---
dataset_info:
  features:
  - name: text
    dtype: string
  - name: audio
    dtype: audio
  - name: dialect
    dtype: string
  - name: speaker_id
    dtype: string
  - name: gender
    dtype: string
  splits:
  - name: train
    num_bytes: 889332942
    num_examples: 4862
  download_size: 480207301
  dataset_size: 889332942
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
language:
- am
license: cc-by-4.0
task_categories:
- automatic-speech-recognition
- text-to-speech
tags:
- amharic
- africa
- ethiopia
- dialect
- speech
- asr
- tts
pretty_name: Leyu Amharic Wello Dialect
---

# Leyu Amharic - Wello Dialect Speech Corpus

## Dataset Description

This dataset is a curated parallel speech corpus consisting of audio recordings paired with corresponding text transcripts, focused on the **Wello dialect** of the Amharic language. It is designed to support speech technology research across multiple tasks, including Automatic Speech Recognition (ASR) and Text-to-Speech (TTS). The corpus captures dialect-specific phonetic variations, accent patterns, and prosodic characteristics unique to the Wello region.

All recordings were collected in real-world environments using mobile devices, introducing natural acoustic variability that improves model generalization. Each audio–text pair underwent manual review to ensure transcript accuracy and audio clarity.

## Dataset Summary
| Metadata Field | Value |
|----------------|--------|
| Language | Amharic (am-ET) |
| Dialect | Wello |
| Audio Format | .wav |
| Total Hours | 114.7 Hours |
| Recording Environment | Mobile / Crowdsourced (Verified) |

## Key Statistics
- **Total Duration:** 114.7 Hours  
- **Speaker Count:** 68 Speakers (Male:27, Female: 41)
- **Data Quality:** Mobile-recorded and manually verified for transcript alignment.

## Data Collection & Quality Assurance
- Recorded by contributors using mobile devices in real-world environments.
- Manually reviewed to ensure transcript alignment and audio clarity.

## Data Fields
Each sample contains:
- `text` (string): transcript
- `audio` (Audio): waveform/audio file
- `dialect` (string): `wello`
- `speaker_id` (string): anonymized speaker identifier
- `gender` (string): `male` / `female` / `unknown`

## iCog Blogs

- [Leyu: Crowdsourcing Datasets for Ethiopian Languages](https://leyu.ai/blog/1)
- [Dialects & Socioeconomics: Shaping Inclusive Language Models](https://leyu.ai/blog/3)
- [Progress of Natural Language Processing (NLP) for Ethiopian Languages – Part One](https://leyu.ai/blog/4)
- [Progress of Natural Language Processing (NLP) for Ethiopian Languages – Part Two](https://leyu.ai/blog/22)
- [Data Collection with Purpose: The Leyu Approach](https://leyu.ai/blog/23)