Sample Ghomala-TTS-Dataset

Language

Ghomala (ISO 639-3: bbj) is a Grassfields Bantu language belonging to the Niger-Congo phylum, classified within the Mbam-Nkam branch of the Bantoid family. It is spoken in the West Region of Cameroon, predominantly in the département de la Mifi, and constitutes one of the largest Bamiléké languages in terms of speaker population. Despite its sociolinguistic significance within the Bamiléké cultural area of Cameroon, Ghomala remains substantially underrepresented in language technology resources.

Variants

According to the Administrative Atlas of Cameroon's Languages (Breton & Bikia Fohtung 1991), Ghomala comprises four dialect sub-areas:

• ghɔmala-nord: parlers fʉ'sap (Bafoussam) and laŋ (Baleng)

• ngemba (ghɔmala-ouest): parlers mugum (Bamugum), meka (Bameka) and mɔnjɔ (Bamenju)

• ghɔmala-central: parlers jo (Bandjoun), we (Bahuan), hɔm (Baham) and yogam (Bayangam)

• ghɔmala-sud: parlers tɛ' (Batiɛ), pa (Bapa) and denkwop (Badenkop)

The present dataset represents the ghɔmala-central sub-area, specifically the jo parler of Bandjoun.

Writing System

The writing system used for the transcription of Ghomala in this dataset is the General Alphabet of Cameroon's Languages (AGLC). The AGLC provides a phonologically motivated orthographic standard for Cameroonian national languages and serves as the reference framework for Ghomala literacy materials.

1. Vowels

The vowel system attested in the dataset includes the following oral vowels:

a, e, ə, ɛ, i, o, ɔ, u, ʉ

Where:

• ə (schwa, U+0259): mid central unrounded vowel; one of the highest-frequency vowels in the dataset

• ɛ (epsilon, U+025B): open-mid front unrounded vowel

• ɔ (open-o, U+0254): open-mid back rounded vowel

• ʉ (barred u, U+0289): high central rounded vowel

The digraph aə functions as a distinct complex vowel grapheme in Ghomala, frequently appearing in roots and affixes (e.g., gaə̂, gaə̂kə́, da'gaə́, maə́, yaə̌, pfaə̌). Tone diacritics are applied to either or both elements of this sequence as required.

Long vowels may be represented by vowel repetition in certain morphological contexts (e.g., áa as a sentence-final particle; pa'aa, ta'a in extended forms).

2. Consonants

The consonant inventory reflected in the dataset includes simple, digraph, labialized, palatalised, and nasal-onset consonants:

b, bv, c, cw, cy, d, dj, dz, f, g, gh, gw, gwy, h, j, k, kh, kp, l, m, mb, n, nd, ng, nj, nk, nt, nw, ny, ŋ, p, pf, s, sh, t, ts, v, w, y, z, zh

Notable consonantal features:

• gh: voiced velar fricative [ɣ] (e.g., ghɔ́, ghə, ghɔm)

• sh: voiceless postalveolar fricative [ʃ] (e.g., shimnyə, shyə, shə́)

• zh: voiced postalveolar fricative [ʒ] (e.g., zhʉ́zhʉ̂m, zhí'tə, mzhəŋ)

• kp: voiceless labio-velar stop (e.g., wɔ́kpə, ŋkwítə́)

• pf: voiceless labiodental affricate (e.g., pfaə̌, pfʉ́tə́, pfə̂)

• bv: voiced labiodental fricative (e.g., bvʉ̂m, bvʉ)

• ts: voiceless alveolar affricate (e.g., tsʉ', tsə)

• dz: voiced alveolar affricate (e.g., dzʉ̂, fɔkdzʉ, dzə̀)

• c: palatal/postalveolar affricate [tʃ] or [c] (e.g., cyə, cwə, cʉ')

• gw, gwy: labialized voiced velar stop (e.g., gwyə̌, gwəp, gwp)

• kh: aspirated or velar fricative (e.g., khʉ̂, kḥʉ)

• cw, cy: labialized and palatalised affricates (e.g., cwǝlɔ̌, cyə̌pa')

Nasal-onset clusters constitute a distinctive structural feature of Ghomala. They are formed by combining the nasals m or n with a following consonant, yielding sequences such as: mh (mhɔgnə), mn (mnɔm, mnə́), mk (mkwɛ́nyə̀, mkámtə́), mgh (mghətsə́, mghɛ̌və́), mf (mfʉ̌), mc (mcʉ̀m, mco'), ms (msəku, msətùk), mt (mtap, mtɔ, mtâp), mzh (mzhəŋ, mzhinyə), mj (mjwǐ, mjyə̂), mw (mnwə, mwə), nt (ntǎknyə, ntɔ̌knyə́), nk (nkáp, nkwítə́), nw (nwə́), nj (njɔm). These clusters correspond to nasal onset consonant sequences in the Ghomala phonological system — including prenasalised stops and nasal-initial consonant clusters arising from the combination of nominal class prefixes with root-initial consonants — and appear throughout the dataset.

Special symbols:

• ŋ (eng, U+014B): velar nasal consonant (e.g., ŋkáp, ŋwak, ŋwə̌)

• ' (apostrophe, U+0027): glottal stop / glottal closure marker (e.g., lá', cʉ', bɔ'ɔ, da'gaə́, bâ'ba', na'kə́tâm, zhi'tə)

3. Syllabic nasals

Ghomala attests syllabic nasal consonants that function as independent phonological units and tone-bearing positions. The nasals m, n, and ŋ appear in syllabic function and may carry tone diacritics. The eng ŋ also functions as a nasal onset in clusters such as ŋk and ŋw.

