bli://models/bli-asr-0v0.1
/ Model release · 001 · ASR

BLI ASR 0. Lingala speech, transcribed.

Our first automatic speech-recognition model for Lingala. Built on OpenAI Whisper large-v3 and adapted with LoRA, BLI ASR 0 turns Lingala audio into text — an early, community-oriented model for an under-resourced Bantu language.

Task
01
Speech recognition
Language
02
Lingala
Base model
03
Whisper large-v3
CER · norm.
04
0.1703
Based on openai/whisper-large-v3 · adapted with LoRA / PEFT[01]Open on Hugging Face ↗
/ 01 · Overview

What it does.

BLI ASR 0 transcribes Lingala speech into Lingala text. It is not a translation model — it listens in Lingala and writes in Lingala.

It is intended as a first-pass transcription model: useful for research on low-resource ASR, for bootstrapping new datasets, and as assisted transcription before review by human annotators.

This is an early release. We publish it openly, with its limitations stated plainly, because shipping and documenting beats waiting for perfect.

/ 02 · Dataset

Trained on Waxal Lingala.

The model was trained on the Waxal Lingala ASR dataset, split into training, validation and held-out test sets.

Split
Samples
Usage
Train
≈ 14,400
Model training
Validation
1,844
Validation during development
Test
1,866
Final held-out evaluation

Text post-processing

We applied a light normalization pipeline — the goal was to reduce noise, not to impose a strict orthography. We deliberately avoided aggressive spelling correction, because Lingala has substantial orthographic variation across speakers, regions and sources.

  • Unicode normalization
  • Lowercasing
  • Whitespace normalization
  • Punctuation & symbol cleanup
  • Original raw transcription preserved when available
  • Normalized transcription field used for training / evaluation
/ 03 · Training

How it was built.

Fine-tuned from openai/whisper-large-v3 with LoRA, using the transcribe task and the Lingala language token.

Base model
openai/whisper-large-v3
Fine-tuning method
LoRA / PEFT
Task token
transcribe
Language token
Lingala
Precision
bf16
Optimizer
AdamW
Evaluation
Random validation subsets during training; full split at the end
Dataset
Waxal Lingala ASR
/ 04 · Performance
Character Error Rate · normalized
0.1703

Why we report CER, not WER

We report CER rather than WER for this first release. Lingala does not yet have a single widely enforced normalized orthography in our data, and WER strongly penalizes spelling variants, segmentation differences, and silence-related insertions and deletions. We plan to release a corrected WER metric that better accounts for linguistic and contextual variation.

/ 05 · Use & limits

Honest about both.

Intended use
  • +Lingala speech transcription
  • +Research on low-resource ASR
  • +Dataset bootstrapping
  • +Assisted transcription before human correction
  • +Evaluation of ASR pipelines for Bantu languages
Known limitations
  • !Silence handling still needs work — may hallucinate text in long silent regions
  • !Degrades with music, jingles, intros / outros and strong background noise
  • !Limited on overlapping speech in real-world media
  • !Training data does not yet cover all Lingala varieties
  • !May struggle with recent slang, urban expressions and code-switching
  • !Not yet robust across all domains (news, sermons, street interviews, music-heavy content)
/ 06 · Proof

See it transcribe.

A subtitled demonstration of BLI ASR 0 on real Lingala audio — the subtitles are produced from the model's own transcription.

bli://demo/lingala/mahele.mp4 asr
Demo · Lingala transcription · subtitles from BLI ASR 0
/ Get the model

Open weights, open to feedback.

BLI ASR 0 lives on Hugging Face. Try it, break it, and tell us where it fails — every reported failure becomes the next dataset and the next model.