| --- |
| license: cc-by-sa-4.0 |
| language: |
| - am |
| - om |
| - so |
| - ti |
| task_categories: |
| - automatic-speech-recognition |
| pretty_name: Horn of Africa ASR Benchmark |
| tags: |
| - speech |
| - asr |
| - low-resource |
| - africa |
| - amharic |
| - oromo |
| - somali |
| - tigrinya |
| size_categories: |
| - 1K<n<10K |
| configs: |
| - config_name: am |
| data_files: |
| - split: test |
| path: am/test/metadata.csv |
| - config_name: om |
| data_files: |
| - split: test |
| path: om/test/metadata.csv |
| - config_name: so |
| data_files: |
| - split: test |
| path: so/test/metadata.csv |
| - config_name: ti |
| data_files: |
| - split: test |
| path: ti/test/metadata.csv |
| --- |
| |
| # Horn of Africa ASR Benchmark |
|
|
| A multilingual evaluation benchmark for automatic speech recognition |
| covering four under-served languages of the Horn of Africa: |
| **Amharic**, **Oromo**, **Somali**, and **Tigrinya**. Each language |
| ships **1,000 evaluation utterances** drawn from spontaneous |
| interview-style speech with reference transcripts post-edited and |
| QC-validated by native-speaker annotators. |
|
|
| - **Total**: 4000 utterances, 15.44 hours, 975 distinct speakers (interview proxy). |
| - **Audio**: 16 kHz mono PCM 16-bit WAV (metadata stripped). |
| - **License**: CC-BY-SA 4.0. |
|
|
| ## Dataset Statistics |
|
|
| | Language | Code | Utterances | Hours | Speakers | Gender (gender_qa) | |
| |---|---|---|---|---|---| |
| | Amharic | am | 1000 | 4.38 | 524 | male 518 / female 482 | |
| | Oromo | om | 1000 | 4.3 | 572 | male 503 / female 497 | |
| | Somali | so | 1000 | 3.32 | 294 | male 523 / female 477 | |
| | Tigrinya | ti | 1000 | 3.44 | 320 | male 500 / female 498 / unknown 2 | |
| |
| ### Dialect coverage |
| |
| | Language | Dialect labels | |
| |---|---| |
| | Amharic | Addis Ababa, Gojam, Gonder, L2, Shewa, Unknown, Wollo | |
| | Oromo | Eastern Oromo, Southern Oromo, West Central Oromo | |
| | Somali | Benadiri Somali, Northern Somali (Ogaadeen), Northern Somali (Puntland), Northern Somali (Somaliland), Other | |
| | Tigrinya | D, L, Z | |
| |
| ## Dataset Structure |
| |
| ``` |
| hf/ |
| ├── am/test/ 1,000 .wav files + metadata.csv |
| ├── om/test/ 1,000 .wav files + metadata.csv |
| ├── so/test/ 1,000 .wav files + metadata.csv |
| └── ti/test/ 1,000 .wav files + metadata.csv |
| ``` |
| |
| A single `test` split per language. Loading via `datasets`: |
| |
| ```python |
| from datasets import load_dataset |
| am = load_dataset("LesanAI/Horn-ASR", "am", split="test") |
| print(am[0]["audio"]["array"].shape, am[0]["transcript"]) |
| ``` |
| |
| ### Fields (per row) |
| |
| | field | type | description | |
| |---|---|---| |
| | `audio` | `Audio(sampling_rate=16000)` | 16 kHz mono PCM WAV | |
| | `utterance_id` | str | `{lang}_{phase}_{interview}_{segment}` | |
| | `transcript` | str | reference transcript (annotator-edited, single-line) | |
| | `speaker_id` | str | interview ID; upper-bound speaker proxy (interviews are filtered to single-speaker but the ID itself is the interview ID) | |
| | `gender` | str | `male`, `female`, or `unknown` (annotator-verified) | |
| | `dialect` | str | per-language dialect label (see classifications below) | |
| | `domain` | str | content domain bucket (15-way controlled vocabulary) | |
| | `duration_s` | str | utterance duration in seconds | |
| | `lang` | str | ISO 639 code (`am` / `om` / `so` / `ti`) | |
|
|
| ## Data Collection |
|
|
| Source audio is interview-style spontaneous speech segmented from |
| public-domain interview material. Native-speaker annotators |
| post-edited automatic-transcription drafts (or transcribed from |
| scratch where no draft existed) and a separate QC pass verified |
| transcript correctness. A second annotation round added/verified |
| gender, dialect, domain, and region labels. Segments flagged for |
| multi-speaker contamination, mis-aligned transcripts, hate speech, |
