dondza-xitsonga-asr-wav2vec2

Dondza-Xitsonga Wav2Vec2 is a Xitsonga Automatic Speech Recognition (ASR) model fine-tuned in the context of the Dondza project (Raggio AI) from the pretrained checkpoint facebook/wav2vec2-xls-r-300m.

To our knowledge, this is among the first Mozambican-developed and publicly released end-to-end ASR models for Xitsonga.
(If you are aware of earlier Mozambican releases, please share them — we welcome corrections.)

---

License (READ FIRST) — Dual licensing for the model checkpoint

This repository provides a fine-tuned ASR checkpoint and related files.

1) Non-commercial use (default, free)

The model weights/checkpoint and associated files in this repo are released under the Dondza Non-Commercial Model License.

• ✅ Allowed: research, academic use, personal projects, demos, non-profit experiments
• ❌ Not allowed without a commercial license: use in a paid product, paid service, internal business operations, monetized apps, paid APIs, SaaS, enterprise deployments, or any other commercial exploitation
• Full terms: see LICENSE in this repository.

2) Commercial use (paid)

If you want to use this model commercially (including deploying it in a commercial product/service or using it to power a paid offering), you must obtain a Commercial License from Raggio AI.

How to request a commercial license:

• Open a discussion on this repo: https://huggingface.co/Raggio/dondza-xitsonga-asr-wav2vec2/discussions
• Or open an issue on the Dondza project tracker (if provided by the team)

We can offer:

• Commercial license for self-hosting the model weights, and/or
• Subscription access to a hosted ASR API (commercial terms/SLA available upon request)

---

Upstream / third-party licenses & attribution (important)

Even though this repo uses a custom license for the fine-tuned checkpoint, it also depends on upstream components with their own licenses.

Base model and libraries

• Base checkpoint: facebook/wav2vec2-xls-r-300m (Apache-2.0)
• Transformers library: Apache-2.0

When redistributing or using this repo, you must also comply with any applicable upstream notice and attribution requirements.

Training dataset (not redistributed here)

This model was fine-tuned on the NCHLT Xitsonga Speech Corpus (approx. 56 hours of read speech).

• We do not redistribute the dataset in this repository.
• You must obtain the dataset from its official source.
• Dataset license: CC BY 3.0 (attribution required).

Required attribution (dataset):
The Department of Arts and Culture of the government of the Republic of South Africa (DAC), the Council for Scientific and Industrial Research (CSIR), and North-West University (NWU), for the NCHLT Speech Corpus.

---

Base model

• Base checkpoint: facebook/wav2vec2-xls-r-300m
• Architecture: Wav2Vec2 + CTC
• Note: the CTC output head (lm_head.weight, lm_head.bias) is newly initialized for the Xitsonga vocabulary and learned during fine-tuning. This is expected for CTC fine-tuning.

---

Dataset

This model was fine-tuned on the NCHLT Speech Corpus (Xitsonga):

• Dataset page: https://repo.sadilar.org/items/5a6587ad-3067-49ff-bf2a-05ab171f4807
• Project page: https://sites.google.com/site/nchltspeechcorpus/home/xitsonga

The NCHLT corpus is primarily read speech. Real-world conversational audio (noise, code-switching, phone recordings) may produce higher error rates.

---

Data splits (utterances, speakers, duration)

split	utterances	unique_speakers	duration_hours	duration_hms
train	44924	190	52.658	52:39:29
val	2247	188	2.595	02:35:42
test	2905	8	3.609	03:36:31

Speaker separation

The train and validation speaker sets are disjoint (no overlap), so validation results reflect generalisation to unseen speakers.

---

Results

Best checkpoint (selected by lowest validation WER):

• Validation WER: 0.0719 (≈ 7.19%)
• Validation loss: 0.0721
• Step: 26,500
• Epoch: 29.7757
• Eval runtime: 27.787s
• Eval throughput: 80.865 samples/s

---

Intended use

• Voice input for the Dondza app (Xitsonga ASR)
• Research and prototyping for Xitsonga speech technology
• Batch transcription of read or clean speech recordings in Xitsonga

---

Limitations

• Domain mismatch: trained on read speech; conversational or noisy audio may degrade performance.
• Dialect/region: NCHLT Xitsonga data is largely from South Africa; Mozambican varieties (e.g., Changana/Ronga/Tswa influences) may differ in pronunciation and orthography.
• Not for high-stakes use: not validated for medical, legal, or other safety-critical transcription tasks.

---

How to use

import torch
import librosa
from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC

model_id = "Raggio/dondza-xitsonga-asr-wav2vec2"

processor = Wav2Vec2Processor.from_pretrained(model_id)
model = Wav2Vec2ForCTC.from_pretrained(model_id)

audio, sr = librosa.load("audio.wav", sr=16000)
inputs = processor(audio, sampling_rate=16000, return_tensors="pt", padding=True)

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

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