---
dataset_info:
  features:
  - name: text
    dtype: string
  - name: audio
    dtype: audio
  - name: language
    dtype: string
  - name: dialect
    dtype: string
  - name: speaker_id
    dtype: string
  - name: gender
    dtype: string
  splits:
  - name: train
    num_examples: 2750
  download_size: 118587955
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*.parquet
language:
- am
- om
- sid
- ti
license: cc-by-4.0
task_categories:
- automatic-speech-recognition
- text-to-speech
tags:
- amharic
- afaan-oromo
- sidama
- tigrinya
- africa
- ethiopia
- dialect
- speech
- asr
- tts
pretty_name: Leyu Ethiopian Languages Speech Dataset
---

# Leyu Ethiopian Languages Speech Dataset

Audio recordings paired with corresponding text transcripts, collected on the [Leyu Data Collection Platform](https://leyu.ai) — an open-source platform for crowdsourced speech data collection — for the Leyu Platform Competition, covering 4 languages: Amharic, Afaan Oromo, Sidama, Tigrinya.

## Dataset Summary

- **Languages:** Amharic (`am`), Afaan Oromo (`om`), Sidama (`sid`), Tigrinya (`ti`)
- **Total examples:** 2750
- **License:** [CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/)
- **Task categories:** Automatic Speech Recognition (ASR), Text-to-Speech (TTS)
- **Audio format:** M4A, 10–60 seconds per clip
- **Source:** Live human recordings via the Leyu mobile app — no synthetic or externally-sourced audio

## Prompt Themes & Real-World Use Cases

Prompts were team-authored (see Source Data below), not scraped, so the collection reflects deliberate scenario design rather than whatever happened to be scrapeable. Four themes recur across the dataset, each written as natural, first-person or narrative speech rather than isolated sentences -- closer to how people actually talk than to a word list.

**1. Weather, climate, and environmental impact.** The most represented theme. Prompts cover day-to-day conditions (rain, temperature swings, fog, thunderstorms) and their direct consequences for both cities and farms: rainfall timing and its effect on planting season, flooding risk to water supplies and property, and drought's effect on food security, alongside purely descriptive appreciation of weather (rainbows, changing light, mountain snow).

> *"የዝናብ ውሃ መጠን በዝናብ ምክንያት እየጨመረ ነው። ትላንት ማታ ከባድ ዝናብ ከጣለ በኋላ የከተማችን ወንዞች የውሃ መጠን በከፍተኛ ሁኔታ ጨምሯል። የጎርፍ መጥለቅለቅን ለመከላከል የከተማዋ የአደጋ ጊዜ አስተዳደር ቡድን ጥንቃቄ እያደረገ ነው..."*
> — *"Water levels are rising because of the rain. After heavy rain fell last night, our city's rivers rose sharply. The city's emergency management team is taking precautions to prevent flooding..."* (Amharic, Shewa dialect)

> *"በተደጋጋሚ የሚከሰተው ድርቅ የግብርናውን ዘርፍ ኢኮኖሚያዊ መሰረት እያናጋው ነው። አርሶ አደሮች ከውሃ እጦት የተነሳ ሰብል ማብቀል ተስኗቸዋል። ይህ ደግሞ የምግብ ዋስትና ላይ ከፍተኛ ስጋት ፈጥሯል።"*
> — *"Recurring drought is undermining the agricultural sector's economic foundation. Farmers have been unable to grow crops due to water scarcity, creating a serious threat to food security."* (Amharic, Shewa dialect)

**2. Education and academic life.** Exam-week stress, study routines, forming study groups, and — a recurring, deliberate counterpoint — explicit advice on rest, sleep, and avoiding burnout while studying.

> *"የፈተናዬ ቀን ተቃርቧል። ዛሬ ጠዋት ከቡናዬ ጋር ተቀምጬ የነበረ ሲሆን፣ የጥናት ማስታወሻዎቼን የመጨረሻ ግምገማ እያደረግኩ ነው... ትንሽ ፍርሃት ቢሰማኝም፣ ላለፉት ወራት ያጠፋሁት ጥረት ውጤታማ እንደሚሆን አምናለሁ።"*
> — *"My exam day is approaching. This morning I sat with my coffee doing a final review of my study notes... Even though I feel a little nervous, I believe the effort I've put in over the past months will pay off."* (Amharic, Shewa dialect)

**3. Ethiopian culture, tradition, and history.** The coffee ceremony as a social and conversational center, holiday food preparation (doro wat, tibs, injera) and gatherings, and descriptions of pre-modern travel, house-building, and craftsmanship.

