Whisper Small Lingala QLoRA Adapters
Model Description
These are the LoRA adapter weights obtained by fine-tuning openai/whisper-small for automatic speech recognition (ASR) in Lingala, using QLoRA (4-bit quantized base model + low-rank adapters). This repository contains adapter weights only — it must be loaded on top of the base Whisper Small model using the peft library; it is not a standalone model.
Intended Use
Automatic transcription of spoken Lingala audio to text, including the two Lingala-specific characters absent from the standard Latin alphabet: ɔ (U+0254) and ɛ (U+025B).
Training Data
Fine-tuned on
Congo-digital-service/audios-lingala-annotatees, a corpus of approximately 52 hours of annotated Lingala speech (Radio Rurale and Voice-a-thon recordings).
Evaluation
| Metric | Value |
|---|
| WER (Word Error Rate) | 53.38% (as of September 2026) |
This figure reflects the current QLoRA fine-tune on the existing 52h training set. A remediation effort is in progress (data cleaning, followed by a full fine-tuning trial) — this metric should be refreshed once new results are available, rather than left to go stale.
How to Use
1from peft import PeftModel
2from transformers import WhisperForConditionalGeneration, WhisperProcessor
3
4base_model_id = "openai/whisper-small"
5adapter_id = "Congo-digital-service/whisper-small-lingala-qlora-adapters"
6
7processor = WhisperProcessor.from_pretrained(base_model_id)
8base_model = WhisperForConditionalGeneration.from_pretrained(base_model_id)
9model = PeftModel.from_pretrained(base_model, adapter_id)
License
The adapter weights in this repository are released under the same Apache 2.0 License as the base model (openai/whisper-small).
In addition, because this model was fine-tuned on data released under the Nwulite Obodo Open Data License (NOODL-1.0), the following attribution requirement — carried over from the training dataset's license — applies to this model as well:
This model was created by Congo Digital Services (CDS SARL) (
https://www.congo-digital.com/) using training data licensed under the Nwulite Obodo Open Data License (
https://licensingafricandatasets.com/nwulite-obodo-license). Users from high-income countries or commercial entities are required to publicly acknowledge and credit the Maloba Project (UNDP Republic of Congo — language digitalisation initiative) in any publication, product, model, or output derived from this model. To fulfil this requirement, contact
contact@congo-digital.com or visit
https://www.congo-digital.com/contact.
Considerations for Using the Model
The current WER (53.38%) is well above the project's target (below 15%) — this model should be treated as a work-in-progress checkpoint, not a production-ready transcription tool, until the remediation effort (data cleaning, extended fine-tuning) is complete.
Creators
- Congo Digital Services (CDS SARL) — https://www.congo-digital.com/
- In collaboration with:
- the MALOBA community — https://maloba.congo-digital.com/
- Radio Rurale
- the Service National des Grandes Endémies de Brazzaville
- the Faculté des Lettres et des Langues Vivantes, Université Marien Ngouabi (UMNG)
- the Ministry of Posts, Telecommunications and Digital Economy of the Republic of Congo
- UNDP Congo
- Created in August 2026
