| --- |
| language: |
| - ha |
| - hau |
| license: cc-by-nc-4.0 |
| task_categories: |
| - automatic-speech-recognition |
| size_categories: |
| - 1K<n<10K |
| tags: |
| - ha |
| - audio |
| - speech |
| - asr |
| - africa |
| - nigeria |
| - hausa |
| - gamayun |
| - low-resource |
| - read-speech |
| pretty_name: "Gamayun Hausa Voice Mini-kit" |
| download_size: 229465898 |
| dataset_size: 230429993 |
| configs: |
| - config_name: read |
| data_files: |
| - split: train |
| path: data/read/train-*.parquet |
| --- |
| |
| # Gamayun Hausa Voice Mini-kit |
|
|
| 4,986 read-speech recordings in Hausa (`ha`) with their English source sentences, recorded by paid CLEAR Global voice talent on mobile phones or desktop computers in 2021. The clips cover the sentence set of the Hausa–English mini-kit (`kit5k`), the recording arm of CLEAR Global's Gamayun language-data initiative. |
|
|
| ## Languages |
|
|
| - `ha` — Hausa |
|
|
| ## Dataset Structure |
|
|
| Recordings are embedded in the Parquet shards under `data/read/`, one row per clip, in the `audio` column (HuggingFace `Audio` feature, mono 48 kHz MP3). |
|
|
| ```python |
| from datasets import load_dataset |
| |
| ds = load_dataset("CLEAR-Global/Gamayun-hausa-voice-minikit", "read", split="train") |
| ``` |
|
|
| | field | description | |
| |---|---| |
| | `audio` | The recording (HuggingFace `Audio` feature). `audio.path` is the original file name. | |
| | `id` | Tatoeba sentence ID of the English source sentence. | |
| | `eng` | English source sentence (Tatoeba-derived, from the `kit5k` core). | |
| | `hau` | Hausa sentence that was read aloud (from the Hausa `kit5k` translation). | |
| | `duration` | Clip duration in seconds. | |
|
|
| ## Statistics |
|
|
| ### `read` |
|
|
| | metric | value | |
| |---|---:| |
| | Rows | 4986 | |
| | Distinct audio files | 4986 | |
| | Total hours (approved) | 6.68 | |
| | Approved / pending / rejected | 4986 / 0 / 0 | |
| | Median / p95 duration (s) | 4.36 / 8.65 | |
|
|
| **Clips and hours by split** |
|
|
| | split | clips | hours | |
| |---|---:|---:| |
| | `train` | 4986 | 6.68 | |
|
|
| **Hours by approval status** |
|
|
| | status | hours | |
| |---|---:| |
| | approved | 6.68 | |
|
|
| ## Dataset Creation |
|
|
| ### Data Collection |
|
|
| Produced by CLEAR Global (formerly Translators without Borders) in **2021** as part of its [Gamayun](https://huggingface.co/datasets/CLEAR-Global/Gamayun-kits) initiative. English sentences were drawn from the [Tatoeba](https://tatoeba.org) corpus; the Hausa translations were produced by paid linguists in the CLEAR Global translator community, and paid voice talent read the sentences aloud on mobile phones or desktop computers so that the audio reflects real-world, low-resource and noisy usage. Recordings were checked in-house; this release keeps only the sentences that come from the Gamayun kits and keeps one recording per sentence. |
|
|
| ### Quality Control |
|
|
| Every recording in this release was checked in-house by CLEAR Global for audio quality and for agreement with the prompt text, and all 4,986 clips passed that review, so every row is published (`train`); no pending or rejected rows are included. The audio decodes cleanly and is mono MP3 at 48 kHz; durations range from roughly 1 second to 18 seconds (median ~4.4 s). |
|
|
| ## Ethical Considerations |
|
|
| - All speakers were paid, professional voice talent who consented to recording |
| and to open publication of the recordings. |
| - Speakers are not identified in this release; no speaker or personal data is |
| included. |
| - It is forbidden to attempt to identify individual speakers in this dataset. |
| - The data should be used responsibly, for research and educational purposes. |
| - The audio is derived from prompts that CLEAR Global has the right to license; |
| the English source sentences come from the Tatoeba corpus. |
|
|
| ## Caveats |
|
|
| - The recordings were captured on mobile phones or desktop computers and are not studio quality; expect channel noise, reverberation and varying levels. |
| - This is a curated subset. It was carved out of the larger *Hausa Kato* voice kit, keeping only the recordings whose transcript is a Gamayun `kit5k` sentence; rows that trace to a different, non-Gamayun source were removed and are not published here. |
| - Where a Gamayun sentence had more than one recording, only one is kept, so there is exactly one clip per unique sentence (4,986 clips for 4,986 sentences). |
| - No speaker or gender metadata exists for these recordings, so, unlike the Congolese Swahili mini-kit, none is provided and the recordings cannot be partitioned by speaker. Attempting to identify individual speakers is forbidden. |
| - The sibling text release is the Hausa–English kit itself; see [Gamayun Language Data Kits](https://huggingface.co/datasets/CLEAR-Global/Gamayun-kits). |
| - The Congolese Swahili mini-kit recorded from the same initiative is published as [`CLEAR-Global/Gamayun-swc-voice-minikit`](https://huggingface.co/datasets/CLEAR-Global/Gamayun-swc-voice-minikit). |
|
|
| ## License |
|
|
| CC BY-NC 4.0 (Creative Commons Attribution-NonCommercial 4.0 International) (SPDX: `cc-by-nc-4.0`) |
|
|
| ## Citation |
|
|
| ```bibtex |
| @inproceedings{oktem2020gamayun, |
| author = {Öktem, Alp and Albayk Jaam, Muhannad and DeLuca, Eric and Tang, Grace}, |
| title = {Gamayun - Language Technology for Humanitarian Response}, |
| booktitle = {2020 IEEE Global Humanitarian Technology Conference (GHTC)}, |
| year = {2020}, |
| doi = {10.1109/GHTC46280.2020.9342939}, |
| url = {https://ieeexplore.ieee.org/document/9342939} |
| } |
| ``` |
|
|
| ## Contact |
|
|
| CLEAR Global — `tech@clearglobal.org` |
|
|
| ## Acknowledgments |
|
|
| Recorded by paid CLEAR Global voice talent in Nigeria. Hausa translations by paid linguists in the CLEAR Global (formerly Translators without Borders) translator community; English source sentences from [Tatoeba](https://tatoeba.org). Produced as part of the [Gamayun](https://huggingface.co/datasets/CLEAR-Global/Gamayun-kits) language-data initiative. |
|
|