> *"በተለይ በዓላት ላይ ባህላዊ ምግቦች ልዩ ቦታ አላቸው። የትንሳኤ፣ የገና፣ የፋሲካ በዓላት ላይ የሚዘጋጁት የዶሮ ወጥ፣ የጥብስ፣ የእንጀራና የዳቦ አይነቶች ቤተሰብንና ወዳጆችን ያገናኛሉ። የቡና ሥነ-ሥርዓትም የደስታና የውይይት ማዕከል ይሆናል።"*
> — *"Traditional foods hold a special place, especially during holidays. The doro wat, tibs, injera, and bread prepared for Easter, Christmas, and other holidays bring family and friends together. The coffee ceremony becomes a center of joy and conversation."* (Amharic, Shewa dialect)

**4. Health, clothing, and daily routines tied to environment.** The everyday reasoning behind dressing for conditions (light cotton for heat, layering for cold snaps), and health guidance tied to weather — hydration and sun exposure in hot months, staying warm during sudden temperature drops.

> *"ዛሬ የነበረው እርጥበት በጣም ከፍተኛ ነበር፤ ለዚህም ነው የሰውነት ሙቀት መጨመር የተሰማኝ። ጠንካራ ልብሶችን ከመልበስ ይልቅ ቀለል ያሉና አየር የሚያስገቡ ልብሶችን መልበስ ይመከራል። ብዙ ውሃ መጠጣት ሰውነትን ለማቀዝቀዝ ይረዳል፤ ይህም ለጤና ጠቃሚ ነው።"*
> — *"Today's humidity was very high, which is why I felt my body temperature rising. Instead of heavy clothing, it's advisable to wear light, breathable clothes. Drinking plenty of water helps cool the body, which is good for health."* (Amharic, Shewa dialect)

**Why this matters beyond speech-tech benchmarking**: recordings built around real weather, agricultural, and health scenarios — read aloud in the exact dialects spoken by the communities most exposed to flooding, drought, and heat stress — are directly reusable training/evaluation data for applications those communities can't currently access in their own language:

- **Voice-based early-warning and disaster-preparedness systems** (flood/drought alerts) for low-literacy, low-connectivity populations who are underserved by text-based or English/Amharic-only alerting.
- **Agricultural advisory tools** — rainfall-timing and planting-season guidance delivered by voice, in-dialect, to smallholder farmers.
- **Accessible climate and weather information delivery** for rural communities, where radio and voice remain the dominant information channel.
- **ASR/TTS model training and evaluation** for Ethiopian languages that are essentially absent from existing benchmark datasets (Afaan Oromo, Tigrinya, Sidama, alongside Amharic).
- **Education-focused conversational AI** (study companions, exam-prep voice assistants) grounded in how students actually describe their own study habits and stress.
- **Cultural-heritage digitization and multilingual content generation** — the coffee ceremony, holiday traditions, and oral-history-adjacent material in this dataset are exactly the kind of content most at risk of never being digitized in the original language.
- **Public-health voice guidance** (heat/hydration advisories, seasonal health reminders) deliverable in-language at the community level.

## Breakdown by language and dialect

| Language | Dialect | Clips | Hours | Speakers | Male | Female |
|---|---|---|---|---|---|---|
| Afaan Oromo | Western (Wollega/Shoa) Dialect | 278 | 1.58 | 2 | 0 | 278 |
| Afaan Oromo | East/Western (Hararghe) | 108 | 1.12 | 1 | 108 | 0 |
| Amharic | Shewa Dialect | 1854 | 12.57 | 7 | 1386 | 468 |
| Amharic | Wollo Dialect | 348 | 2.09 | 2 | 0 | 348 |
| Sidama | Hawella | 78 | 0.59 | 2 | 78 | 0 |
| Tigrinya | Standard Tigrinya | 84 | 0.44 | 2 | 0 | 84 |
| **Total** | | **2750** | **18.39** | | | |

## Dataset Structure

Each row is one read-aloud recording:

| Column | Type | Description |
|---|---|---|
| `text` | string | The prompt text the speaker read aloud |
| `audio` | audio | The recorded clip (10–60 seconds) |
| `language` | string | ISO 639 code of the language (`am`, `om`, `ti`, `sid`) |
| `dialect` | string | Named dialect of the recording |
| `speaker_id` | string | Anonymized contributor identifier |
| `gender` | string | Speaker-reported gender (`Male` / `Female`) |

## Source Data

Prompts were team-authored (not scraped or extracted from copyrighted third-party text) and recorded by volunteer contributors through the Leyu mobile app. Every submission passed contributor recording, then human reviewer/PM approval, before being included here — this repo only contains rows with `status = 'Approved'` on the live Leyu platform database, matching the platform's own audit logs.

## Considerations

- Data collection is ongoing for the Leyu Platform Competition (through 2026-08-21); this dataset is updated as more audio is approved.
- Dialect and speaker coverage is still growing — see the breakdown table above for current representation per language.

## Licensing and Attribution

Released under [CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/). Collected for the **Leyu Platform Competition**, sponsored by gheero (Leyu Platform Team), Addis Ababa, Ethiopia.