XIMA-2122

DOVEXAI Nigerian Languages Speech & Translation Dataset (XIMA-2122)

Curated by DOVEXAI LTD (Nigeria) · dovexai.africa · dovexai.io

A provenance-tracked dataset of Nigerian-language voice recordings paired with human transcriptions and translations, built for evaluation and supervised fine-tuning (SFT) of speech and translation models. Every recording is cryptographically receipted, pseudonymised, and quality-checked before release.

Dataset snapshot

Note: This release is a 52-example evaluation and pipeline verification sample, demonstrating DOVEXAI's end-to-end cryptographic provenance and SFT structure. It is designed for pipeline testing, format validation, and evaluation—not as a standalone training corpus.

Stated Limitations & Intended Use

• Evaluation Sample Scope: This open sample is intentionally limited to 52 clips across 5 contributors to demonstrate our data format, transcription standards, and cryptographic verification protocol.
• Demographic Scope: 100% of contributors in this open sample are female speakers. DOVEXAI's full commercial datasets cover gender-balanced, multi-region, and multi-dialect demographics across Nigeria and West Africa.
• Commercial Holdings: Production datasets (thousands of hours) remain unreleased open-source and are available for enterprise licensing and custom AI training.

How to use

from datasets import load_dataset

ds = load_dataset("DOVEXAI/XIMA-2122")

# Play/access audio
sample = ds["sample"][0]
print(sample["audio"])          # {'array': [...], 'sampling_rate': 44100, 'path': ...}
print(sample["text_en_ng"])     # English prompt
print(sample["text_local"])     # Local-language transcription
print(sample["language"])       # e.g. "Yoruba", "Hausa", "Igbo"

Need custom data?

DOVEXAI builds training and evaluation data across every modality - text, audio, image, video, document and multimodal - including collection, transcription, translation, data generation, SFT and RLHF preference data, human evaluation and red-teaming packs. Our differentiated coverage is Nigerian and other African languages, accents and code-switching. XIMA-2122 is a small public sample of that coverage; our unreleased holdings are substantially larger.

• Fastest route: Book a 30-minute scoping call
• Or email: [email protected]

Community & Feedback

We'd love to hear from you! If you use this dataset, please share:

• Your feedback on the quality, usefulness, and potential improvements
• Research or projects built with this data
• Contributions - if you'd like to contribute additional recordings, transcriptions, or translations for Nigerian or other African languages

Join the DOVEXAI Early Access Contributor Community via our website at dovexai.africa or dovexai.io.

Dataset columns

Important: verbatim vs. standardized. text_local is the ground truth for what was actually spoken and should be used as the transcription label. text_local_standard, when present, is an editorial standardization and may differ from the audio.

Recording structure

Each recording falls into one of three types, indicated by the speech_type column:

1. Sequential bilingual parallel corpus (speech_type = "sequential bilingual parallel corpus")

The contributor first reads the prompt in Nigerian English, then immediately speaks the standard local-language translation in the same clip:

[ Nigerian English sentence ]  →  [ Hausa / Igbo / Nigerian Pidgin translation ]

This applies to: - DVX-XIMA-2122-001 (all 10 clips): English → Hausa - DVX-XIMA-2122-002 (all 10 clips): English → Igbo - DVX-XIMA-2122-003 clips 2, 7, 8, 9, 10: English → Nigerian Pidgin - DVX-XIMA-2122-004 (all 11 clips): English → Nigerian Pidgin

2. Sequential bilingual code-switched parallel corpus (speech_type = "sequential bilingual code-switched parallel corpus")

The contributor first reads the English prompt, followed immediately by a natural, intra-sententially code-switched Yoruba translation with cultural loanwords/metaphors, and where applicable, a concluding Yoruba proverb:

[ Nigerian English sentence ]  →  [ Code-switched Yoruba translation ]  (→ [ Yoruba proverb ])

This applies to: - DVX-XIMA-2122-005 (all 11 clips): English → Code-switched Yoruba (+ Proverbs)

3. Monolingual (speech_type = "monolingual")

The contributor speaks Nigerian English only. There is no local-language translation in the clip.

• DVX-XIMA-2122-003 clips 1, 3, 4, 5, 6: purely Nigerian English

For these recordings, local_lang_key is nigerian-english and text_local contains the same English text.

Additional files

This dataset also ships the following raw files for advanced use cases:

sft.jsonl                          supervised pairs: audio path → text  (raw format)
audio/<receipt_id>/audio_NNN.m4a   original recordings, byte-for-byte as receipted (renamed)
audio/<receipt_id>/audio_NNN.wav   16 kHz 16-bit mono WAV converted from the matching .m4a
receipts/<receipt_id>.json         per-contributor, hash-locked provenance receipts
receipts/index.json                master manifest (tamper-evident)

Provenance & integrity

• Each receipt lists every source file with its SHA-256 and carries a receipt_hash computed over the receipt's contents; index.json carries an index_hash over the manifest. Recomputing these detects any tampering. Both hashes are SHA-256 of the JSON object with that hash field removed, serialised with sorted keys and compact separators (Python: json.dumps(obj, sort_keys=True, separators=(",", ":"))).
• The original recordings in audio/*.m4a are byte-for-byte the files that were receipted, so each one's SHA-256 matches its receipt entry. The .wav files and the Parquet audio are 16 kHz 16-bit mono conversions of those originals, with no denoising, trimming or normalisation; each row's source_sha256 links it to its receipt entry.
• Every recording passed automated noise-floor and clipping checks before inclusion; each receipt records the measured values and thresholds per file.

Privacy & consent

• Pseudonymised. Contributors appear only as opaque IDs (contributor_NNN); real names and original filenames are never included. The key linking IDs to identities is held privately by DOVEXAI and never shipped. Voice recordings remain personal data under the Nigeria Data Protection Act (NDPA) 2023.
• Contributors recorded under signed consent and NDA agreements. Those documents are not shipped. Each receipt proves they exist via their SHA-256 hash alone, with no personal data exposed.

Changelog

• 2026-10-15:
• Expanded evaluation sample from 41 to 52 audio clips across 5 contributors by onboarding contributor_004 (DVX-XIMA-2122-004, Nigerian Pidgin).
• Shipped byte-for-byte original .m4a recordings alongside 16 kHz 16-bit mono .wav files under audio/.
• Added source_sha256 linking each dataset row directly to its entry in the contributor's cryptographic provenance receipt.
• Recorded automated acoustic quality checks (RMS noise floor and peak clipping) across all receipts.
• Removed unpopulated preferences.jsonl and deprecated legacy Parquet split.

License

Released under Creative Commons Attribution 4.0 International (CC BY 4.0). You are free to use, share, and adapt this dataset, including for commercial purposes and for training models, provided you give appropriate credit to DOVEXAI LTD (Nigeria). Full terms: https://creativecommons.org/licenses/by/4.0/

Citation

If you use this dataset in your research, product evaluation, or model training, please cite it as follows:

@misc{dovexai2026xima2122,
  author       = {DOVEXAI LTD (Nigeria)},
  title        = {XIMA-2122: Nigerian Languages Speech \& Translation Dataset},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/DOVEXAI/XIMA-2122}},
  note         = {Licensed under CC BY 4.0}
}

Plain text:

DOVEXAI LTD (Nigeria) (2026). XIMA-2122: Nigerian Languages Speech & Translation Dataset. Hugging Face Datasets. https://huggingface.co/datasets/DOVEXAI/XIMA-2122

Contact

For questions or commercial dataset licensing, contact DOVEXAI LTD (Nigeria) via [email protected] or schedule a call at https://calendly.com/dovexai-io/30min.