4. Tone system

Ghomala is a tonal language with multiple contrastive pitch levels and contour tones. The dataset employs systematic tone marking on vowels and syllabic nasals in accordance with the AGLC convention. The following diacritics are attested in the dataset:

Level tones:

• High tone (H): acute accent — á, é, ə́, ɛ́, í, ó, ɔ́, ú, ʉ́

• Low tone (L): grave accent — à, è, ə̀, ɛ̀, ì, ò, ɔ̀, ù, ʉ̀

Contour tones:

• Falling tone (HL): circumflex — â, ê, ə̂, ɛ̂, î, ô, ɔ̂, û, ʉ̂

• Rising tone (LH): caron — ǎ, ě, ə̌, ɛ̌, ǐ, ǒ, ɔ̌, ǔ, ʉ̌

Mid tone is generally left unmarked in the Ghomala AGLC orthography. Tone marking is systematic and applies to all nine vowels (a, e, ə, ɛ, i, o, ɔ, u, ʉ) as well as to syllabic nasals. Tone diacritics may accumulate within close vowel sequences and across morpheme boundaries.

Source

The dataset was compiled from scripted speech prompt lists read by a native speaker of Ghomala (jo parler, Bandjoun). Sentences were selected to provide broad phonological coverage and were transcribed in accordance with the AGLC orthographic standard.

Domain

The dataset represents scripted speech in Ghomala (jo parler, Bandjoun variety), covering a broad range of everyday sentence types drawn from a general-purpose TTS/ASR prompt list. All utterances are scripted rather than spontaneous.

Size

Total audio duration: 7,888 seconds (≈2h 11m 28s), distributed across 1,994 accepted audio clips in 20 recording sessions from 2 speakers (MP3 format).

• Speaker 1 (audio_1): 3,564 seconds (59m 24s) across 997 clips in 10 sessions. An additional 15 audio files (1m 07s) are retained in a dedicated attempts subfolder within session tts_dataset_bbj_08, corresponding to rejected recording attempts.

• Speaker 2 (audio_2): 4,324 seconds (≈1h 12m 04s) across 997 clips in 10 sessions. Rejected recording attempts are retained in dedicated attempts subfolders within each session (292 files total, ≈25m 53s).

Total dataset size (audio + mapping files): approximately 95 MB (audio_1: ~34 MB; audio_2: ~61 MB).

Structure

The dataset is organised into two speaker collections, each comprising 10 recording sessions.

Speaker 1 (audio_1) — 10 sessions:

• Session asr-tts_dataset_bbj_01: 100 clips (3m 26s)

• Session asr-tts_dataset_bbj_02: 98 clips (3m 22s)

• Session asr-tts_dataset_bbj_03: 99 clips (3m 43s)

• Session asr-tts_dataset_bbj_04: 100 clips (6m 16s)

• Session asr-tts_dataset_bbj_05: 100 clips (6m 51s)

• Session asr-tts_dataset_bbj_06: 100 clips (5m 27s)

• Session asr-tts_dataset_bbj_07: 100 clips (6m 45s)

• Session tts_dataset_bbj_08: 100 clips (7m 10s)

• Session tts_dataset_bbj_09: 100 clips (8m 38s)

• Session tts_dataset_bbj_10: 100 clips (7m 46s)

Sessions 01–07 are designated for both ASR and TTS use (prefix asr-tts_dataset_bbj); sessions 08–10 are designated for TTS use only (prefix tts_dataset_bbj). Session tts_dataset_bbj_08 additionally contains an attempts subfolder holding 15 rejected recording takes (total duration: 1m 07s).

Speaker 2 (audio_2) — 10 sessions:

• Session asr-tts_dataset_bbj_01: 100 clips (5m 05s)

• Session asr-tts_dataset_bbj_02: 98 clips (5m 30s)

• Session asr-tts_dataset_bbj_03: 99 clips (5m 04s)

• Session asr-tts_dataset_bbj_04: 100 clips (8m 21s)

• Session asr-tts_dataset_bbj_05: 100 clips (8m 48s)

• Session asr-tts_dataset_bbj_06: 100 clips (7m 32s)

• Session asr-tts_dataset_bbj_07: 100 clips (8m 04s)

• Session asr-tts_dataset_bbj_08: 100 clips (7m 55s)

• Session asr-tts_dataset_bbj_09: 100 clips (7m 41s)

• Session asr-tts_dataset_bbj_10: 100 clips (8m 04s)

All 10 sessions are designated for both ASR and TTS use (prefix asr-tts_dataset_bbj). Each session additionally contains an attempts subfolder holding rejected recording takes (292 files total across all sessions, ≈25m 53s).

Each session folder (in both collections) contains:

1. Audio clips (MP3 format)

2. One per-session sentence-to-audio mapping file (mapping.tsv), with 4 columns

Description of columns (mapping.tsv)

• #audio_filename: filename of the audio clip (MP3)

• #key: unique hash identifier of the recording

• #sentence: sentence text as read by the speaker, transcribed in AGLC orthography

• #attempts: number of recording attempts before acceptance

Sample