~37 columns such as feat_tone, feat_syllabic, feat_consonantal, feat_nasal, etc.
realization_ids	list[string]	IDs of the individual realizations in the companion realizations dataset.

The NA rule for formants

• If a phoneme has no matching recorded audio, all of F1–F5 are NA (has_audio = false, n_segments = 0).
• If audio exists but Praat could measure only k < 5 formants, then k numeric values are kept and the remaining 5 − k are NA.
• In CSV exports NA is written literally as NA; in the parquet / Hugging Face version it is a true null value.

Usage

from datasets import load_dataset

ds = load_dataset("Mikimi/phoible-voxangeles-formants-aggregated", split="train")

# only phonemes that have measured formants
with_audio = ds.filter(lambda r: r["has_audio"])

# all vowels
vowels = ds.filter(lambda r: r["segment_class"] == "vowel")

Limitations

• Audio coverage is partial. Only languages present in VoxAngeles (~95 languages) can receive formant values; all other PHOIBLE phonemes have NA formants.
• Formants describe realizations, not abstract phonemes. A formant value is a property of a specific recorded token produced by a specific speaker. Values here are averages over the available tokens and should be read as a summary, not as a canonical property of the phoneme.
• Matching is symbol-based. Realizations are linked to PHOIBLE by exact (language, IPA symbol) pairs; transcription differences between sources can cause missed matches.
• Measurement settings matter. Formants depend on the Praat ceiling frequency; a single global setting cannot be optimal for every speaker.
• The effort columns are proxies, not energy. formant_distance_* and effort_proxy are derived from formant geometry (deviation from a neutral schwa), not from the speech signal. The schwa reference is calibrated to an adult male vocal tract, so cross-speaker comparison is more reliable in the Bark column than in raw Hz.
• Normalization assumes one speaker per language. Lobanov/Nearey statistics are pooled within iso639_3, matching VoxAngeles' typical one-speaker-per-language structure; if a language actually contains several speakers, the normalized values mix them. Normalization also needs several vowel tokens per speaker, so sparsely sampled languages may have NA.

Sources and licences

• PHOIBLE — Moran, Steven & McCloy, Daniel (eds.) 2019. PHOIBLE 2.0. Jena: Max Planck Institute for the Science of Human History. https://phoible.org — licence: CC-BY-SA.
• VoxAngeles — a phonetically audited, segment-aligned version of the UCLA Phonetics Lab Archive — licence: CC-BY-NC 4.0.

This derived dataset is released under CC-BY-SA-4.0. Note that the audio it is based on (VoxAngeles) is non-commercial (CC-BY-NC 4.0); any redistribution of the underlying audio must respect that non-commercial term.

Citation

If you use this dataset, please cite both source projects:

@book{phoible2019,
  address   = {Jena},
  editor    = {Moran, Steven and McCloy, Daniel},
  publisher = {Max Planck Institute for the Science of Human History},
  title     = {PHOIBLE 2.0},
  url       = {https://phoible.org},
  year      = {2019}
}

Please also cite the VoxAngeles corpus and the UCLA Phonetics Lab Archive according to their current citation guidelines.