Badjia-ALCAM-MultimodalDataset

Language

Badja (also spelled Badjia; dialect name Bakjo) is classified in standard typological work as a dialect of Ewondo (also known as Beti, Yaunde, Cameroonian; ISO 639-3: ewo; Guthrie code A.72), a Bantu language (Niger-Congo > Atlantic-Congo > Volta-Congo > Benue-Congo > Bantoid > Southern Bantoid > Bantu, Zone A > Beti) spoken by an estimated 577,700 people (1982 figure; more recent secondary estimates put the number of speakers above 500,000). Ewondo speakers are concentrated in the departments of Mfoundi, Mefou-et-Afamba, Mefou-et-Akono, Nyong-et-So'o and Nyong-et-Mfoumou in the Centre Region, and in part of the Océan department in the South Region. Ewondo is mutually intelligible, or closely related, to other Beti-Fang varieties such as Bulu, Eton and Fang, and is widely used as a trade language in and around Yaoundé. This dataset documents the Badja (Badjia) dialect specifically.

Variants

Published descriptions of Ewondo list a substantial number of named dialects alongside Badjia (Bakjo) — including Bafeuk, Bemvele (Mvele, Yezum, Yesoum), Bane, Beti, Enoah, Evouzom, Mbida-Bani, Mvete, Mvog-Niengue, Omvang, Yabekolo (Yebekolo), Yabeka and Yabekanga — reflecting the fine-grained, clan- and lineage-based dialectal fragmentation characteristic of Ewondo/Beti. This release documents a single, localized dialect (Badja/Badjia) as elicited from a native speaker; no further internal sub-variety split is documented within the dataset, and the precise village- or locality-level provenance of the variety documented here has not been recorded and should be clarified with the dataset's creator before downstream use.

Writing System

The writing system used for the transcription of Badja in this dataset is the International Phonetic Alphabet (IPA), as reflected in the Word, LangEx and LangPars columns of the datasheet and in the sentence field of the audio-sentence mapping files. The phonological inventory below draws on the 343 datasheet entries and the 326 audio-aligned sentences (the mapping.tsv files across the dataset's four recording batches). A harmonized practical orthography for Ewondo, developed for national-language teaching and Bible translation, is well established, but is not the transcription system used in this dataset.

1. Vowels

The data attests eight oral vowel qualities: a, e, ɛ, ə, i, o, ɔ, u, matching the vowel inventory reported for Ewondo in the wider literature. Vowel length is occasionally marked with a colon diacritic (5 tokens); no nasalized vowels are systematically marked in the transcribed data. A combining tilde-below diacritic (◌̰), which in IPA conventionally marks creaky voice, occurs on 73 tokens across the Word, LangEx and LangPars columns — the same convention observed in other datasets elicited with the same ALCAM questionnaire; its precise phonetic conditioning has not been established here and should be clarified with the dataset's creator before downstream use.

2. Consonants

The core simple-consonant inventory attested in the datasheet and mapping files is: b, d, g, j, k, l, m, n, p, r, s, t, v, w, y, z. Notably, f and h, both attested in published Ewondo consonant inventories, do not occur in the transcribed data.

Two affricates are attested: dʒ (voiced postalveolar, 4 tokens, occurring only in prenasalized position, i.e. always as ndʒ) and dz (voiced alveolar, 10 tokens, occasionally prenasalized as ndz, 3 tokens).

Prenasalized consonants are well attested: mb (75 tokens), nd (31), ŋg (88) and nz (76), consistent with the prenasalized consonant series widely reported for Ewondo and other Beti-Fang languages.

A labial-velar stop with a tie bar, k͡p, is attested (18 tokens); a corresponding tied voiced labial-velar g͡b does not occur — the plain sequence "gb" is found only twice and more likely reflects an adjacent, non-clustered g and b than a phonemic labial-velar.

Labialized consonants are only sparsely attested (sw, 3 tokens; tw, 2 tokens) and do not form a fully developed phonemic series comparable to the labialized series documented for some other languages in the same ALCAM series.

Word-final voiceless stops (t, p, k) frequently carry a superscript ʳ marking an audible release (117 tokens, e.g. títʳ "animals", kápʳ). A small number of forms instead mark aspiration (ʰ, 2 tokens, e.g. kʰu, pʰu) or an explicitly unreleased stop (◌̚, 2 tokens, e.g. bòk̚), suggesting variability in the realization of word-final stops that has not been further analysed here.