| read-speech (scripted reading instead of spontaneous), or content |
| in another language were dropped and replaced from a tier-2 |
| diversity-aware backup pool, preserving per-language gender balance. |
|
|
| Audio was originally encoded at low bitrate (24 kbps mp3) which |
| constrains acoustic quality; the WAV files in this release are |
| upsampled to 16 kHz mono but inherit the source bandwidth. ASR |
| systems should expect noisy, conversational, accented speech. |
|
|
| ## Dialect Classification |
|
|
| Per-language label sets follow these references: |
|
|
| - **Amharic**: Mengistu Tadese (2018). *Documentation and Description of Amharic Dialects.* PhD thesis, Department of Linguistics: Documentary Linguistics and Culture Program, Addis Ababa University. Labels: Addis Ababa, Shewa, Gojam, Gonder, Wollo, L2 (non-native), Unknown. |
| - **Oromo**: Stroomer (1995), Owens (1985), Griefenow-Mewis (2001). Three principal dialect zones: West Central Oromo (Mecha-Tulama), Eastern Oromo (Hararghe), Southern Oromo (Borana-Arsi-Guji). |
| - **Somali**: Lamberti (1986); Saeed (1999). Northern Somali sub-varieties (Somaliland, Puntland, Ogaadeen), Benadiri Somali (Mogadishu coastal), and Other (diaspora + minor varieties). Maay is not represented. |
| - **Tigrinya**: Asfaw Gedamu Haileslasie, Asmelash Teka Hadgu, Solomon Teferra Abate (2023). *Tigrinya Dialect Identification.* AfricaNLP @ ICLR 2023. Three abstract dialect codes: Z, L, D. |
|
|
| ## Intended Use |
|
|
| Benchmarking ASR systems on under-resourced Horn-of-Africa |
| languages — single-language evaluation, multilingual |
| joint-training evaluation, and zero-shot transfer studies. The |
| test-only structure is deliberate: this is an evaluation |
| benchmark, not a fine-tuning corpus. |
|
|
| ## Limitations and Bias |
|
|
| - **Source bandwidth**: 24 kbps mp3 source; high-frequency content |
| is band-limited. |
| - **Speaker coverage**: speaker IDs are interview IDs and are an |
| upper bound on distinct speakers; some interviews host multiple |
| speakers despite the single-speaker filter. |
| - **Dialect imbalance**: dialect cell sizes within a language are |
| uneven (e.g., Tigrinya Z=506 vs D=196). Per-dialect WER reports |
| on small cells will have wide confidence intervals. |
| - **Domain skew**: Politics and Social are overrepresented; some |
| domains have <50 utterances per language. |
| - **Maay** Somali and **Wallo / Menz** Amharic are not represented. |
| - **Gender** labels are annotator audio judgments; mixed-gender |
| interview segments were filtered out via single-speaker selection. |
|
|
| ## License |
|
|
| This dataset is released under the **Creative Commons |
| Attribution-ShareAlike 4.0 International License** (CC-BY-SA 4.0). |
| You may share and adapt the work for any purpose, including |
| commercially, provided you give appropriate credit and distribute |
| your contributions under the same license. |
|
|
| License text: `LICENSE` in this repository, or |
| https://creativecommons.org/licenses/by-sa/4.0/. |
|
|
| ## Citation |
|
|
| If you use this benchmark, please cite: |
|
|
| ``` |
| @inproceedings{horn_asr_2026, |
| title = {Horn of Africa ASR Benchmark}, |
| author = {Anonymous}, |
| year = {2026}, |
| note = {Under review}, |
| } |
| ``` |
|
|
| And the dialect-classification sources cited in the Dialect |
| Classification section above. |
|
|
| ## Reproducibility |
|
|
| The pipeline that produced this release is publicly available; see |
| the `code/` tree (stages 01–08). To rebuild from scratch: |
|
|
| 1. Run the QC analysis stages (`code/01_clean` → `code/03_qc_analysis`). |
| 2. Run the tag-merge + final-replacement (`code/06_tags`). |
| 3. Run `bash code/08_hf/run.sh` to produce this directory. |
|
|
| The 4,000 frozen utterance IDs are checked in at |
| `data/manifests/ids_frozen.tsv` so the release is bit-exact |
| reproducible from the raw audio. |
|
|
