MALIBA-AI/bambara-asr-v3

Bambara automatic speech recognition model. Built on OpenAI's openai/whisper-large-v3.

Currently ranked #1 on the Bambara ASR Benchmark Leaderboard, making it the best publicly available Bambara ASR model as of January 2026.

⚠️ Non-commercial use only. One of the training sources (Bible recordings) carries a license that restricts commercial use. See License.

Evaluation

Internal Test Set

Evaluated on oza75/bambara-asr (clean-combined split, 2,088 samples), 3h duration.

Condition	WER (%)	CER (%)
Raw (no normalization)	17.99	8.08
Minimal normalization	16.94	7.74
Normalized (expand mode)	13.23	6.89
Normalized (contract mode)	13.79	6.89

Normalization applied using bambara-text-normalization with the for_wer_evaluation() preset. Expand mode handles contraction disambiguation (e.g., k'a → ka a) while contract mode collapses expanded forms (e.g., bɛ a → b'a). See the normalizer documentation for details on contraction modes.

Bambara ASR Leaderboard

Evaluated on the MALIBA-AI/bambara-asr-benchmark — 1 hour of studio-recorded Malian constitutional text (pure bambara), validated by linguists from Mali's DNENF-LN.

Metric	Score
WER	45.73%
CER	13.45%
Combined (0.5 × WER + 0.5 × CER)	29.59%
Rank	🏆 1st / 37 models

See the Leaderboard

What Changed from v1

	v1	v3
Base model	whisper-large-v2	whisper-large-v3
Training data	jeli-asr + Mali-Pense	jeli-asr + Mali-Pense + Bible + Common Voice (fr/en)
Benchmark WER	61.74%	45.73%
Benchmark CER	17.90%	13.45%
License	Apache 2.0	CC-BY-NC-4.0 (non-commercial)

The ~16 point WER improvement comes from both the stronger base model and broader, more diverse training data.

Training Data

Source	Language	Description
RobotsMali/jeli-asr	Bambara	Conversational and read speech
Mali-Pense	Bambara	Transcribed audio from the Mali-Pense linguistic project
Bible recordings	Bambara	Narrated Bible text formal register. Non-commercial license
Common Voice	French, English	Supplementary multilingual data

The mix of conversational (jeli-asr), formal/literary (Bible, Mali-Pense), and multilingual (Common Voice) data gives broader coverage than previous Bambara-only systems. The French and English data helps the model handle code-switching, which is common in everyday Bambara speech in Mali.

Training

• Base model: openai/whisper-large-v3
• Method: LoRA (Low-Rank Adaptation) via PEFT

Usage

pip install git+https://github.com/sudoping01/whosper.git

from whosper import WhosperTranscriber

transcriber = WhosperTranscriber(model_id="MALIBA-AI/bambara-asr-v3")

result = transcriber.transcribe_audio("path/to/audio.wav")
print(result)

Intended Use

For:

• Research on Bambara and low-resource African language ASR
• Benchmarking and comparison of ASR systems
• Transcription assistance
• Educational and non-profit applications
• Linguistic analysis and documentation

Not for:

• Commercial products or services (license restriction from Bible training data)
• Medical, legal, or safety-critical transcription

License

CC-BY-NC-4.0 — non-commercial use only.

This restriction exists because the Bible audio used in training does not permit commercial use. If you need a commercially-licensed Bambara ASR model, see MALIBA-AI/bambara-asr-v1 (Apache 2.0, lower accuracy).

Citation

@misc{maliba_asr_v3,
  author       = {{MALIBA-AI}},
  title        = {MALIBA-AI Bambara ASR v3},
  year         = {2025},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/MALIBA-AI/bambara-asr-v3}}
}

If reporting benchmark results, please also cite:

@misc{BambaraASRBenchmark2025,
  title        = {Where Are We at with Automatic Speech Recognition for the Bambara Language?},
  author       = {Seydou Diallo and Yacouba Diarra and Mamadou K. Keita and Panga Azazia Kamat{\'e} and Adam Bouno Kampo and Aboubacar Ouattara},
  year         = {2025},
  howpublished = {Hugging Face Datasets},
  url          = {https://huggingface.co/datasets/MALIBA-AI/bambara-asr-benchmark}
}