Two further symbols recur without a clearly documented function and should be clarified with the dataset's creator before downstream use: a diaeresis on w or n (6 tokens, e.g. ásẅɔ, ètün), and a word-final q restricted to three lexical entries (6 tokens total, e.g. mə́ndʒɔ̀q, bə̀ndʒɔ̀qʳ). An apostrophe (') also recurs on 5 tokens; as with the symbols above, its precise grammatical or phonetic function has not been formally documented.

3. Tone system

Three level tones are marked in the datasheet and mapping files:

• High tone (H): á

• Mid tone (M): ā

• Low tone (L): à

Unmarked vowels represent tonally neutral or contextually determined syllables. A rising tone, marked with a caron (ǎ), occurs on only 2 tokens; no falling (circumflex-marked) contour tone is attested in this data, even though the harmonized orthographic conventions used for Ewondo in national-language teaching provide for both rising and falling tone marks. This simplification, and the very limited attestation of contour tone generally, should be taken into account when using the dataset for tonal analysis.

Source

The dataset was collected through a questionnaire designed to gather basic information about the Badja lexicon and grammar. This was done as part of the Atlas Linguistique du Cameroun (ALCAM) project.

Domain

The dataset represents a linguistic questionnaire designed to elicit the basic lexicon and grammatical information of Badja (Badjia).

Size

The datasheet comprises 343 elicited lexical/example-sentence entries. The audio subfolder contains 326 audio files distributed across four recording batches (91, 99, 40 and 96 clips respectively; 4m 56s, 3m 50s, 2m 04s and 5m 26s of speech, measured directly from the audio files), for a total of 16 minutes 18 seconds of speech across 326 files. No cross-batch redundancy (repeated sentences recorded across more than one batch) was found: all 326 recorded sentences and all 326 audio filenames are distinct, so the dataset contains 326 unique sentence-audio pairs.

Structure

The dataset comprises: 1) a datasheet (AlCAM_dataset_badjia.tsv) with lexical entries, French glosses, and example sentences with word-for-word parsing; 2) voice clips read by a native speaker of Badja, organized into four recording batches under the audio subfolder; 3) sentence-to-audio mapping files (mapping.tsv, one per recording batch).

Description of columns

Datasheet (AlCAM_dataset_badjia.tsv):

• OrigID: original number of lexical entry on paper questionnaire

• EditID: modification of OrigID

• FrenchRef: reference entry (originally provided in French)

• FrenchComm: original comments about reference entry (FrenchRef)

• French: lexical entry in French (overlaps with FrenchRef)

• Note: note of researcher on the lexical entry

• POS: part of speech

• Class: noun class (where applicable); not populated in this release — Ewondo is a Bantu language with a productive noun-class system in the broader literature, but this was not annotated as part of the present questionnaire

• Morf: morphological attribute (e.g. plural, singular)

• Var: dialect variant tag; not populated in this release, since the dataset represents a single, localized dialect (see Variants above)

• Word: lexical entry in Badja (IPA)

• CrossRef: cross-referencing of lexical entry number

• FrenchEx: example sentence in French

• LangEx: example sentence in Badja (IPA)

• LangExEdit: manual editing of LangEx

• FrenchExEdit: edited French equivalent of FrenchEx

• LangPars: word-for-word parsing in Badja

• LangParsEdit: editing of LangPars

• FrenchPars: French equivalent of LangParsEdit

• FrenchParsEdit: editing of FrenchPars

_ marks fields left empty in the source questionnaire. Of the 343 entries, 342 have a French gloss, 334 have a Badja Word, 263 have a French example sentence, 259 have a Badja example sentence, and 259 have word-for-word parsing. The audio subfolder separately provides recordings for 326 unique sentence-level items overall (see Size, above); a spot-check (allowing for Unicode normalization of combining diacritics) confirms 7 of these correspond exactly to LangEx entries in the datasheet (see Sample, below), but an exhaustive row-by-row correspondence between every audio clip and a specific datasheet entry has not been established.

Sample

The excerpt below shows 5 of the 343 datasheet entries (OrigID x77, x183, x203, x204, x205) for which the Badja example sentence (LangEx) has an exact, verified match (after Unicode normalization) in one of the audio-sentence mapping files, illustrating the parallel Badja/French structure together with the aligned audio.