Model Description 🍋🟩
This model is a fine-tuned version of Wav2Vec2-BERT 2.0 for Zulu automatic speech recognition (ASR). It was trained on the 250 hours of transcribed Zulu speech. The ASR model is robust and the in-domain WER is below 16.3%.
- Developed by: Badr al-Absi
- Model type: Speech Recognition (ASR)
- Language: Zulu (zu)
- License: CC-BY-4.0
- Finetuned from: facebook/w2v-bert-2.0
Examples 🚀
| Audio | Human Transcription | ASR Transcription | |
|---|---|---|---|
| 1 | Yenza isinqumo ngezilimo uzozitshala kumaphi amasimu uphinde idwebe imephu njengereferensi yakho. | yenza isinqumo ngezilimo ozozitshala kumaphi amasimu uphinde igwebe imephu njengereference yakho | |
| 2 | Emdlalweni wokugcina ngokumelene IFrance, wayengumuntu ongasetshenziswanga esikhundleni njengoba i-Argentina inqobe ngo-4-2 nge-penalty ukuze ithole isiqu sayo sesithathu seNdebe Yomhlaba. | emdlalweni wokugqina ngokumelene i-france wayengumuntu ongasetshenziswanga esikhundleni njengoba i-argentina incobe ngo-4-2 ngephelnathi ukuze ithole isiqu sayo sesithathu sendebe yomhlaba | |
| 3 | Amadolobhana angaphandle angaphezu kwamamitha ambalwa, Reneging cishe 140m, amamitha angu-459.3, ngaphezu kogu lolwandle. Le ndawo iningi emahlathini ama-dune asogwini, ikakhulukazi eceleni kwe-zindunduma zasogwini nasedolobheni lase-Meerensee. | amadolobhana angaphandle angaphezu kwamamitha ambalwa reneging cishe 140m amamitha angu 4593 ngaphezu kogulolwandle le ndawo iningi emahlabathini amedum esogwini ikakhulukazi eceleni kwezindunduma zasogwini nasedolobheni lasemerins |
Direct Use
The model can be used directly for automatic speech recognition of Zulu audio:
from transformers import Wav2Vec2BertProcessor, Wav2Vec2BertForCTC
import torch
import torchaudio
# load model and processor
processor = Wav2Vec2BertProcessor.from_pretrained("badrex/w2v-bert-2.0-zulu-asr")
model = Wav2Vec2BertForCTC.from_pretrained("badrex/w2v-bert-2.0-zulu-asr")
# load audio
audio_input, sample_rate = torchaudio.load("path/to/audio.wav")
# preprocess
inputs = processor(audio_input.squeeze(), sampling_rate=sample_rate, return_tensors="pt")
# inference
with torch.no_grad():
logits = model(**inputs).logits
# decode
predicted_ids = torch.argmax(logits, dim=-1)
transcription = processor.batch_decode(predicted_ids)[0]
print(transcription)
Downstream Use
This model can be used as a foundation for:
- building voice assistants for Zulu speakers
- transcription services for Zulu content
- accessibility tools for Zulu-speaking communities
- research in low-resource speech recognition
Model Architecture
- Base model: Wav2Vec2-BERT 2.0
- Architecture: transformer-based with convolutional feature extractor
- Parameters: ~600M (inherited from base model)
- Objective: connectionist temporal classification (CTC)
Funding
The development of this model was supported by CLEAR Global and Gates Foundation.
Citation
@misc{w2v_bert_zulu_asr,
author = {Badr M. Abdullah},
title = {Adapting Wav2Vec2-BERT 2.0 for Zulu ASR},
year = {2025},
publisher = {Hugging Face},
url = {https://huggingface.co/badrex/w2v-bert-2.0-zulu-asr}
}
Model Card Contact
For questions or issues, please contact via the Hugging Face model repository in the community discussion section.