Modelpublic

badrex/Ethio-ASR-multilingual-94M

sourceHugging Facecc-by-nc-sa-4.0updated 6mo agoView on Hugging Face

0likes69downloads

Model Card

<div align="center" style="line-height: 1; transform: scale(1.4);"> <a href="https://huggingface.co/datasets/google/WaxalNLP" target="blank" style="margin: 2px;"> <img alt="Hugging Face" src="https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Dataset-ffc107?color=ffca28&logoColor=white" style="display: inline-block; vertical-align: middle;"/> <a href="https://huggingface.co/spaces/badrex/Ethio-ASR-multilingual-demo" target="blank" style="margin: 2px;"> <img alt="Hugging Face" src="https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Space-ffc107?color=c62828&logoColor=white" style="display: inline-block; vertical-align: middle;"/> <a href="https://creativecommons.org/licenses/by/4.0/deed.en" style="margin: 2px;"> <img alt="License" src="https://img.shields.io/badge/License-CC%20BY%20NC%20SA%204.0-lightgrey.svg" style="display: inline-block; vertical-align: middle;"/> </a> </div> <h3 style="text-align: center; font-size: 24px; color:#C70039;"> arXiv 📖 <a href="https://arxiv.org/pdf/2603.23654">[ preprint ]</a> </h3>

⚒️ Model Description

Ethio-ASR is a suite of multilingual Automatic Speech Recognition (ASR) models that support five Ethiopian languages: Amharic, Tigrinya, Afaan Oromo, Sidama, and Wolaytta. The ASR model in this repo is based on the AfriHuBERT pre-trained model by fine-tuning it on the WAXAL Speech Dataset.

—Developed by: Ethio-ASR Team
—Task: Speech Recognition (ASR) and Language Identification (LID)
—Languages: Amharic, Tigrinya, Afaan Oromo, Sidama, and Wolaytta
—License: CC-BY-NC-SA 4.0
—Finetuned from: ajesujoba/AfriHuBERT

📈 Evaluation on WAXAL Test Set

📌 ASR model in this HF repo | Model | # Params | Amharic | Tigrinya | Oromo | Wolaytta | Sidaama | Avg. | |------|----------|----------|-----------|--------|-----------|----------|------| | Ethio-ASR (afrihubert) 📌 | 94M | 30.95 | 42.42 | 27.57 | 40.44 | 34.02 | 35.08 | | Ethio-ASR (mms-300) | 300M | 30.19 | 41.62 | 26.41 | 39.10 | 32.66 | 33.99 | | Ethio-ASR (mms-1b) | 1B | 26.14 | 37.63 | 23.69 | 37.51 | 31.02 | 31.20 | | Ethio-ASR (w2v-bert-2.0) | 600M | 22.92 | 35.22 | 24.44 | 38.19 | 31.65 | 30.48 |

🎧 Direct Use

python

from transformers import AutoModelForCTC, AutoProcessor
import torchaudio, torch

processor = AutoProcessor.from_pretrained("badrex/Ethio-ASR-multilingual-94M")
model = AutoModelForCTC.from_pretrained("badrex/Ethio-ASR-multilingual-94M")

audio, sr = torchaudio.load("audio.wav")
inputs = processor(audio.squeeze(), sampling_rate=sr, return_tensors="pt")

with torch.no_grad():
    logits = model(**inputs).logits

pred_ids = torch.argmax(logits, dim=-1)
transcription = processor.batch_decode(pred_ids)[0]

print(transcription)

🔧 Downstream Use

—Voice assistants
—Accessibility tools
—Research baselines

🚫 Out‑of‑Scope Use

—Languages outside Amharic, Tigrinya, Afaan Oromo, Sidama, and Wolaytta.
—High‑stakes deployments without human review
—Noisy audio without speech enhancement

⚠️ Risks & Limitations

Performance might vary across dialects, genders, ages, and recording quality.

📌 Citation

bibtex

@misc{ethio_asr_2026,
  author = {
    Abdullah, Badr M. and
    Azime, Israel Abebe and
    Tonja, Atnafu Lambebo and
    Alabi, Jesujoba O. and
    Alemu, Abel Mulat and
    Hagos, Eyob G. and
    Balcha, Bontu Fufa and
    Nerea, Mulubrhan A. and
    Yadeta, Debela Desalegn and
    Marilign, Dagnachew Mekonnen and
    Fentahun, Amanuel Temesgen and
    Kebede, Tadesse and
    Gebru, Israel D. and
    Woldeyohannis, Michael Melese and
    Sewunetie, Walelign Tewabe and
    Möbius, Bernd and
    Klakow, Dietrich
  },
  title = {Ethio-ASR: Joint Multilingual Speech Recognition and Language Identification for Ethiopian Languages},
  year = {2026},
  howpublished = {\url{https://huggingface.co/badrex/Ethio-ASR-multilingual-94M}}
}